promise-contract.json

{
  "last_updated": "2026-09-18",
  "claims": [
    {
      "id": "root_readme_hnsw_latency_microseconds",
      "file": "README.md",
      "must_contain": "HNSW Search index-only (10K/768D, k=10)",
      "validation_command": "cargo bench -p velesdb-core --bench hnsw_benchmark",
      "executable": false,
      "source": "benchmarks/results/pr363_365_comparison.md (47-56µs measured) — relabeled as 'index-only' for transparency vs the canonical 450µs end-to-end number. Corpus label corrected 5K→10K on 2026-07-20: hnsw_benchmark.rs inserts 10,000 vectors; re-measured quiet (Apple M5 Pro) at 53.2 µs.",
      "last_updated": "2026-04-27",
      "measured_on": "2026-07-20",
      "measured_machine": "Apple M5 Pro (quiet re-run, 53.2 us)",
      "measured_version": "3.12.0",
      "stale_accepted": {
        "reason": "Measured 2026-07-20: develop carried workspace 3.12.0 from v3.12.0 (2026-07-15) until the 4.0.0 bump (2026-07-24), so any build of that day reports 3.12.0; the run itself names no commit. Re-measuring belongs on the idle Apple M5 Pro it was taken on, not done in this change.",
        "granted": "2026-09-23",
        "for_workspace_major": 6
      }
    },
    {
      "id": "root_readme_simd_dot",
      "file": "README.md",
      "must_contain": "SIMD Dot Product (768D, AVX2)",
      "validation_command": "cargo bench -p velesdb-core --bench simd_benchmark",
      "executable": false,
      "source": "The label of the README's index-only dot-product figure, 21.7 ns (pinned by readme_simd_dot_product_latency). commit df217860 (2026-03-27, 'docs: update README performance section with verified production metrics'): 'Core engine: Rust index-only for architecture reference (5K/768D: 55us HNSW search, 21.7ns dot product)'; 'Measured 2026-03-27, i9-14900KF, 64GB DDR5, Windows 11, Rust 1.92.' The April 3, 2026 run in docs/BENCHMARKS.md section 1 measured 21.8 ns for the same kernel; the README's distance-metrics line quotes that run. measured_version is the workspace version of df217860.",
      "measured_on": "2026-03-27",
      "measured_machine": "Intel Core i9-14900KF, 64GB DDR5, Windows 11, Rust 1.92 (commit df217860)",
      "measured_version": "1.7.2",
      "stale_accepted": {
        "reason": "Dated by commit df217860 (2026-03-27, 1.7.2); re-measuring belongs on the idle i9-14900KF reference machine, not done in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "root_readme_ci_quality_gate",
      "file": "README.md",
      "must_contain": "cargo test --workspace",
      "validation_command": "cargo test --workspace",
      "executable": false,
      "source": "The README names the command CI's test job runs (`cargo test --workspace`): a gate, not a measured figure.",
      "notes": "Documentary: running the full workspace test suite from inside this fast contract-check script would duplicate the dedicated CI test job and be far too slow/costly to run on every doc/registry change.",
      "measured_on": "n/a",
      "measured_machine": "n/a",
      "measured_version": "n/a",
      "last_updated": "2026-09-13"
    },
    {
      "id": "core_readme_hnsw_latency",
      "file": "crates/velesdb-core/README.md",
      "must_contain": "450µs p50",
      "validation_command": "python benchmarks/velesdb_benchmark.py --recall",
      "executable": false,
      "measured_on": "2026-03-27",
      "measured_machine": "Intel Core i9-14900KF, 64 GB, Windows (benchmarks/report_1.7.2_2026-03-27.json)",
      "measured_version": "1.7.2",
      "source": "benchmarks/report_1.7.2_2026-03-27.json, written by benchmarks/velesdb_benchmark.py --recall (Python SDK, WAL ON, 10K x 384D, 200 searches after warm-up): search p50 447 us, p99 662 us; recall@10 mean 0.976 (min 0.8) on its clustered 10K x 384D recall set. CHANGELOG.md [1.8.0] (2026-03-27) publishes the run as '450us p50 search' with 'recall@10 >= 96%'. The collection default path ran HnswParams::auto and Balanced at ef_search 128 then (160 since 70b33d33, 2026-04-01).",
      "stale_accepted": {
        "reason": "Published with its original date and version (2026-03-27, 1.7.2, Balanced then at ef 128). Re-measuring means re-running velesdb_benchmark.py on the idle i9-14900KF it was taken on; the 3.3.0 Apple Silicon run (benchmarks/results/report_3.3.0_apple-silicon_2026-06-24.json, p50 366 us) is another machine and does not replace it.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "core_readme_columnstore_claim",
      "file": "crates/velesdb-core/README.md",
      "must_contain": "`ColumnStore` filtering",
      "validation_command": "cargo bench -p velesdb-core --bench column_filter_benchmark",
      "executable": false,
      "measured_on": "2026-03-19",
      "measured_machine": "Intel Core i9-14900KF (x86_64, AVX2) - docs/BENCHMARKS.md reference machine",
      "measured_version": "1.6.0",
      "source": "docs/BENCHMARKS.md section 6 '#### Integer Equality Filter (`filter_eq_int`, measured 2026-03-19)': JSON scan 3.84 ms vs ColumnStore 29.5 us at 100K rows = 130x, recorded by the v1.6.0 release commit 92cc35bd ('Re-run all 6 benchmark suites sequentially on idle machine (i9-14900KF, 64GB DDR5, Windows 11, Rust 1.92.0) — March 19, 2026'). This claim pins the label of the core README row that states it ('`ColumnStore` filtering vs. scanning JSON payloads | up to **130x** faster | integer equality, 100K rows'); core_readme_columnstore_speedup pins the number.",
      "stale_accepted": {
        "reason": "Published with its date in docs/BENCHMARKS.md section 6; the ratio is hardware-dependent and re-measuring it belongs on the idle i9-14900KF reference machine.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "core_readme_columnstore_speedup",
      "file": "crates/velesdb-core/README.md",
      "must_contain": "130x",
      "validation_command": "cargo bench -p velesdb-core --bench column_filter_benchmark",
      "executable": false,
      "source": "docs/BENCHMARKS.md section 6 '#### Integer Equality Filter (`filter_eq_int`, measured 2026-03-19)': JSON scan 3.84 ms vs ColumnStore 29.5 us at 100K rows = 130x, recorded by the v1.6.0 release commit 92cc35bd ('Re-run all 6 benchmark suites sequentially on idle machine (i9-14900KF, 64GB DDR5, Windows 11, Rust 1.92.0) — March 19, 2026'). Same run and figure as readme_columnstore_130x_claim.",
      "measured_on": "2026-03-19",
      "measured_machine": "Intel Core i9-14900KF (x86_64, AVX2) - docs/BENCHMARKS.md reference machine",
      "measured_version": "1.6.0",
      "stale_accepted": {
        "reason": "Published with its date in docs/BENCHMARKS.md section 6; the ratio is hardware-dependent and re-measuring it belongs on the idle i9-14900KF reference machine.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "readme_simd_dot_product_latency",
      "file": "README.md",
      "must_contain": "**21.7 ns**",
      "validation_command": "cargo bench -p velesdb-core --bench simd_benchmark",
      "executable": false,
      "source": "commit df217860 (2026-03-27, 'docs: update README performance section with verified production metrics'): 'Core engine: Rust index-only for architecture reference (5K/768D: 55us HNSW search, 21.7ns dot product)'; 'Measured 2026-03-27, i9-14900KF, 64GB DDR5, Windows 11, Rust 1.92.' The April 3, 2026 run in docs/BENCHMARKS.md section 1 measured 21.8 ns for the same kernel; the README's distance-metrics line quotes that run. measured_version is the workspace version of df217860.",
      "last_updated": "2026-09-13",
      "measured_on": "2026-03-27",
      "measured_machine": "Intel Core i9-14900KF, 64GB DDR5, Windows 11, Rust 1.92 (commit df217860)",
      "measured_version": "1.7.2",
      "stale_accepted": {
        "reason": "Dated by commit df217860 (2026-03-27, 1.7.2); re-measuring belongs on the idle i9-14900KF reference machine, not done in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      }
    },
    {
      "id": "readme_production_search_latency",
      "file": "README.md",
      "must_contain": "**450 us**",
      "validation_command": "python benchmarks/velesdb_benchmark.py --recall",
      "executable": false,
      "source": "benchmarks/report_1.7.2_2026-03-27.json, written by benchmarks/velesdb_benchmark.py --recall (Python SDK, WAL ON, 10K x 384D, 200 searches after warm-up): search p50 447 us, p99 662 us; recall@10 mean 0.976 (min 0.8) on its clustered 10K x 384D recall set. CHANGELOG.md [1.8.0] (2026-03-27) publishes the run as '450us p50 search' with 'recall@10 >= 96%'. The collection default path ran HnswParams::auto and Balanced at ef_search 128 then (160 since 70b33d33, 2026-04-01).",
      "last_updated": "2026-09-13",
      "measured_on": "2026-03-27",
      "measured_machine": "Intel Core i9-14900KF, 64 GB, Windows (benchmarks/report_1.7.2_2026-03-27.json)",
      "measured_version": "1.7.2",
      "stale_accepted": {
        "reason": "Published with its original date and version (2026-03-27, 1.7.2, Balanced then at ef 128). Re-measuring means re-running velesdb_benchmark.py on the idle i9-14900KF it was taken on; the 3.3.0 Apple Silicon run (benchmarks/results/report_3.3.0_apple-silicon_2026-06-24.json, p50 366 us) is another machine and does not replace it.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      }
    },
    {
      "id": "readme_hnsw_core_search_latency",
      "file": "README.md",
      "must_contain": "**55 us**",
      "validation_command": "cargo bench -p velesdb-core --bench hnsw_benchmark",
      "executable": false,
      "source": "commit df217860 (2026-03-27, 'docs: update README performance section with verified production metrics'): 'Core engine: Rust index-only for architecture reference (5K/768D: 55us HNSW search, 21.7ns dot product)' and 'Full production path: Python SDK, WAL ON, recall@10 >= 95% (10K/384D: 571us p50, 15K vec/s insert)'; 'Measured 2026-03-27, i9-14900KF, 64GB DDR5, Windows 11, Rust 1.92.' hnsw_benchmark's hnsw_search_latency group inserted 10,000 vectors at that commit, so the '5K' label was wrong; commit 45cb8021 relabeled it 10K and re-measured 53.2 us on an Apple M5 Pro (2026-07-20). measured_version is the workspace version of df217860.",
      "last_updated": "2026-09-13",
      "measured_on": "2026-03-27",
      "measured_machine": "Intel Core i9-14900KF, 64GB DDR5, Windows 11, Rust 1.92 (commit df217860)",
      "measured_version": "1.7.2",
      "stale_accepted": {
        "reason": "Dated by commit df217860 (2026-03-27, 1.7.2, i9-14900KF). The same day commit 126d5a9a recorded 47.0 us for the same bench on the same machine, and a quiet Apple M5 Pro re-run on 2026-07-20 measured 53.2 us; re-measuring belongs on the idle i9-14900KF, not done in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      }
    },
    {
      "id": "readme_recall_at_10_accurate",
      "file": "README.md",
      "must_contain": "**100%**",
      "validation_command": "cargo bench -p velesdb-core --bench recall_benchmark",
      "executable": false,
      "source": "cargo bench -p velesdb-core --bench recall_benchmark: its recall report (10K random 128-D vectors, HnswParams::max_recall index, 100 queries, k=10) at the current presets, 2026-09-10 on 6.0.0 (Accurate ef=512: 100.0%)",
      "last_updated": "2026-09-10",
      "measured_on": "2026-09-10",
      "measured_machine": "Apple M5 Pro",
      "measured_version": "6.0.0"
    },
    {
      "id": "readme_recall_fast_mode",
      "file": "README.md",
      "must_contain": "| Fast | 96 | 97.4% |",
      "validation_command": "cargo bench -p velesdb-core --bench recall_benchmark",
      "executable": false,
      "source": "cargo bench -p velesdb-core --bench recall_benchmark: its recall report (10K random 128-D vectors, HnswParams::max_recall index, 100 queries, k=10) at the current presets, 2026-09-10 on 6.0.0 (Fast ef=96: 97.4%)",
      "last_updated": "2026-09-10",
      "measured_on": "2026-09-10",
      "measured_machine": "Apple M5 Pro",
      "measured_version": "6.0.0"
    },
    {
      "id": "readme_recall_balanced_mode",
      "file": "README.md",
      "must_contain": "| Balanced (default) | 160 | 99.8% |",
      "validation_command": "cargo bench -p velesdb-core --bench recall_benchmark",
      "executable": false,
      "source": "cargo bench -p velesdb-core --bench recall_benchmark: its recall report (10K random 128-D vectors, HnswParams::max_recall index, 100 queries, k=10) at the current presets, 2026-09-10 on 6.0.0 (Balanced ef=160: 99.8%)",
      "last_updated": "2026-09-10",
      "measured_on": "2026-09-10",
      "measured_machine": "Apple M5 Pro",
      "measured_version": "6.0.0"
    },
    {
      "id": "readme_recall_accurate_mode",
      "file": "README.md",
      "must_contain": "| Accurate | 512 | 100% |",
      "validation_command": "cargo bench -p velesdb-core --bench recall_benchmark",
      "executable": false,
      "source": "cargo bench -p velesdb-core --bench recall_benchmark: its recall report (10K random 128-D vectors, HnswParams::max_recall index, 100 queries, k=10) at the current presets, 2026-09-10 on 6.0.0 (Accurate ef=512: 100.0%)",
      "last_updated": "2026-09-10",
      "measured_on": "2026-09-10",
      "measured_machine": "Apple M5 Pro",
      "measured_version": "6.0.0"
    },
    {
      "id": "readme_columnstore_130x_claim",
      "file": "README.md",
      "must_contain": "130x faster",
      "validation_command": "cargo bench -p velesdb-core --bench column_filter_benchmark",
      "executable": false,
      "source": "docs/BENCHMARKS.md section 6 '#### Integer Equality Filter (`filter_eq_int`, measured 2026-03-19)': JSON scan 3.84 ms vs ColumnStore 29.5 us at 100K rows = 130x, recorded by the v1.6.0 release commit 92cc35bd ('Re-run all 6 benchmark suites sequentially on idle machine (i9-14900KF, 64GB DDR5, Windows 11, Rust 1.92.0) — March 19, 2026'). The README names that machine and those two times. The Apple M5 Pro cross-check (~50–105x, 2026-07-20, commit 45cb8021) is a separate run, pinned by readme_columnstore_apple_ratio_range.",
      "last_updated": "2026-09-13",
      "measured_on": "2026-03-19",
      "measured_machine": "Intel Core i9-14900KF (x86_64, AVX2) - docs/BENCHMARKS.md reference machine",
      "measured_version": "1.6.0",
      "stale_accepted": {
        "reason": "Published with its date in docs/BENCHMARKS.md section 6; the ratio is hardware-dependent and re-measuring it belongs on the idle i9-14900KF reference machine.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      }
    },
    {
      "id": "benchmarks_sift1m_recall_at_10",
      "file": "docs/BENCHMARKS.md",
      "must_contain": "SIFT1M",
      "validation_command": "cargo bench -p velesdb-core --bench sift1m_recall --features bench-sift1m",
      "executable": false,
      "source": "docs/BENCHMARKS.md section 11 (SIFT1M — Standard ANN Benchmark)",
      "last_updated": "2026-04-18",
      "notes": "Gated by --features bench-sift1m + cached download or VELESDB_SIFT1M_DIR. CI skips (feature off); run locally or on a dedicated bench runner. The `must_contain` is intentionally loose (dataset name only) until numeric ranges are populated in section 11.3 from a first stable run on reference hardware — tightening it prematurely would lock the contract around literature reference points that are not VelesDB measurements.",
      "measured_on": "n/a",
      "measured_machine": "n/a",
      "measured_version": "n/a"
    },
    {
      "id": "readme_sparse_search_latency",
      "file": "README.md",
      "must_contain": "BM25 Sparse Search index-only (10K docs, top-10) 57.6 us",
      "validation_command": "cargo bench -p velesdb-core --bench sparse_benchmark -- top10_10k_corpus",
      "executable": false,
      "source": "commit f4999a74 (PR #621, 2026-04-21): 'Measured impact (i9-14900K, SPLADE 10K corpus): top-10: 956 us -> 57.6 us (-94%, 16.6x speedup); top-100: 927 us -> 75.1 us (-92%, 12.3x speedup)'. measured_version is the workspace version of that commit.",
      "last_updated": "2026-09-13",
      "measured_on": "2026-04-21 (date of commit f4999a74, which records the run)",
      "measured_machine": "i9-14900K, as commit f4999a74 names it (docs/BENCHMARKS.md's reference machine is the i9-14900KF)",
      "measured_version": "1.13.0",
      "stale_accepted": {
        "reason": "Recorded by PR #621 with its before and after on one machine (1.13.0); re-measuring belongs on the idle reference machine, not done in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      }
    },
    {
      "id": "readme_simd_cosine_latency",
      "file": "README.md",
      "must_contain": "Cosine 32 ns",
      "validation_command": "cargo bench -p velesdb-core --bench simd_benchmark",
      "executable": false,
      "source": "docs/BENCHMARKS.md section 1, SIMD Kernel Latency: Cosine 32.4 ns at 768D, set by commit 81b371ba (2026-04-03) under 'Measured April 3, 2026 on Intel Core i9-14900KF (24C/32T, AVX2+FMA), 64GB DDR5, Rust 1.94.1, Windows 11 Pro, sequential run on idle machine'. The README's distance-metrics line names that run and rounds the figure to 32 ns. measured_version is the workspace version of 81b371ba.",
      "last_updated": "2026-09-13",
      "measured_on": "2026-04-03",
      "measured_machine": "Intel Core i9-14900KF (24C/32T, AVX2+FMA), 64GB DDR5, Windows 11 Pro, Rust 1.94.1, idle (docs/BENCHMARKS.md section 1)",
      "measured_version": "1.11.0",
      "stale_accepted": {
        "reason": "Dated in docs/BENCHMARKS.md section 1 (2026-04-03, 1.11.0); re-measuring belongs on the idle i9-14900KF reference machine, not done in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      }
    },
    {
      "id": "readme_simd_euclidean_latency",
      "file": "README.md",
      "must_contain": "Euclidean 26 ns",
      "validation_command": "cargo bench -p velesdb-core --bench simd_benchmark",
      "executable": false,
      "source": "docs/BENCHMARKS.md section 1, SIMD Kernel Latency: Euclidean 26.0 ns at 768D, set by commit 81b371ba (2026-04-03) under 'Measured April 3, 2026 on Intel Core i9-14900KF (24C/32T, AVX2+FMA), 64GB DDR5, Rust 1.94.1, Windows 11 Pro, sequential run on idle machine'. The README's distance-metrics line names that run and rounds the figure to 26 ns. The line said 20 ns until 2026-09-13, while the run its claim was dated to measured 26.0 ns. measured_version is the workspace version of 81b371ba.",
      "last_updated": "2026-09-13",
      "measured_on": "2026-04-03",
      "measured_machine": "Intel Core i9-14900KF (24C/32T, AVX2+FMA), 64GB DDR5, Windows 11 Pro, Rust 1.94.1, idle (docs/BENCHMARKS.md section 1)",
      "measured_version": "1.11.0",
      "stale_accepted": {
        "reason": "Dated in docs/BENCHMARKS.md section 1 (2026-04-03, 1.11.0); re-measuring belongs on the idle i9-14900KF reference machine, not done in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      }
    },
    {
      "id": "readme_simd_dot_product_per_metric_latency",
      "file": "README.md",
      "must_contain": "Dot Product 22 ns",
      "validation_command": "cargo bench -p velesdb-core --bench simd_benchmark",
      "executable": false,
      "source": "docs/BENCHMARKS.md section 1, SIMD Kernel Latency: Dot Product 21.8 ns at 768D, set by commit 81b371ba (2026-04-03) under 'Measured April 3, 2026 on Intel Core i9-14900KF (24C/32T, AVX2+FMA), 64GB DDR5, Rust 1.94.1, Windows 11 Pro, sequential run on idle machine'. The README's distance-metrics line names that run and rounds the figure to 22 ns. Distinct from readme_simd_dot_product_latency (21.7 ns), the 2026-03-27 run the README's index-only line quotes. measured_version is the workspace version of 81b371ba.",
      "last_updated": "2026-09-13",
      "measured_on": "2026-04-03",
      "measured_machine": "Intel Core i9-14900KF (24C/32T, AVX2+FMA), 64GB DDR5, Windows 11 Pro, Rust 1.94.1, idle (docs/BENCHMARKS.md section 1)",
      "measured_version": "1.11.0",
      "stale_accepted": {
        "reason": "Dated in docs/BENCHMARKS.md section 1 (2026-04-03, 1.11.0); re-measuring belongs on the idle i9-14900KF reference machine, not done in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      }
    },
    {
      "id": "readme_simd_hamming_latency",
      "file": "README.md",
      "must_contain": "Hamming 35 ns",
      "validation_command": "cargo bench -p velesdb-core --bench simd_benchmark",
      "executable": false,
      "source": "docs/BENCHMARKS.md section 1, SIMD Kernel Latency: Hamming 35.2 ns at 768D, set by commit 81b371ba (2026-04-03) under 'Measured April 3, 2026 on Intel Core i9-14900KF (24C/32T, AVX2+FMA), 64GB DDR5, Rust 1.94.1, Windows 11 Pro, sequential run on idle machine'. The README's distance-metrics line names that run and rounds the figure to 35 ns. measured_version is the workspace version of 81b371ba.",
      "measured_on": "2026-04-03",
      "measured_machine": "Intel Core i9-14900KF (24C/32T, AVX2+FMA), 64GB DDR5, Windows 11 Pro, Rust 1.94.1, idle (docs/BENCHMARKS.md section 1)",
      "measured_version": "1.11.0",
      "stale_accepted": {
        "reason": "Dated in docs/BENCHMARKS.md section 1 (2026-04-03, 1.11.0); re-measuring belongs on the idle i9-14900KF reference machine, not done in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "readme_simd_jaccard_latency",
      "file": "README.md",
      "must_contain": "Jaccard 27 ns",
      "validation_command": "cargo bench -p velesdb-core --bench simd_benchmark",
      "executable": false,
      "source": "docs/BENCHMARKS.md section 1, SIMD Kernel Latency: Jaccard 27.3 ns at 768D, set by commit 81b371ba (2026-04-03) under 'Measured April 3, 2026 on Intel Core i9-14900KF (24C/32T, AVX2+FMA), 64GB DDR5, Rust 1.94.1, Windows 11 Pro, sequential run on idle machine'. The README's distance-metrics line names that run and rounds the figure to 27 ns. measured_version is the workspace version of 81b371ba.",
      "measured_on": "2026-04-03",
      "measured_machine": "Intel Core i9-14900KF (24C/32T, AVX2+FMA), 64GB DDR5, Windows 11 Pro, Rust 1.94.1, idle (docs/BENCHMARKS.md section 1)",
      "measured_version": "1.11.0",
      "stale_accepted": {
        "reason": "Dated in docs/BENCHMARKS.md section 1 (2026-04-03, 1.11.0); re-measuring belongs on the idle i9-14900KF reference machine, not done in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "readme_columnstore_speedup_baseline_rows",
      "file": "README.md",
      "must_contain": "100K rows",
      "validation_command": "cargo bench -p velesdb-core --bench column_filter_benchmark",
      "executable": false,
      "source": "README.md ColumnStore 130x speedup baseline. Footnote *¹ scopes the claim: filtering-API micro-benchmark, integer equality, 130x at 100K rows / 55x at 10K rows (docs/BENCHMARKS.md section 6); no 1M-row measurement exists. The filtering API serves SELECT WHERE via the adaptive payload mirror (collection/payload_mirror).",
      "last_updated": "2026-06-10",
      "measured_on": "2026-03-19",
      "measured_machine": "Intel Core i9-14900KF (x86_64, AVX2) — docs/BENCHMARKS.md reference machine",
      "measured_version": "1.6.0",
      "stale_accepted": {
        "reason": "Published with its date in docs/BENCHMARKS.md section 6 (filter_eq_int, measured 2026-03-19, 1.6.0, as the sibling readme_columnstore_130x_claim records); the 2026-07-20 run was the Apple M5 Pro cross-check, not this figure. The ratio is hardware-dependent and re-measuring it belongs on the idle i9-14900KF reference machine.",
        "granted": "2026-09-23",
        "for_workspace_major": 6
      }
    },
    {
      "id": "readme_binary_size_9mb",
      "file": "README.md",
      "must_contain": "~14 MB binary",
      "validation_command": "cargo build --release -p velesdb-server && stat -c%s target/release/velesdb-server  # measured 2026-09-02 on x86_64-unknown-linux-gnu: 14345984 bytes = 13.68 MiB",
      "executable": false,
      "source": "Measured 2026-09-02 from `cargo build --release -p velesdb-server` on x86_64-unknown-linux-gnu: 14345984 bytes (13.68 MiB); [profile.release] already sets strip = true. The Binary Size Gate measured 13.78 MiB for the same commit on ubuntu-latest with toolchain 1.90, a 10 KB delta. The figure this replaces (~10 MB) was measured at 4.0.0 on aarch64-apple-darwin and was never true for the Linux x86_64 build users download: v5.2.0, already published, builds to 14339664 bytes (13.68 MiB) on the same machine, so 6.0.0 adds 6320 bytes (0.04%) rather than regressing. Two things kept that invisible. binary-size.yml piped the gate's exit status into tee under a shell without pipefail, so the required check could not fail on an over-ceiling binary (#2193). And this entry's previous source cited a \"README footnote [3]\" stating a 7-13 MB provenance range - that footnote does not exist; README.md carries no footnotes at all, so readers saw a flat ~10 MB with no range. Apple Silicon is smaller and has not been re-measured since 4.0.0; ~14 MB is quoted for the largest target rather than the smallest.",
      "last_updated": "2026-09-02",
      "measured_on": "2026-09-02",
      "measured_machine": "x86_64-unknown-linux-gnu (size varies by target; this is the largest)",
      "measured_version": "6.0.0"
    },
    {
      "id": "readme_wasm_bundle_gzipped",
      "file": "README.md",
      "must_contain": "~710 KB gzipped",
      "validation_command": "cd crates/velesdb-wasm && wasm-pack build --target web --release && gzip -c pkg/velesdb_wasm_bg.wasm | wc -c  # measured 2026-09-02: 726700 bytes = 710 KiB (raw 2114045 = 2.02 MiB)",
      "executable": false,
      "source": "Measured 2026-09-02 at 6.0.0 with `wasm-pack build --target web --release`, the command release.yml:623 publishes from: pkg/velesdb_wasm_bg.wasm is 2114045 bytes raw and 726700 gzipped (level 6) = 710 KiB. The ~674 KB it replaces was measured on the published 4.0.0 npm package and had drifted 5.4% low over two majors. This entry's previous waiver said the figure could only come from the published npm artifact of the next release, and that \"the WASM Build Check job bounds regressions meanwhile\" - that job runs cargo check, quantization-parity tests and wasm-bindgen tests and never measures bundle size, so nothing bounded anything. The validation_command above removes the chicken-and-egg by measuring the same artifact from source, before the tag rather than after it. Caveat on the last percent: the measurement used wasm-pack 0.13.1 with binaryen 117, while CI installs the latest wasm-pack via taiki-e/install-action, which may pin a different binaryen; wasm-opt settings come from [package.metadata.wasm-pack.profile.release] (-Os, SIMD, bulk-memory, nontrapping-float-to-int) either way. wasm-opt MUST run for this number to mean anything - skipping it yields 703535 gzipped from a larger 2294177-byte raw bundle, which is not the published artifact.",
      "last_updated": "2026-09-02",
      "measured_on": "2026-09-02",
      "measured_machine": "x86_64-unknown-linux-gnu, wasm-pack 0.13.1 + binaryen 117 (platform-independent artifact; toolchain-sensitive at the percent level)",
      "measured_version": "6.0.0"
    },
    {
      "id": "readme_rest_endpoint_count",
      "file": "README.md",
      "must_contain": "54 REST endpoints",
      "validation_command": "test \"$(grep -cE '^  /' docs/openapi.yaml)\" -eq 54",
      "executable": true,
      "source": "docs/openapi.yaml on this branch: 54 paths / 61 operations, including the relations + TTL endpoints and GET /metrics. Corroborated by crates/velesdb-server/src/routes.rs (54 .route() registrations). Endpoint count convention = OpenAPI paths.",
      "last_updated": "2026-07-10",
      "measured_on": "re-derived on every run",
      "measured_machine": "platform-independent (derived from docs/openapi.yaml)",
      "measured_version": "re-derived on every run",
      "re_derives": true
    },
    {
      "id": "readme_economics_cropped_dollars_saved",
      "file": "README.md",
      "must_contain": "**10.9 %**",
      "validation_command": "grep -qF '10.9 %' README.md && grep -qF '10.9%' examples/real-session-benchmark/README.md",
      "executable": true,
      "source": "examples/real-session-benchmark/README.md:405 (billed campaign 2026-07-19, 19-turn cropped-screenshot session, cli runner: $0.4442 -> $0.3960 = 10.9% $ saved; raw logs examples/real-session-benchmark/results/2026-07-19-vibe-cli/) — consistency check only, no live billed re-run.",
      "last_updated": "2026-07-19",
      "measured_on": "2026-07-19",
      "measured_machine": "billed API campaign (Claude Code cli runner / direct Messages API) — not CPU-bound",
      "measured_version": "3.12.0",
      "stale_accepted": {
        "reason": "Dated by commit 0d461a502, which records the billed campaign's vibe-cli raw logs (2026-07-19, workspace 3.12.0); the logs name no version, so the version is that commit's. Re-measuring means re-running the billed API campaign, not done in this change.",
        "granted": "2026-09-23",
        "for_workspace_major": 6
      }
    },
    {
      "id": "readme_economics_retina_dollars_and_tokens_saved",
      "file": "README.md",
      "must_contain": "**21.9 %**",
      "validation_command": "grep -qF '21.9 %' README.md && grep -qF '21.9%' examples/real-session-benchmark/README.md && grep -qF '17.6%' examples/real-session-benchmark/README.md",
      "executable": true,
      "source": "examples/real-session-benchmark/README.md:406 (billed campaign 2026-07-19, 19-turn real-Retina-screenshot session, cli runner: $0.5693 -> $0.4444 = 21.9% $ saved, 17.6% tokens saved; raw logs examples/real-session-benchmark/results/2026-07-19-day-scale/) — consistency check only, no live billed re-run.",
      "last_updated": "2026-07-19",
      "measured_on": "2026-07-19",
      "measured_machine": "billed API campaign (Claude Code cli runner / direct Messages API) — not CPU-bound",
      "measured_version": "3.12.0",
      "stale_accepted": {
        "reason": "Dated by commit cf4e14f7d, which records the billed campaign's day-scale raw logs (2026-07-19, workspace 3.12.0); the logs name no version, so the version is that commit's. Re-measuring means re-running the billed API campaign, not done in this change.",
        "granted": "2026-09-23",
        "for_workspace_major": 6
      }
    },
    {
      "id": "readme_economics_day_scale_dollars_saved",
      "file": "README.md",
      "must_contain": "**14.7 %**",
      "validation_command": "grep -qF '14.7 %' README.md && grep -qF '14.7%' examples/real-session-benchmark/README.md",
      "executable": true,
      "source": "examples/real-session-benchmark/README.md:407 (billed campaign 2026-07-19, 36-turn day-scale session, cli runner: $1.8613 -> $1.5876 = 14.7% $ saved; raw log examples/real-session-benchmark/results/2026-07-19-day-scale/billed-long36-cli.log) — consistency check only, no live billed re-run.",
      "last_updated": "2026-07-19",
      "measured_on": "2026-07-19",
      "measured_machine": "billed API campaign (Claude Code cli runner / direct Messages API) — not CPU-bound",
      "measured_version": "3.12.0",
      "stale_accepted": {
        "reason": "Dated by commit cf4e14f7d, which records the billed campaign's day-scale raw logs (2026-07-19, workspace 3.12.0); the logs name no version, so the version is that commit's. Re-measuring means re-running the billed API campaign, not done in this change.",
        "granted": "2026-09-23",
        "for_workspace_major": 6
      }
    },
    {
      "id": "readme_economics_api_input_tokens_saved",
      "file": "README.md",
      "must_contain": "**15.1 %**",
      "validation_command": "grep -qF '15.1 %' README.md && grep -qF '15.1%' examples/real-session-benchmark/README.md",
      "executable": true,
      "source": "examples/real-session-benchmark/README.md:408 (billed campaign 2026-07-19, 19-turn session via direct Messages API runner: input_tokens 15.1% saved; raw log examples/real-session-benchmark/results/2026-07-19-day-scale/billed-vibe-api.log) — consistency check only, no live billed re-run.",
      "last_updated": "2026-07-19",
      "measured_on": "2026-07-19",
      "measured_machine": "billed API campaign (Claude Code cli runner / direct Messages API) — not CPU-bound",
      "measured_version": "3.12.0",
      "stale_accepted": {
        "reason": "Dated by commit cf4e14f7d, which records the billed campaign's day-scale raw logs (2026-07-19, workspace 3.12.0); the logs name no version, so the version is that commit's. Re-measuring means re-running the billed API campaign, not done in this change.",
        "granted": "2026-09-23",
        "for_workspace_major": 6
      }
    },
    {
      "id": "readme_economics_context_savings_corpus",
      "file": "README.md",
      "must_contain": "**82.5 %**",
      "validation_command": "grep -qF '82.5 %' README.md && grep -qF '82.5% saved' crates/velesdb-memory/examples/context_savings/README.md && grep -qF '26533 raw -> 4647 compiled' crates/velesdb-memory/examples/context_savings/README.md",
      "executable": true,
      "source": "crates/velesdb-memory/examples/context_savings/README.md:58 (committed 12-turn agent-session corpus, real cl100k BPE: 26533 raw -> 4647 compiled tokens = 82.5% saved). Re-measured 2026-09-23 on 6.0.0 with real_measures/agent_session.mjs (velesdb-node built from the tree, gpt-tokenizer 4.0.0 installed --no-save as the harness prescribes): the same 26533 -> 4647, every turn compiled twice byte-identical. The validation_command only greps the figure, so the claim ages like a documentary one.",
      "last_updated": "2026-09-23",
      "measured_on": "2026-09-23",
      "measured_machine": "platform-independent (committed corpus, deterministic tokenizer)",
      "measured_version": "6.0.0"
    },
    {
      "id": "mcp_same_named_scalar_wire_tolerance",
      "file": "docs/reference/MCP_TOOLS.md",
      "must_contain": "**Invariant:** same field name + same advertised scalar type means identical tolerance for the JSON-encoded string form.",
      "validation_command": "cargo test -p velesdb-memory --test mcp_schema_bdd same_named_scalar_input_slots_share_wire_tolerance -- --exact --test-threads=1",
      "executable": false,
      "source": "Issue #1686: the live-schema MCP regression test compares the accepted wire form for every repeated field-name and scalar-type pair.",
      "last_updated": "2026-08-15",
      "measured_on": "2026-09-02",
      "measured_machine": "platform-independent (wire-schema contract)",
      "measured_version": "6.0.0"
    },
    {
      "id": "mcp_lenient_scalar_rejects_incompatible_forms",
      "file": "docs/reference/MCP_TOOLS.md",
      "must_contain": "**Rejection invariant:** leniency never accepts an invalid JSON string or a value of another scalar or container type.",
      "validation_command": "cargo test -p velesdb-memory --test mcp_schema_bdd every_non_string_scalar_input_rejects_incompatible_forms -- --exact --test-threads=1",
      "executable": false,
      "source": "Issue #1686: the live MCP server must reject invalid strings and incompatible scalar or container values at every non-string scalar input slot.",
      "last_updated": "2026-08-15",
      "measured_on": "2026-09-02",
      "measured_machine": "platform-independent (wire-schema contract)",
      "measured_version": "6.0.0"
    },
    {
      "id": "core_perf_e2e_headline_recall",
      "file": "docs/guides/CORE_PERFORMANCE.md",
      "must_contain": "recall ≥ 96%), measured 2026-03-27 on 1.7.2, when Balanced ran at ef 128",
      "validation_command": "python benchmarks/velesdb_benchmark.py --recall",
      "executable": false,
      "source": "benchmarks/report_1.7.2_2026-03-27.json, written by benchmarks/velesdb_benchmark.py --recall (Python SDK, WAL ON, 10K x 384D, 200 searches after warm-up): search p50 447 us, p99 662 us; recall@10 mean 0.976 (min 0.8) on its clustered 10K x 384D recall set. CHANGELOG.md [1.8.0] (2026-03-27) publishes the run as '450us p50 search' with 'recall@10 >= 96%'. The collection default path ran HnswParams::auto and Balanced at ef_search 128 then (160 since 70b33d33, 2026-04-01).",
      "measured_on": "2026-03-27",
      "measured_machine": "Intel Core i9-14900KF, 64 GB, Windows (benchmarks/report_1.7.2_2026-03-27.json)",
      "measured_version": "1.7.2",
      "stale_accepted": {
        "reason": "Published with its original date and version (2026-03-27, 1.7.2, Balanced then at ef 128). Re-measuring means re-running velesdb_benchmark.py on the idle i9-14900KF it was taken on; the 3.3.0 Apple Silicon run (benchmarks/results/report_3.3.0_apple-silicon_2026-06-24.json, p50 366 us) is another machine and does not replace it.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "core_perf_e2e_key_characteristic",
      "file": "docs/guides/CORE_PERFORMANCE.md",
      "must_contain": "- End-to-end search latency: **450 µs p50** (10K/384D, WAL ON, recall ≥ 96%) —",
      "validation_command": "python benchmarks/velesdb_benchmark.py --recall",
      "executable": false,
      "source": "benchmarks/report_1.7.2_2026-03-27.json, written by benchmarks/velesdb_benchmark.py --recall (Python SDK, WAL ON, 10K x 384D, 200 searches after warm-up): search p50 447 us, p99 662 us; recall@10 mean 0.976 (min 0.8) on its clustered 10K x 384D recall set. CHANGELOG.md [1.8.0] (2026-03-27) publishes the run as '450us p50 search' with 'recall@10 >= 96%'. The collection default path ran HnswParams::auto and Balanced at ef_search 128 then (160 since 70b33d33, 2026-04-01).",
      "measured_on": "2026-03-27",
      "measured_machine": "Intel Core i9-14900KF, 64 GB, Windows (benchmarks/report_1.7.2_2026-03-27.json)",
      "measured_version": "1.7.2",
      "stale_accepted": {
        "reason": "Published with its original date and version (2026-03-27, 1.7.2, Balanced then at ef 128). Re-measuring means re-running velesdb_benchmark.py on the idle i9-14900KF it was taken on; the 3.3.0 Apple Silicon run (benchmarks/results/report_3.3.0_apple-silicon_2026-06-24.json, p50 366 us) is another machine and does not replace it.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "core_perf_e2e_headline_p50",
      "file": "docs/guides/CORE_PERFORMANCE.md",
      "must_contain": "**450 µs p50** end-to-end vector search (10K vectors, 384D, WAL ON,",
      "validation_command": "python benchmarks/velesdb_benchmark.py --recall",
      "executable": false,
      "source": "benchmarks/report_1.7.2_2026-03-27.json, written by benchmarks/velesdb_benchmark.py --recall (Python SDK, WAL ON, 10K x 384D, 200 searches after warm-up): search p50 447 us, p99 662 us; recall@10 mean 0.976 (min 0.8) on its clustered 10K x 384D recall set. CHANGELOG.md [1.8.0] (2026-03-27) publishes the run as '450us p50 search' with 'recall@10 >= 96%'. The collection default path ran HnswParams::auto and Balanced at ef_search 128 then (160 since 70b33d33, 2026-04-01).",
      "measured_on": "2026-03-27",
      "measured_machine": "Intel Core i9-14900KF, 64 GB, Windows (benchmarks/report_1.7.2_2026-03-27.json)",
      "measured_version": "1.7.2",
      "stale_accepted": {
        "reason": "Published with its original date and version (2026-03-27, 1.7.2, Balanced then at ef 128). Re-measuring means re-running velesdb_benchmark.py on the idle i9-14900KF it was taken on; the 3.3.0 Apple Silicon run (benchmarks/results/report_3.3.0_apple-silicon_2026-06-24.json, p50 366 us) is another machine and does not replace it.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "core_perf_e2e_canonical_note",
      "file": "docs/guides/CORE_PERFORMANCE.md",
      "must_contain": "> The canonical end-to-end latency remains **450 µs p50** (2026-03-27, 1.7.2).",
      "validation_command": "python benchmarks/velesdb_benchmark.py --recall",
      "executable": false,
      "source": "benchmarks/report_1.7.2_2026-03-27.json, written by benchmarks/velesdb_benchmark.py --recall (Python SDK, WAL ON, 10K x 384D, 200 searches after warm-up): search p50 447 us, p99 662 us; recall@10 mean 0.976 (min 0.8) on its clustered 10K x 384D recall set. CHANGELOG.md [1.8.0] (2026-03-27) publishes the run as '450us p50 search' with 'recall@10 >= 96%'. The collection default path ran HnswParams::auto and Balanced at ef_search 128 then (160 since 70b33d33, 2026-04-01).",
      "measured_on": "2026-03-27",
      "measured_machine": "Intel Core i9-14900KF, 64 GB, Windows (benchmarks/report_1.7.2_2026-03-27.json)",
      "measured_version": "1.7.2",
      "stale_accepted": {
        "reason": "Published with its original date and version (2026-03-27, 1.7.2, Balanced then at ef 128). Re-measuring means re-running velesdb_benchmark.py on the idle i9-14900KF it was taken on; the 3.3.0 Apple Silicon run (benchmarks/results/report_3.3.0_apple-silicon_2026-06-24.json, p50 366 us) is another machine and does not replace it.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "python_engine_e2e_p50",
      "file": "docs/guides/PYTHON_ENGINE_BENCHMARKS.md",
      "must_contain": "**~450 µs** (canonical, recall ≥ 96%; measured 2026-03-27 on 1.7.2)",
      "validation_command": "python benchmarks/velesdb_benchmark.py --recall",
      "executable": false,
      "source": "benchmarks/report_1.7.2_2026-03-27.json, written by benchmarks/velesdb_benchmark.py --recall (Python SDK, WAL ON, 10K x 384D, 200 searches after warm-up): search p50 447 us, p99 662 us; recall@10 mean 0.976 (min 0.8) on its clustered 10K x 384D recall set. CHANGELOG.md [1.8.0] (2026-03-27) publishes the run as '450us p50 search' with 'recall@10 >= 96%'. The collection default path ran HnswParams::auto and Balanced at ef_search 128 then (160 since 70b33d33, 2026-04-01).",
      "measured_on": "2026-03-27",
      "measured_machine": "Intel Core i9-14900KF, 64 GB, Windows (benchmarks/report_1.7.2_2026-03-27.json)",
      "measured_version": "1.7.2",
      "stale_accepted": {
        "reason": "Published with its original date and version (2026-03-27, 1.7.2, Balanced then at ef 128). Re-measuring means re-running velesdb_benchmark.py on the idle i9-14900KF it was taken on; the 3.3.0 Apple Silicon run (benchmarks/results/report_3.3.0_apple-silicon_2026-06-24.json, p50 366 us) is another machine and does not replace it.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "examples_ecommerce_e2e_p50_450us",
      "file": "examples/ecommerce_recommendation/README.md",
      "must_contain": "**~450µs** (10K/384D, WAL ON; measured 2026-03-27 on 1.7.2)",
      "validation_command": "python benchmarks/velesdb_benchmark.py --recall",
      "executable": false,
      "source": "benchmarks/report_1.7.2_2026-03-27.json, written by benchmarks/velesdb_benchmark.py --recall (Python SDK, WAL ON, 10K x 384D, 200 searches after warm-up): search p50 447 us, p99 662 us; recall@10 mean 0.976 (min 0.8) on its clustered 10K x 384D recall set. CHANGELOG.md [1.8.0] (2026-03-27) publishes the run as '450us p50 search' with 'recall@10 >= 96%'. The collection default path ran HnswParams::auto and Balanced at ef_search 128 then (160 since 70b33d33, 2026-04-01).",
      "measured_on": "2026-03-27",
      "measured_machine": "Intel Core i9-14900KF, 64 GB, Windows (benchmarks/report_1.7.2_2026-03-27.json)",
      "measured_version": "1.7.2",
      "stale_accepted": {
        "reason": "Published with its original date and version (2026-03-27, 1.7.2, Balanced then at ef 128). Re-measuring means re-running velesdb_benchmark.py on the idle i9-14900KF it was taken on; the 3.3.0 Apple Silicon run (benchmarks/results/report_3.3.0_apple-silicon_2026-06-24.json, p50 366 us) is another machine and does not replace it.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "server_tour_e2e_p50",
      "file": "docs/guides/SERVER_REST_TOUR.md",
      "must_contain": "~450 µs (10K/384D, WAL on, recall ≥ 96%), measured 2026-03-27 on 1.7.2",
      "validation_command": "python benchmarks/velesdb_benchmark.py --recall",
      "executable": false,
      "source": "benchmarks/report_1.7.2_2026-03-27.json, written by benchmarks/velesdb_benchmark.py --recall (Python SDK, WAL ON, 10K x 384D, 200 searches after warm-up): search p50 447 us, p99 662 us; recall@10 mean 0.976 (min 0.8) on its clustered 10K x 384D recall set. CHANGELOG.md [1.8.0] (2026-03-27) publishes the run as '450us p50 search' with 'recall@10 >= 96%'. The collection default path ran HnswParams::auto and Balanced at ef_search 128 then (160 since 70b33d33, 2026-04-01).",
      "measured_on": "2026-03-27",
      "measured_machine": "Intel Core i9-14900KF, 64 GB, Windows (benchmarks/report_1.7.2_2026-03-27.json)",
      "measured_version": "1.7.2",
      "stale_accepted": {
        "reason": "Published with its original date and version (2026-03-27, 1.7.2, Balanced then at ef 128). Re-measuring means re-running velesdb_benchmark.py on the idle i9-14900KF it was taken on; the 3.3.0 Apple Silicon run (benchmarks/results/report_3.3.0_apple-silicon_2026-06-24.json, p50 366 us) is another machine and does not replace it.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "architecture_e2e_p50",
      "file": "docs/reference/ARCHITECTURE.md",
      "must_contain": "| Search end-to-end p50 (10K/384D, WAL ON, recall ≥ 96%) | ~450 µs (2026-03-27, 1.7.2) |",
      "validation_command": "python benchmarks/velesdb_benchmark.py --recall",
      "executable": false,
      "source": "benchmarks/report_1.7.2_2026-03-27.json, written by benchmarks/velesdb_benchmark.py --recall (Python SDK, WAL ON, 10K x 384D, 200 searches after warm-up): search p50 447 us, p99 662 us; recall@10 mean 0.976 (min 0.8) on its clustered 10K x 384D recall set. CHANGELOG.md [1.8.0] (2026-03-27) publishes the run as '450us p50 search' with 'recall@10 >= 96%'. The collection default path ran HnswParams::auto and Balanced at ef_search 128 then (160 since 70b33d33, 2026-04-01).",
      "measured_on": "2026-03-27",
      "measured_machine": "Intel Core i9-14900KF, 64 GB, Windows (benchmarks/report_1.7.2_2026-03-27.json)",
      "measured_version": "1.7.2",
      "stale_accepted": {
        "reason": "Published with its original date and version (2026-03-27, 1.7.2, Balanced then at ef 128). Re-measuring means re-running velesdb_benchmark.py on the idle i9-14900KF it was taken on; the 3.3.0 Apple Silicon run (benchmarks/results/report_3.3.0_apple-silicon_2026-06-24.json, p50 366 us) is another machine and does not replace it.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "tuning_new_recall_default_path",
      "file": "docs/guides/TUNING_GUIDE.md",
      "must_contain": "measured 97.6% recall@10 end to end (10K × 384D clustered set",
      "validation_command": "python benchmarks/velesdb_benchmark.py --recall",
      "executable": false,
      "source": "benchmarks/report_1.7.2_2026-03-27.json, written by benchmarks/velesdb_benchmark.py --recall (Python SDK, WAL ON, 10K x 384D, 200 searches after warm-up): search p50 447 us, p99 662 us; recall@10 mean 0.976 (min 0.8) on its clustered 10K x 384D recall set. CHANGELOG.md [1.8.0] (2026-03-27) publishes the run as '450us p50 search' with 'recall@10 >= 96%'. The collection default path ran HnswParams::auto and Balanced at ef_search 128 then (160 since 70b33d33, 2026-04-01).",
      "measured_on": "2026-03-27",
      "measured_machine": "Intel Core i9-14900KF, 64 GB, Windows (benchmarks/report_1.7.2_2026-03-27.json)",
      "measured_version": "1.7.2",
      "stale_accepted": {
        "reason": "Published with its original date and version (2026-03-27, 1.7.2, Balanced then at ef 128). Re-measuring means re-running velesdb_benchmark.py on the idle i9-14900KF it was taken on; the 3.3.0 Apple Silicon run (benchmarks/results/report_3.3.0_apple-silicon_2026-06-24.json, p50 366 us) is another machine and does not replace it.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "native_hnsw_new_recall_default_path",
      "file": "docs/reference/NATIVE_HNSW.md",
      "must_contain": "measured 97.6% recall@10 end to end (10K × 384D clustered set",
      "validation_command": "python benchmarks/velesdb_benchmark.py --recall",
      "executable": false,
      "source": "benchmarks/report_1.7.2_2026-03-27.json, written by benchmarks/velesdb_benchmark.py --recall (Python SDK, WAL ON, 10K x 384D, 200 searches after warm-up): search p50 447 us, p99 662 us; recall@10 mean 0.976 (min 0.8) on its clustered 10K x 384D recall set. CHANGELOG.md [1.8.0] (2026-03-27) publishes the run as '450us p50 search' with 'recall@10 >= 96%'. The collection default path ran HnswParams::auto and Balanced at ef_search 128 then (160 since 70b33d33, 2026-04-01).",
      "measured_on": "2026-03-27",
      "measured_machine": "Intel Core i9-14900KF, 64 GB, Windows (benchmarks/report_1.7.2_2026-03-27.json)",
      "measured_version": "1.7.2",
      "stale_accepted": {
        "reason": "Published with its original date and version (2026-03-27, 1.7.2, Balanced then at ef 128). Re-measuring means re-running velesdb_benchmark.py on the idle i9-14900KF it was taken on; the 3.3.0 Apple Silicon run (benchmarks/results/report_3.3.0_apple-silicon_2026-06-24.json, p50 366 us) is another machine and does not replace it.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "benchmarks_balanced_recall_design_target",
      "file": "docs/BENCHMARKS.md",
      "must_contain": "the 10K table above measures Balanced at 99.8%",
      "validation_command": "cargo bench -p velesdb-core --bench recall_benchmark",
      "executable": false,
      "source": "cargo bench -p velesdb-core --bench recall_benchmark: its recall report (10K random 128-D vectors, HnswParams::max_recall index, 100 queries, k=10) at the current presets, 2026-09-10 on 6.0.0 (Fast ef=96: 97.4%, Balanced ef=160: 99.8%, Accurate ef=512: 100.0%) - the run behind readme_recall_fast_mode, readme_recall_balanced_mode, readme_recall_accurate_mode and docs/BENCHMARKS.md section 5 'HNSW Recall Profiles (10K/128D)'.",
      "measured_on": "2026-09-10",
      "measured_machine": "Apple M5 Pro",
      "measured_version": "6.0.0",
      "last_updated": "2026-09-13"
    },
    {
      "id": "mobile_api_fast_recall",
      "file": "docs/guides/MOBILE_API.md",
      "must_contain": "| `Fast` | 96 | 97.4% recall@10*, lowest latency |",
      "validation_command": "cargo bench -p velesdb-core --bench recall_benchmark",
      "executable": false,
      "source": "cargo bench -p velesdb-core --bench recall_benchmark: its recall report (10K random 128-D vectors, HnswParams::max_recall index, 100 queries, k=10) at the current presets, 2026-09-10 on 6.0.0 (Fast ef=96: 97.4%, Balanced ef=160: 99.8%, Accurate ef=512: 100.0%) - the run behind readme_recall_fast_mode, readme_recall_balanced_mode, readme_recall_accurate_mode and docs/BENCHMARKS.md section 5 'HNSW Recall Profiles (10K/128D)'.",
      "measured_on": "2026-09-10",
      "measured_machine": "Apple M5 Pro",
      "measured_version": "6.0.0",
      "last_updated": "2026-09-13"
    },
    {
      "id": "mobile_api_balanced_recall",
      "file": "docs/guides/MOBILE_API.md",
      "must_contain": "| `Balanced` | 160 | 99.8% recall@10*, default |",
      "validation_command": "cargo bench -p velesdb-core --bench recall_benchmark",
      "executable": false,
      "source": "cargo bench -p velesdb-core --bench recall_benchmark: its recall report (10K random 128-D vectors, HnswParams::max_recall index, 100 queries, k=10) at the current presets, 2026-09-10 on 6.0.0 (Fast ef=96: 97.4%, Balanced ef=160: 99.8%, Accurate ef=512: 100.0%) - the run behind readme_recall_fast_mode, readme_recall_balanced_mode, readme_recall_accurate_mode and docs/BENCHMARKS.md section 5 'HNSW Recall Profiles (10K/128D)'.",
      "measured_on": "2026-09-10",
      "measured_machine": "Apple M5 Pro",
      "measured_version": "6.0.0",
      "last_updated": "2026-09-13"
    },
    {
      "id": "mobile_api_accurate_recall",
      "file": "docs/guides/MOBILE_API.md",
      "must_contain": "| `Accurate` | 512 | 100% recall@10*;",
      "validation_command": "cargo bench -p velesdb-core --bench recall_benchmark",
      "executable": false,
      "source": "cargo bench -p velesdb-core --bench recall_benchmark: its recall report (10K random 128-D vectors, HnswParams::max_recall index, 100 queries, k=10) at the current presets, 2026-09-10 on 6.0.0 (Fast ef=96: 97.4%, Balanced ef=160: 99.8%, Accurate ef=512: 100.0%) - the run behind readme_recall_fast_mode, readme_recall_balanced_mode, readme_recall_accurate_mode and docs/BENCHMARKS.md section 5 'HNSW Recall Profiles (10K/128D)'.",
      "measured_on": "2026-09-10",
      "measured_machine": "Apple M5 Pro",
      "measured_version": "6.0.0",
      "last_updated": "2026-09-13"
    },
    {
      "id": "python_perf_accurate_recall",
      "file": "docs/guides/PYTHON_PERFORMANCE.md",
      "must_contain": "| 512 | 100%*;",
      "validation_command": "cargo bench -p velesdb-core --bench recall_benchmark",
      "executable": false,
      "source": "cargo bench -p velesdb-core --bench recall_benchmark: its recall report (10K random 128-D vectors, HnswParams::max_recall index, 100 queries, k=10) at the current presets, 2026-09-10 on 6.0.0 (Fast ef=96: 97.4%, Balanced ef=160: 99.8%, Accurate ef=512: 100.0%) - the run behind readme_recall_fast_mode, readme_recall_balanced_mode, readme_recall_accurate_mode and docs/BENCHMARKS.md section 5 'HNSW Recall Profiles (10K/128D)'.",
      "measured_on": "2026-09-10",
      "measured_machine": "Apple M5 Pro",
      "measured_version": "6.0.0",
      "last_updated": "2026-09-13"
    },
    {
      "id": "python_perf_fast_recall",
      "file": "docs/guides/PYTHON_PERFORMANCE.md",
      "must_contain": "Lower latency at 97.4% recall@10 in `recall_benchmark`",
      "validation_command": "cargo bench -p velesdb-core --bench recall_benchmark",
      "executable": false,
      "source": "cargo bench -p velesdb-core --bench recall_benchmark: its recall report (10K random 128-D vectors, HnswParams::max_recall index, 100 queries, k=10) at the current presets, 2026-09-10 on 6.0.0 (Fast ef=96: 97.4%, Balanced ef=160: 99.8%, Accurate ef=512: 100.0%) - the run behind readme_recall_fast_mode, readme_recall_balanced_mode, readme_recall_accurate_mode and docs/BENCHMARKS.md section 5 'HNSW Recall Profiles (10K/128D)'.",
      "measured_on": "2026-09-10",
      "measured_machine": "Apple M5 Pro",
      "measured_version": "6.0.0",
      "last_updated": "2026-09-13"
    },
    {
      "id": "server_tour_fast_recall",
      "file": "docs/guides/SERVER_REST_TOUR.md",
      "must_contain": "| `fast` | Low latency: 97.4% recall@10* |",
      "validation_command": "cargo bench -p velesdb-core --bench recall_benchmark",
      "executable": false,
      "source": "cargo bench -p velesdb-core --bench recall_benchmark: its recall report (10K random 128-D vectors, HnswParams::max_recall index, 100 queries, k=10) at the current presets, 2026-09-10 on 6.0.0 (Fast ef=96: 97.4%, Balanced ef=160: 99.8%, Accurate ef=512: 100.0%) - the run behind readme_recall_fast_mode, readme_recall_balanced_mode, readme_recall_accurate_mode and docs/BENCHMARKS.md section 5 'HNSW Recall Profiles (10K/128D)'.",
      "measured_on": "2026-09-10",
      "measured_machine": "Apple M5 Pro",
      "measured_version": "6.0.0",
      "last_updated": "2026-09-13"
    },
    {
      "id": "server_tour_balanced_recall",
      "file": "docs/guides/SERVER_REST_TOUR.md",
      "must_contain": "| `balanced` | Default: 99.8% recall@10* |",
      "validation_command": "cargo bench -p velesdb-core --bench recall_benchmark",
      "executable": false,
      "source": "cargo bench -p velesdb-core --bench recall_benchmark: its recall report (10K random 128-D vectors, HnswParams::max_recall index, 100 queries, k=10) at the current presets, 2026-09-10 on 6.0.0 (Fast ef=96: 97.4%, Balanced ef=160: 99.8%, Accurate ef=512: 100.0%) - the run behind readme_recall_fast_mode, readme_recall_balanced_mode, readme_recall_accurate_mode and docs/BENCHMARKS.md section 5 'HNSW Recall Profiles (10K/128D)'.",
      "measured_on": "2026-09-10",
      "measured_machine": "Apple M5 Pro",
      "measured_version": "6.0.0",
      "last_updated": "2026-09-13"
    },
    {
      "id": "server_tour_accurate_recall",
      "file": "docs/guides/SERVER_REST_TOUR.md",
      "must_contain": "| `accurate` | High precision: 100% recall@10*;",
      "validation_command": "cargo bench -p velesdb-core --bench recall_benchmark",
      "executable": false,
      "source": "cargo bench -p velesdb-core --bench recall_benchmark: its recall report (10K random 128-D vectors, HnswParams::max_recall index, 100 queries, k=10) at the current presets, 2026-09-10 on 6.0.0 (Fast ef=96: 97.4%, Balanced ef=160: 99.8%, Accurate ef=512: 100.0%) - the run behind readme_recall_fast_mode, readme_recall_balanced_mode, readme_recall_accurate_mode and docs/BENCHMARKS.md section 5 'HNSW Recall Profiles (10K/128D)'.",
      "measured_on": "2026-09-10",
      "measured_machine": "Apple M5 Pro",
      "measured_version": "6.0.0",
      "last_updated": "2026-09-13"
    },
    {
      "id": "rustdoc_config_fast_recall",
      "file": "crates/velesdb-core/src/config.rs",
      "must_contain": "`ef_search=96`: 97.4% recall@10 in `recall_benchmark`",
      "validation_command": "cargo bench -p velesdb-core --bench recall_benchmark",
      "executable": false,
      "source": "cargo bench -p velesdb-core --bench recall_benchmark: its recall report (10K random 128-D vectors, HnswParams::max_recall index, 100 queries, k=10) at the current presets, 2026-09-10 on 6.0.0 (Fast ef=96: 97.4%, Balanced ef=160: 99.8%, Accurate ef=512: 100.0%) - the run behind readme_recall_fast_mode, readme_recall_balanced_mode, readme_recall_accurate_mode and docs/BENCHMARKS.md section 5 'HNSW Recall Profiles (10K/128D)'.",
      "measured_on": "2026-09-10",
      "measured_machine": "Apple M5 Pro",
      "measured_version": "6.0.0",
      "last_updated": "2026-09-13"
    },
    {
      "id": "rustdoc_config_balanced_recall",
      "file": "crates/velesdb-core/src/config.rs",
      "must_contain": "`ef_search=160`: 99.8% recall@10 in",
      "validation_command": "cargo bench -p velesdb-core --bench recall_benchmark",
      "executable": false,
      "source": "cargo bench -p velesdb-core --bench recall_benchmark: its recall report (10K random 128-D vectors, HnswParams::max_recall index, 100 queries, k=10) at the current presets, 2026-09-10 on 6.0.0 (Fast ef=96: 97.4%, Balanced ef=160: 99.8%, Accurate ef=512: 100.0%) - the run behind readme_recall_fast_mode, readme_recall_balanced_mode, readme_recall_accurate_mode and docs/BENCHMARKS.md section 5 'HNSW Recall Profiles (10K/128D)'.",
      "measured_on": "2026-09-10",
      "measured_machine": "Apple M5 Pro",
      "measured_version": "6.0.0",
      "last_updated": "2026-09-13"
    },
    {
      "id": "rustdoc_config_accurate_recall",
      "file": "crates/velesdb-core/src/config.rs",
      "must_contain": "`ef_search=512`: 100% recall@10 in",
      "validation_command": "cargo bench -p velesdb-core --bench recall_benchmark",
      "executable": false,
      "source": "cargo bench -p velesdb-core --bench recall_benchmark: its recall report (10K random 128-D vectors, HnswParams::max_recall index, 100 queries, k=10) at the current presets, 2026-09-10 on 6.0.0 (Fast ef=96: 97.4%, Balanced ef=160: 99.8%, Accurate ef=512: 100.0%) - the run behind readme_recall_fast_mode, readme_recall_balanced_mode, readme_recall_accurate_mode and docs/BENCHMARKS.md section 5 'HNSW Recall Profiles (10K/128D)'.",
      "measured_on": "2026-09-10",
      "measured_machine": "Apple M5 Pro",
      "measured_version": "6.0.0",
      "last_updated": "2026-09-13"
    },
    {
      "id": "rustdoc_params_fast_recall",
      "file": "crates/velesdb-core/src/index/hnsw/params.rs",
      "must_contain": "`ef_search=96`: 97.4% recall@10 in `recall_benchmark`",
      "validation_command": "cargo bench -p velesdb-core --bench recall_benchmark",
      "executable": false,
      "source": "cargo bench -p velesdb-core --bench recall_benchmark: its recall report (10K random 128-D vectors, HnswParams::max_recall index, 100 queries, k=10) at the current presets, 2026-09-10 on 6.0.0 (Fast ef=96: 97.4%, Balanced ef=160: 99.8%, Accurate ef=512: 100.0%) - the run behind readme_recall_fast_mode, readme_recall_balanced_mode, readme_recall_accurate_mode and docs/BENCHMARKS.md section 5 'HNSW Recall Profiles (10K/128D)'.",
      "measured_on": "2026-09-10",
      "measured_machine": "Apple M5 Pro",
      "measured_version": "6.0.0",
      "last_updated": "2026-09-13"
    },
    {
      "id": "rustdoc_params_balanced_recall",
      "file": "crates/velesdb-core/src/index/hnsw/params.rs",
      "must_contain": "`ef_search=160`: 99.8% recall@10 in",
      "validation_command": "cargo bench -p velesdb-core --bench recall_benchmark",
      "executable": false,
      "source": "cargo bench -p velesdb-core --bench recall_benchmark: its recall report (10K random 128-D vectors, HnswParams::max_recall index, 100 queries, k=10) at the current presets, 2026-09-10 on 6.0.0 (Fast ef=96: 97.4%, Balanced ef=160: 99.8%, Accurate ef=512: 100.0%) - the run behind readme_recall_fast_mode, readme_recall_balanced_mode, readme_recall_accurate_mode and docs/BENCHMARKS.md section 5 'HNSW Recall Profiles (10K/128D)'.",
      "measured_on": "2026-09-10",
      "measured_machine": "Apple M5 Pro",
      "measured_version": "6.0.0",
      "last_updated": "2026-09-13"
    },
    {
      "id": "rustdoc_params_accurate_recall",
      "file": "crates/velesdb-core/src/index/hnsw/params.rs",
      "must_contain": "`ef_search=512`: 100% recall@10 in",
      "validation_command": "cargo bench -p velesdb-core --bench recall_benchmark",
      "executable": false,
      "source": "cargo bench -p velesdb-core --bench recall_benchmark: its recall report (10K random 128-D vectors, HnswParams::max_recall index, 100 queries, k=10) at the current presets, 2026-09-10 on 6.0.0 (Fast ef=96: 97.4%, Balanced ef=160: 99.8%, Accurate ef=512: 100.0%) - the run behind readme_recall_fast_mode, readme_recall_balanced_mode, readme_recall_accurate_mode and docs/BENCHMARKS.md section 5 'HNSW Recall Profiles (10K/128D)'.",
      "measured_on": "2026-09-10",
      "measured_machine": "Apple M5 Pro",
      "measured_version": "6.0.0",
      "last_updated": "2026-09-13"
    },
    {
      "id": "rustdoc_hnsw_index_fast_recall",
      "file": "crates/velesdb-core/src/index/hnsw/index/mod.rs",
      "must_contain": "`ef_search=96`, 97.4% recall@10 in `recall_benchmark`",
      "validation_command": "cargo bench -p velesdb-core --bench recall_benchmark",
      "executable": false,
      "source": "cargo bench -p velesdb-core --bench recall_benchmark: its recall report (10K random 128-D vectors, HnswParams::max_recall index, 100 queries, k=10) at the current presets, 2026-09-10 on 6.0.0 (Fast ef=96: 97.4%, Balanced ef=160: 99.8%, Accurate ef=512: 100.0%) - the run behind readme_recall_fast_mode, readme_recall_balanced_mode, readme_recall_accurate_mode and docs/BENCHMARKS.md section 5 'HNSW Recall Profiles (10K/128D)'.",
      "measured_on": "2026-09-10",
      "measured_machine": "Apple M5 Pro",
      "measured_version": "6.0.0",
      "last_updated": "2026-09-13"
    },
    {
      "id": "rustdoc_hnsw_index_balanced_recall",
      "file": "crates/velesdb-core/src/index/hnsw/index/mod.rs",
      "must_contain": "`ef_search=160`, 99.8% recall@10 there",
      "validation_command": "cargo bench -p velesdb-core --bench recall_benchmark",
      "executable": false,
      "source": "cargo bench -p velesdb-core --bench recall_benchmark: its recall report (10K random 128-D vectors, HnswParams::max_recall index, 100 queries, k=10) at the current presets, 2026-09-10 on 6.0.0 (Fast ef=96: 97.4%, Balanced ef=160: 99.8%, Accurate ef=512: 100.0%) - the run behind readme_recall_fast_mode, readme_recall_balanced_mode, readme_recall_accurate_mode and docs/BENCHMARKS.md section 5 'HNSW Recall Profiles (10K/128D)'.",
      "measured_on": "2026-09-10",
      "measured_machine": "Apple M5 Pro",
      "measured_version": "6.0.0",
      "last_updated": "2026-09-13"
    },
    {
      "id": "rustdoc_hnsw_index_accurate_recall",
      "file": "crates/velesdb-core/src/index/hnsw/index/mod.rs",
      "must_contain": "`ef_search=512`, 100% recall@10 there",
      "validation_command": "cargo bench -p velesdb-core --bench recall_benchmark",
      "executable": false,
      "source": "cargo bench -p velesdb-core --bench recall_benchmark: its recall report (10K random 128-D vectors, HnswParams::max_recall index, 100 queries, k=10) at the current presets, 2026-09-10 on 6.0.0 (Fast ef=96: 97.4%, Balanced ef=160: 99.8%, Accurate ef=512: 100.0%) - the run behind readme_recall_fast_mode, readme_recall_balanced_mode, readme_recall_accurate_mode and docs/BENCHMARKS.md section 5 'HNSW Recall Profiles (10K/128D)'.",
      "measured_on": "2026-09-10",
      "measured_machine": "Apple M5 Pro",
      "measured_version": "6.0.0",
      "last_updated": "2026-09-13"
    },
    {
      "id": "rustdoc_mobile_fast_recall",
      "file": "crates/velesdb-mobile/src/types.rs",
      "must_contain": "(`ef_search=96`): 97.4% recall@10 in the core's",
      "validation_command": "cargo bench -p velesdb-core --bench recall_benchmark",
      "executable": false,
      "source": "cargo bench -p velesdb-core --bench recall_benchmark: its recall report (10K random 128-D vectors, HnswParams::max_recall index, 100 queries, k=10) at the current presets, 2026-09-10 on 6.0.0 (Fast ef=96: 97.4%, Balanced ef=160: 99.8%, Accurate ef=512: 100.0%) - the run behind readme_recall_fast_mode, readme_recall_balanced_mode, readme_recall_accurate_mode and docs/BENCHMARKS.md section 5 'HNSW Recall Profiles (10K/128D)'.",
      "measured_on": "2026-09-10",
      "measured_machine": "Apple M5 Pro",
      "measured_version": "6.0.0",
      "last_updated": "2026-09-13"
    },
    {
      "id": "rustdoc_mobile_balanced_recall",
      "file": "crates/velesdb-mobile/src/types.rs",
      "must_contain": "(`ef_search=160`): 99.8% recall@10 in the core's",
      "validation_command": "cargo bench -p velesdb-core --bench recall_benchmark",
      "executable": false,
      "source": "cargo bench -p velesdb-core --bench recall_benchmark: its recall report (10K random 128-D vectors, HnswParams::max_recall index, 100 queries, k=10) at the current presets, 2026-09-10 on 6.0.0 (Fast ef=96: 97.4%, Balanced ef=160: 99.8%, Accurate ef=512: 100.0%) - the run behind readme_recall_fast_mode, readme_recall_balanced_mode, readme_recall_accurate_mode and docs/BENCHMARKS.md section 5 'HNSW Recall Profiles (10K/128D)'.",
      "measured_on": "2026-09-10",
      "measured_machine": "Apple M5 Pro",
      "measured_version": "6.0.0",
      "last_updated": "2026-09-13"
    },
    {
      "id": "rustdoc_mobile_accurate_recall",
      "file": "crates/velesdb-mobile/src/types.rs",
      "must_contain": "(`ef_search=512`): 100% recall@10 in the core's",
      "validation_command": "cargo bench -p velesdb-core --bench recall_benchmark",
      "executable": false,
      "source": "cargo bench -p velesdb-core --bench recall_benchmark: its recall report (10K random 128-D vectors, HnswParams::max_recall index, 100 queries, k=10) at the current presets, 2026-09-10 on 6.0.0 (Fast ef=96: 97.4%, Balanced ef=160: 99.8%, Accurate ef=512: 100.0%) - the run behind readme_recall_fast_mode, readme_recall_balanced_mode, readme_recall_accurate_mode and docs/BENCHMARKS.md section 5 'HNSW Recall Profiles (10K/128D)'.",
      "measured_on": "2026-09-10",
      "measured_machine": "Apple M5 Pro",
      "measured_version": "6.0.0",
      "last_updated": "2026-09-13"
    },
    {
      "id": "benchmarks_accurate_sift1m_recall",
      "file": "docs/BENCHMARKS.md",
      "must_contain": "§11.3 measures Accurate at 0.98",
      "validation_command": "cargo bench -p velesdb-core --bench sift1m_recall --features bench-sift1m",
      "executable": false,
      "source": "docs/BENCHMARKS.md section 11.3, production path: Accurate recall@10 0.9803 on SIFT1M (1M base x 10K queries, M=16, ef_construction=200, L2), 'First reproducible run, VelesDB v3.3.0' on Apple Silicon, otherwise idle. Recorded by #1225, squashed as 4b62338f (2026-06-24), whose message reads 'Adds report_recall_quality() printing RECALL_REPORT_QUALITY lines for Accurate (ef~1024 -> 0.9803)'; the PR commit that added that report, b33d7ebd, is dated 2026-06-24 too. Section 11.3 names the version and host but not the day, so measured_on is the day of the commit that records the run.",
      "measured_on": "2026-06-24 (date of commit 4b62338f, which records the run)",
      "measured_machine": "Apple Silicon (M-series, ARM64/NEON), otherwise idle - docs/BENCHMARKS.md section 11.3",
      "measured_version": "3.3.0",
      "stale_accepted": {
        "reason": "Published with its version (3.3.0) and host in docs/BENCHMARKS.md section 11.3. No full 1M SIFT1M re-run has been recorded since: the nightly SIFT1M job gates recall@10 >= 0.90 at ef=128 but commits no figures, so the figure stays pinned to the run that produced it.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "mobile_api_accurate_sift1m",
      "file": "docs/guides/MOBILE_API.md",
      "must_contain": "0.98 on SIFT1M's 1M |",
      "validation_command": "cargo bench -p velesdb-core --bench sift1m_recall --features bench-sift1m",
      "executable": false,
      "source": "docs/BENCHMARKS.md section 11.3, production path: Accurate recall@10 0.9803 on SIFT1M (1M base x 10K queries, M=16, ef_construction=200, L2), 'First reproducible run, VelesDB v3.3.0' on Apple Silicon, otherwise idle. Recorded by #1225, squashed as 4b62338f (2026-06-24), whose message reads 'Adds report_recall_quality() printing RECALL_REPORT_QUALITY lines for Accurate (ef~1024 -> 0.9803)'; the PR commit that added that report, b33d7ebd, is dated 2026-06-24 too. Section 11.3 names the version and host but not the day, so measured_on is the day of the commit that records the run.",
      "measured_on": "2026-06-24 (date of commit 4b62338f, which records the run)",
      "measured_machine": "Apple Silicon (M-series, ARM64/NEON), otherwise idle - docs/BENCHMARKS.md section 11.3",
      "measured_version": "3.3.0",
      "stale_accepted": {
        "reason": "Published with its version (3.3.0) and host in docs/BENCHMARKS.md section 11.3. No full 1M SIFT1M re-run has been recorded since: the nightly SIFT1M job gates recall@10 >= 0.90 at ef=128 but commits no figures, so the figure stays pinned to the run that produced it.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "python_perf_accurate_sift1m",
      "file": "docs/guides/PYTHON_PERFORMANCE.md",
      "must_contain": "100%*; 0.98 on SIFT1M's 1M",
      "validation_command": "cargo bench -p velesdb-core --bench sift1m_recall --features bench-sift1m",
      "executable": false,
      "source": "docs/BENCHMARKS.md section 11.3, production path: Accurate recall@10 0.9803 on SIFT1M (1M base x 10K queries, M=16, ef_construction=200, L2), 'First reproducible run, VelesDB v3.3.0' on Apple Silicon, otherwise idle. Recorded by #1225, squashed as 4b62338f (2026-06-24), whose message reads 'Adds report_recall_quality() printing RECALL_REPORT_QUALITY lines for Accurate (ef~1024 -> 0.9803)'; the PR commit that added that report, b33d7ebd, is dated 2026-06-24 too. Section 11.3 names the version and host but not the day, so measured_on is the day of the commit that records the run.",
      "measured_on": "2026-06-24 (date of commit 4b62338f, which records the run)",
      "measured_machine": "Apple Silicon (M-series, ARM64/NEON), otherwise idle - docs/BENCHMARKS.md section 11.3",
      "measured_version": "3.3.0",
      "stale_accepted": {
        "reason": "Published with its version (3.3.0) and host in docs/BENCHMARKS.md section 11.3. No full 1M SIFT1M re-run has been recorded since: the nightly SIFT1M job gates recall@10 >= 0.90 at ef=128 but commits no figures, so the figure stays pinned to the run that produced it.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "server_tour_accurate_sift1m",
      "file": "docs/guides/SERVER_REST_TOUR.md",
      "must_contain": "100% recall@10*; 0.98 on SIFT1M's 1M",
      "validation_command": "cargo bench -p velesdb-core --bench sift1m_recall --features bench-sift1m",
      "executable": false,
      "source": "docs/BENCHMARKS.md section 11.3, production path: Accurate recall@10 0.9803 on SIFT1M (1M base x 10K queries, M=16, ef_construction=200, L2), 'First reproducible run, VelesDB v3.3.0' on Apple Silicon, otherwise idle. Recorded by #1225, squashed as 4b62338f (2026-06-24), whose message reads 'Adds report_recall_quality() printing RECALL_REPORT_QUALITY lines for Accurate (ef~1024 -> 0.9803)'; the PR commit that added that report, b33d7ebd, is dated 2026-06-24 too. Section 11.3 names the version and host but not the day, so measured_on is the day of the commit that records the run.",
      "measured_on": "2026-06-24 (date of commit 4b62338f, which records the run)",
      "measured_machine": "Apple Silicon (M-series, ARM64/NEON), otherwise idle - docs/BENCHMARKS.md section 11.3",
      "measured_version": "3.3.0",
      "stale_accepted": {
        "reason": "Published with its version (3.3.0) and host in docs/BENCHMARKS.md section 11.3. No full 1M SIFT1M re-run has been recorded since: the nightly SIFT1M job gates recall@10 >= 0.90 at ef=128 but commits no figures, so the figure stays pinned to the run that produced it.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "rustdoc_hnsw_index_accurate_sift1m",
      "file": "crates/velesdb-core/src/index/hnsw/index/mod.rs",
      "must_contain": "100% recall@10 there and 0.98 on SIFT1M's",
      "validation_command": "cargo bench -p velesdb-core --bench sift1m_recall --features bench-sift1m",
      "executable": false,
      "source": "docs/BENCHMARKS.md section 11.3, production path: Accurate recall@10 0.9803 on SIFT1M (1M base x 10K queries, M=16, ef_construction=200, L2), 'First reproducible run, VelesDB v3.3.0' on Apple Silicon, otherwise idle. Recorded by #1225, squashed as 4b62338f (2026-06-24), whose message reads 'Adds report_recall_quality() printing RECALL_REPORT_QUALITY lines for Accurate (ef~1024 -> 0.9803)'; the PR commit that added that report, b33d7ebd, is dated 2026-06-24 too. Section 11.3 names the version and host but not the day, so measured_on is the day of the commit that records the run.",
      "measured_on": "2026-06-24 (date of commit 4b62338f, which records the run)",
      "measured_machine": "Apple Silicon (M-series, ARM64/NEON), otherwise idle - docs/BENCHMARKS.md section 11.3",
      "measured_version": "3.3.0",
      "stale_accepted": {
        "reason": "Published with its version (3.3.0) and host in docs/BENCHMARKS.md section 11.3. No full 1M SIFT1M re-run has been recorded since: the nightly SIFT1M job gates recall@10 >= 0.90 at ef=128 but commits no figures, so the figure stays pinned to the run that produced it.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "python_perf_accurate_latency_ratio",
      "file": "docs/guides/PYTHON_PERFORMANCE.md",
      "must_contain": "~2x slower than balanced in BENCHMARKS.md's native index-level p50s (130 vs 57 µs",
      "validation_command": "not named: docs/BENCHMARKS.md section 5 does not say which bench produced its Latency P50 column (recall_benchmark reports recall only)",
      "executable": false,
      "source": "docs/BENCHMARKS.md section 5 'HNSW Recall Profiles (10K/128D)': 'Latencies were measured March 19, 2026, on the reference machine with the ef defaults of that time (Fast=64, Balanced=128)' - Accurate 130 us, Balanced 57 us p50; recorded by the v1.6.0 release commit 92cc35bd ('Re-run all 6 benchmark suites sequentially on idle machine (i9-14900KF').",
      "measured_on": "2026-03-19",
      "measured_machine": "Intel Core i9-14900KF, idle - docs/BENCHMARKS.md reference machine (release commit 92cc35bd)",
      "measured_version": "1.6.0",
      "stale_accepted": {
        "reason": "Published with its date, and with Balanced's ef of that time, in the line that cites it; re-measuring the latencies belongs on the idle i9-14900KF reference machine.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "benchmarks_index_constructor_insert_speed",
      "file": "docs/BENCHMARKS.md",
      "must_contain": "### Index Constructor Insert Speed",
      "covers_section": true,
      "validation_command": "cargo bench -p velesdb-core --bench hnsw_benchmark -- hnsw_insert_fast",
      "executable": false,
      "source": "benchmarks/results/pr363_365_comparison.md section 1, Sequential Insert (clean re-run on a quiet machine, Criterion, 100 samples; hnsw_benchmark group hnsw_insert_fast, 1K x 768D; 'Current' = commit 4eb1c1f0): HnswIndex::new 275.08 ms, new_fast_insert 99.60 ms (2.8x), new_turbo 55.68 ms (4.9x). auto(), fast_indexing() and turbo() compute the same parameters today as at 4eb1c1f0. measured_version is the workspace version of commit 4eb1c1f0.",
      "measured_on": "2026-03-23",
      "measured_machine": "Intel Core i9-14900KF, 64GB, Windows 11 Pro, quiet machine (benchmarks/results/pr363_365_comparison.md)",
      "measured_version": "1.6.0",
      "stale_accepted": {
        "reason": "Published with its date and version (2026-03-23, 1.6.0); the three parameter formulas it compares are unchanged since, and re-measuring the ratio belongs on the idle i9-14900KF reference machine the report used.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "tuning_fast_insert_speed",
      "file": "docs/guides/TUNING_GUIDE.md",
      "must_contain": "| 2.8x faster ² |",
      "validation_command": "cargo bench -p velesdb-core --bench hnsw_benchmark -- hnsw_insert_fast",
      "executable": false,
      "source": "benchmarks/results/pr363_365_comparison.md section 1, Sequential Insert (clean re-run on a quiet machine, Criterion, 100 samples; hnsw_benchmark group hnsw_insert_fast, 1K x 768D; 'Current' = commit 4eb1c1f0): HnswIndex::new 275.08 ms, new_fast_insert 99.60 ms (2.8x), new_turbo 55.68 ms (4.9x). auto(), fast_indexing() and turbo() compute the same parameters today as at 4eb1c1f0. measured_version is the workspace version of commit 4eb1c1f0.",
      "measured_on": "2026-03-23",
      "measured_machine": "Intel Core i9-14900KF, 64GB, Windows 11 Pro, quiet machine (benchmarks/results/pr363_365_comparison.md)",
      "measured_version": "1.6.0",
      "stale_accepted": {
        "reason": "Published with its date and version (2026-03-23, 1.6.0); the three parameter formulas it compares are unchanged since, and re-measuring the ratio belongs on the idle i9-14900KF reference machine the report used.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "tuning_turbo_speed",
      "file": "docs/guides/TUNING_GUIDE.md",
      "must_contain": "| 4.9x faster ² |",
      "validation_command": "cargo bench -p velesdb-core --bench hnsw_benchmark -- hnsw_insert_fast",
      "executable": false,
      "source": "benchmarks/results/pr363_365_comparison.md section 1, Sequential Insert (clean re-run on a quiet machine, Criterion, 100 samples; hnsw_benchmark group hnsw_insert_fast, 1K x 768D; 'Current' = commit 4eb1c1f0): HnswIndex::new 275.08 ms, new_fast_insert 99.60 ms (2.8x), new_turbo 55.68 ms (4.9x). auto(), fast_indexing() and turbo() compute the same parameters today as at 4eb1c1f0. measured_version is the workspace version of commit 4eb1c1f0.",
      "measured_on": "2026-03-23",
      "measured_machine": "Intel Core i9-14900KF, 64GB, Windows 11 Pro, quiet machine (benchmarks/results/pr363_365_comparison.md)",
      "measured_version": "1.6.0",
      "stale_accepted": {
        "reason": "Published with its date and version (2026-03-23, 1.6.0); the three parameter formulas it compares are unchanged since, and re-measuring the ratio belongs on the idle i9-14900KF reference machine the report used.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "native_hnsw_turbo_speed",
      "file": "docs/reference/NATIVE_HNSW.md",
      "must_contain": "| 4.9x faster ² |",
      "validation_command": "cargo bench -p velesdb-core --bench hnsw_benchmark -- hnsw_insert_fast",
      "executable": false,
      "source": "benchmarks/results/pr363_365_comparison.md section 1, Sequential Insert (clean re-run on a quiet machine, Criterion, 100 samples; hnsw_benchmark group hnsw_insert_fast, 1K x 768D; 'Current' = commit 4eb1c1f0): HnswIndex::new 275.08 ms, new_fast_insert 99.60 ms (2.8x), new_turbo 55.68 ms (4.9x). auto(), fast_indexing() and turbo() compute the same parameters today as at 4eb1c1f0. measured_version is the workspace version of commit 4eb1c1f0.",
      "measured_on": "2026-03-23",
      "measured_machine": "Intel Core i9-14900KF, 64GB, Windows 11 Pro, quiet machine (benchmarks/results/pr363_365_comparison.md)",
      "measured_version": "1.6.0",
      "stale_accepted": {
        "reason": "Published with its date and version (2026-03-23, 1.6.0); the three parameter formulas it compares are unchanged since, and re-measuring the ratio belongs on the idle i9-14900KF reference machine the report used.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "benchmarks_velesql_simple_parse",
      "file": "docs/BENCHMARKS.md",
      "must_contain": "| Simple Parse | 1.26 µs | 794K QPS |",
      "validation_command": "cargo bench -p velesdb-core --features internal-bench --bench velesql_benchmark -- --noplot",
      "executable": false,
      "source": "docs/BENCHMARKS.md section 7 'Measured March 19, 2026, sequential run on idle machine', recorded by the v1.6.0 release commit 92cc35bd: 'Re-run all 6 benchmark suites sequentially on idle machine (i9-14900KF, ...)', 'VelesQL parser (1.26us simple, ...)'.",
      "measured_on": "2026-03-19",
      "measured_machine": "Intel Core i9-14900KF, idle - docs/BENCHMARKS.md reference machine (release commit 92cc35bd)",
      "measured_version": "1.6.0",
      "stale_accepted": {
        "reason": "Published with its date in docs/BENCHMARKS.md section 7; re-measuring VelesQL parse latencies belongs on the idle i9-14900KF reference machine.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "benchmarks_velesql_vector_query",
      "file": "docs/BENCHMARKS.md",
      "must_contain": "| Vector Query | 1.77 µs | 565K QPS |",
      "validation_command": "cargo bench -p velesdb-core --features internal-bench --bench velesql_benchmark -- --noplot",
      "executable": false,
      "source": "docs/BENCHMARKS.md section 7 'Measured March 19, 2026, sequential run on idle machine', recorded by the v1.6.0 release commit 92cc35bd: 'Re-run all 6 benchmark suites sequentially on idle machine (i9-14900KF, ...)', 'VelesQL parser (1.26us simple, ...)'.",
      "measured_on": "2026-03-19",
      "measured_machine": "Intel Core i9-14900KF, idle - docs/BENCHMARKS.md reference machine (release commit 92cc35bd)",
      "measured_version": "1.6.0",
      "stale_accepted": {
        "reason": "Published with its date in docs/BENCHMARKS.md section 7; re-measuring VelesQL parse latencies belongs on the idle i9-14900KF reference machine.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "benchmarks_velesql_complex_query",
      "file": "docs/BENCHMARKS.md",
      "must_contain": "| Complex Query | 7.47 µs | 134K QPS |",
      "validation_command": "cargo bench -p velesdb-core --features internal-bench --bench velesql_benchmark -- --noplot",
      "executable": false,
      "source": "docs/BENCHMARKS.md section 7 'Measured March 19, 2026, sequential run on idle machine', recorded by the v1.6.0 release commit 92cc35bd: 'Re-run all 6 benchmark suites sequentially on idle machine (i9-14900KF, ...)', 'VelesQL parser (1.26us simple, ...)'.",
      "measured_on": "2026-03-19",
      "measured_machine": "Intel Core i9-14900KF, idle - docs/BENCHMARKS.md reference machine (release commit 92cc35bd)",
      "measured_version": "1.6.0",
      "stale_accepted": {
        "reason": "Published with its date in docs/BENCHMARKS.md section 7; re-measuring VelesQL parse latencies belongs on the idle i9-14900KF reference machine.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "benchmarks_velesql_explain_plan",
      "file": "docs/BENCHMARKS.md",
      "must_contain": "| EXPLAIN Plan (simple) | 65.4 ns | 15.3M QPS |",
      "validation_command": "cargo bench -p velesdb-core --features internal-bench --bench velesql_benchmark -- --noplot",
      "executable": false,
      "source": "docs/BENCHMARKS.md section 7 'Measured March 19, 2026, sequential run on idle machine', recorded by the v1.6.0 release commit 92cc35bd: 'Re-run all 6 benchmark suites sequentially on idle machine (i9-14900KF, ...)', 'VelesQL parser (1.26us simple, ...)'.",
      "measured_on": "2026-03-19",
      "measured_machine": "Intel Core i9-14900KF, idle - docs/BENCHMARKS.md reference machine (release commit 92cc35bd)",
      "measured_version": "1.6.0",
      "stale_accepted": {
        "reason": "Published with its date in docs/BENCHMARKS.md section 7; re-measuring VelesQL parse latencies belongs on the idle i9-14900KF reference machine.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "benchmarks_velesql_cache_hit",
      "file": "docs/BENCHMARKS.md",
      "must_contain": "| **Cache Hit** | **1.08 µs** | **926K QPS** |",
      "validation_command": "cargo bench -p velesdb-core --features internal-bench --bench velesql_benchmark -- --noplot",
      "executable": false,
      "source": "commit dd0d2457 (2026-03-24, 'real measurements March 24, 2026 22:56'): 'All metrics re-measured sequentially on idle i9-14900KF, 64GB DDR5, Windows 11, Rust 1.92.0, Criterion.rs' - 'VelesQL Cache Hit: 1.08 us (~926K QPS)', replacing the March 19 run's 1.06 us. measured_version is the workspace version of that commit.",
      "measured_on": "2026-03-24",
      "measured_machine": "Intel Core i9-14900KF, 64GB DDR5, Windows 11, idle (commit dd0d2457)",
      "measured_version": "1.7.0",
      "stale_accepted": {
        "reason": "Dated in docs/BENCHMARKS.md section 7 (the March 24, 2026 re-run); re-measuring belongs on the idle i9-14900KF reference machine.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "core_perf_velesql_cache_hit",
      "file": "docs/guides/CORE_PERFORMANCE.md",
      "must_contain": "| VelesQL cache hit | **1.08 µs** (~926K QPS) |",
      "validation_command": "cargo bench -p velesdb-core --features internal-bench --bench velesql_benchmark -- --noplot",
      "executable": false,
      "source": "commit dd0d2457 (2026-03-24, 'real measurements March 24, 2026 22:56'): 'All metrics re-measured sequentially on idle i9-14900KF, 64GB DDR5, Windows 11, Rust 1.92.0, Criterion.rs' - 'VelesQL Cache Hit: 1.08 us (~926K QPS)', replacing the March 19 run's 1.06 us. measured_version is the workspace version of that commit.",
      "measured_on": "2026-03-24",
      "measured_machine": "Intel Core i9-14900KF, 64GB DDR5, Windows 11, idle (commit dd0d2457)",
      "measured_version": "1.7.0",
      "stale_accepted": {
        "reason": "Dated in docs/BENCHMARKS.md section 7 (the March 24, 2026 re-run); re-measuring belongs on the idle i9-14900KF reference machine.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "benchmarks_hnsw_search_k10_qps",
      "file": "docs/BENCHMARKS.md",
      "must_contain": "| **Search k=10** (10K/768D) | 47.0 µs | 21.3K QPS |",
      "validation_command": "cargo bench -p velesdb-core --bench hnsw_benchmark -- hnsw_search_latency",
      "executable": false,
      "source": "commit 126d5a9a (2026-03-27): 'Updated metrics (measured March 27, 2026, i9-14900KF, AVX2, Rust 1.92): HNSW Search (10K/768D, k=10): 54.6us -> 47.0us (-14%) ... Search QPS: 18.3K -> 21.3K'. measured_version is the workspace version of that commit.",
      "measured_on": "2026-03-27",
      "measured_machine": "Intel Core i9-14900KF, AVX2, Rust 1.92 (commit 126d5a9a)",
      "measured_version": "1.7.2",
      "stale_accepted": {
        "reason": "Dated in docs/BENCHMARKS.md section 5 (March 27, 2026, commit 126d5a9a); re-measuring belongs on the idle i9-14900KF reference machine.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "benchmarks_sparse_16x_speedup",
      "file": "docs/BENCHMARKS.md",
      "must_contain": "~16× speedup vs v1.12",
      "validation_command": "cargo bench -p velesdb-core --bench sparse_benchmark -- top10_10k_corpus",
      "executable": false,
      "source": "commit f4999a74 (PR #621, 2026-04-21): 'Measured impact (i9-14900K, SPLADE 10K corpus): top-10: 956 us -> 57.6 us (-94%, 16.6x speedup); top-100: 927 us -> 75.1 us (-92%, 12.3x speedup)'. measured_version is the workspace version of that commit.",
      "measured_on": "2026-04-21 (date of commit f4999a74, which records the run)",
      "measured_machine": "i9-14900K, as commit f4999a74 names it (docs/BENCHMARKS.md's reference machine is the i9-14900KF)",
      "measured_version": "1.13.0",
      "stale_accepted": {
        "reason": "Recorded by PR #621 with its before and after on one machine (1.13.0); re-measuring belongs on the idle reference machine, not done in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "core_perf_sparse_16x",
      "file": "docs/guides/CORE_PERFORMANCE.md",
      "must_contain": "**57.6 µs** (v1.13.0, PR #621 — 16x faster than v1.12)",
      "validation_command": "cargo bench -p velesdb-core --bench sparse_benchmark -- top10_10k_corpus",
      "executable": false,
      "source": "commit f4999a74 (PR #621, 2026-04-21): 'Measured impact (i9-14900K, SPLADE 10K corpus): top-10: 956 us -> 57.6 us (-94%, 16.6x speedup); top-100: 927 us -> 75.1 us (-92%, 12.3x speedup)'. measured_version is the workspace version of that commit.",
      "measured_on": "2026-04-21 (date of commit f4999a74, which records the run)",
      "measured_machine": "i9-14900K, as commit f4999a74 names it (docs/BENCHMARKS.md's reference machine is the i9-14900KF)",
      "measured_version": "1.13.0",
      "stale_accepted": {
        "reason": "Recorded by PR #621 with its before and after on one machine (1.13.0); re-measuring belongs on the idle reference machine, not done in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "benchmarks_graph_traversal_v2",
      "file": "docs/BENCHMARKS.md",
      "must_contain": "### Graph Traversal V2 — CSR Snapshot (Issue #491)",
      "covers_section": true,
      "validation_command": "cargo bench -p velesdb-core --bench graph_traversal_v2",
      "executable": false,
      "source": "docs/BENCHMARKS.md section 8 'Measured April 2026 with Criterion on the same test environment', recorded by commit 8939a728 (2026-04-03, 'update BENCHMARKS.md and CONCURRENCY_MODEL.md with perf optimization results (#487, #491, #488)'). measured_version is the workspace version of that commit.",
      "measured_on": "2026-04 (the section records the month only)",
      "measured_machine": "Intel Core i9-14900KF - 'the same test environment' as docs/BENCHMARKS.md's Test Environment",
      "measured_version": "1.11.0",
      "stale_accepted": {
        "reason": "Published with its month in docs/BENCHMARKS.md section 8; re-measuring belongs on the idle i9-14900KF reference machine.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "arch_diagrams_label_pushdown",
      "file": "docs/reference/ARCHITECTURE_DIAGRAMS.md",
      "must_contain": "290ns filtered BFS",
      "validation_command": "cargo bench -p velesdb-core --bench graph_traversal_v2",
      "executable": false,
      "source": "docs/BENCHMARKS.md section 8 'Measured April 2026 with Criterion on the same test environment', recorded by commit 8939a728 (2026-04-03, 'update BENCHMARKS.md and CONCURRENCY_MODEL.md with perf optimization results (#487, #491, #488)'). measured_version is the workspace version of that commit. The diagram quotes that section's 'Predicate pushdown 1/5 labels' row (290 ns).",
      "measured_on": "2026-04 (the section records the month only)",
      "measured_machine": "Intel Core i9-14900KF - 'the same test environment' as docs/BENCHMARKS.md's Test Environment",
      "measured_version": "1.11.0",
      "stale_accepted": {
        "reason": "Published with its month in docs/BENCHMARKS.md section 8; re-measuring belongs on the idle i9-14900KF reference machine.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "arch_diagrams_unfiltered_traversal",
      "file": "docs/reference/ARCHITECTURE_DIAGRAMS.md",
      "must_contain": "3.4µs unfiltered",
      "validation_command": "cargo bench -p velesdb-core --bench graph_traversal_v2",
      "executable": false,
      "source": "docs/BENCHMARKS.md section 8 'Measured April 2026 with Criterion on the same test environment', recorded by commit 8939a728 (2026-04-03, 'update BENCHMARKS.md and CONCURRENCY_MODEL.md with perf optimization results (#487, #491, #488)'). measured_version is the workspace version of that commit. The diagram quotes that section's 'No filter baseline' row (3.4 us).",
      "measured_on": "2026-04 (the section records the month only)",
      "measured_machine": "Intel Core i9-14900KF - 'the same test environment' as docs/BENCHMARKS.md's Test Environment",
      "measured_version": "1.11.0",
      "stale_accepted": {
        "reason": "Published with its month in docs/BENCHMARKS.md section 8; re-measuring belongs on the idle i9-14900KF reference machine.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "benchmarks_pq_full_precision_recall",
      "file": "docs/BENCHMARKS.md",
      "must_contain": "Full-precision recall@10 is 98.4% on this uniform-random setup",
      "validation_command": "cargo bench -p velesdb-core --bench pq_recall_benchmark -- full_precision_baseline",
      "executable": false,
      "source": "commit e35d9612 (2026-03-08, 'restore 0.92 recall thresholds with uniform random dataset'): moved pq_recall_benchmark from clustered to uniform-random data, restored the full-precision assert to recall >= 0.95, and recorded 'All 6 benchmark variants pass: PQ=0.994, Full=0.984, ...'. The 87.6% docs/BENCHMARKS.md showed before came from the clustered data of 7f05d0f0 (2026-03-07). measured_version is the workspace version of commit e35d9612.",
      "measured_on": "2026-03-08 (date of commit e35d9612, which records the run)",
      "measured_machine": "unknown (commit e35d9612 does not name it)",
      "measured_version": "1.5.0",
      "stale_accepted": {
        "reason": "A recall figure published with the commit and date of its run (e35d9612, 2026-03-08); re-running pq_recall_benchmark takes a cargo bench run on a quiet machine, not made in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "native_hnsw_vs_hnsw_rs_history",
      "file": "docs/reference/NATIVE_HNSW.md",
      "must_contain": "## Performance",
      "covers_section": true,
      "validation_command": "none: hnsw_rs left the tree in 1.0.0, so the comparison cannot be re-run",
      "executable": false,
      "source": "CHANGELOG.md [1.0.0] - 2026-01-08, the release that removed hnsw_rs: '1.2x faster search - 26.9ms vs 32.4ms (100 queries, 5K vectors)', '1.07x faster parallel insert - 1.47s vs 1.57s (5K vectors)', '~99% recall parity'; first recorded by commit 8387183e (2026-01-08, 'make native-hnsw default, hnsw_rs optional'). measured_version is the workspace version of that commit. The 'March 20, 2026 - i9-14900KF' stamp the section carried came from the 1.6.0 release commit and did not describe this run.",
      "measured_on": "2026-01-08",
      "measured_machine": "unknown (neither CHANGELOG 1.0.0 nor commit 8387183e names it)",
      "measured_version": "0.8.12",
      "stale_accepted": {
        "reason": "A historical comparison against hnsw_rs, which left the tree in 1.0.0: it cannot be re-run, and docs/reference/NATIVE_HNSW.md presents it as history, with its date.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "architecture_columnstore_130x",
      "file": "docs/reference/ARCHITECTURE.md",
      "must_contain": "Up to 130x faster than JSON filtering",
      "validation_command": "cargo bench -p velesdb-core --bench column_filter_benchmark",
      "executable": false,
      "source": "docs/BENCHMARKS.md section 6 '#### Integer Equality Filter (`filter_eq_int`, measured 2026-03-19)': JSON scan 3.84 ms vs ColumnStore 29.5 us at 100K rows = 130x, recorded by the v1.6.0 release commit 92cc35bd. readme_columnstore_130x_claim pins the same figure (and an Apple M5 Pro cross-check at ~50-105x).",
      "measured_on": "2026-03-19",
      "measured_machine": "Intel Core i9-14900KF (x86_64, AVX2) - docs/BENCHMARKS.md reference machine",
      "measured_version": "1.6.0",
      "stale_accepted": {
        "reason": "Published with its date in docs/BENCHMARKS.md section 6; the ratio is hardware-dependent and re-measuring it belongs on the idle i9-14900KF reference machine.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "server_tour_hnsw_index_only",
      "file": "docs/guides/SERVER_REST_TOUR.md",
      "must_contain": "~55 µs (10K vectors, 768D, Balanced mode, k=10), measured 2026-03-27 on the i9-14900KF",
      "validation_command": "cargo bench -p velesdb-core --bench hnsw_benchmark",
      "executable": false,
      "source": "commit df217860 (2026-03-27, 'docs: update README performance section with verified production metrics'): 'Core engine: Rust index-only for architecture reference (5K/768D: 55us HNSW search, 21.7ns dot product)' and 'Full production path: Python SDK, WAL ON, recall@10 >= 95% (10K/384D: 571us p50, 15K vec/s insert)'; 'Measured 2026-03-27, i9-14900KF, 64GB DDR5, Windows 11, Rust 1.92.' hnsw_benchmark's hnsw_search_latency group inserted 10,000 vectors at that commit, so the '5K' label was wrong; commit 45cb8021 relabeled it 10K and re-measured 53.2 us on an Apple M5 Pro (2026-07-20). measured_version is the workspace version of df217860.",
      "measured_on": "2026-03-27",
      "measured_machine": "Intel Core i9-14900KF, 64GB DDR5, Windows 11, Rust 1.92 (commit df217860)",
      "measured_version": "1.7.2",
      "stale_accepted": {
        "reason": "Dated by commit df217860 (2026-03-27, 1.7.2, i9-14900KF). The same day commit 126d5a9a recorded 47.0 us for the same bench on the same machine, and a quiet Apple M5 Pro re-run on 2026-07-20 measured 53.2 us; re-measuring belongs on the idle i9-14900KF, not done in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "server_tour_hnsw_index_only_m5_rerun",
      "file": "docs/guides/SERVER_REST_TOUR.md",
      "must_contain": "a quiet re-run on an Apple M5 Pro measured 53.2 µs (2026-07-20)",
      "validation_command": "cargo bench -p velesdb-core --bench hnsw_benchmark -- hnsw_search_latency",
      "executable": false,
      "source": "commit 45cb8021 (#1502, 2026-07-20, 're-verified numbers'): hnsw_benchmark's hnsw_search_latency group (10,000 x 768D, k=10) re-run on a quiet Apple M5 Pro at 53.2 us, recorded with the root README's index-only line (root_readme_hnsw_latency_microseconds). Another machine than the 55 us i9-14900KF figure it sits beside; neither replaces the other. measured_version is the workspace version of that commit.",
      "measured_on": "2026-07-20",
      "measured_machine": "Apple M5 Pro, quiet (commit 45cb8021)",
      "measured_version": "3.12.0",
      "stale_accepted": {
        "reason": "Recorded by commit 45cb8021 (3.12.0) on an Apple M5 Pro; re-running hnsw_benchmark on that machine at rest was not done in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "server_tour_simd_cosine",
      "file": "docs/guides/SERVER_REST_TOUR.md",
      "must_contain": "~33 ns per operation (768D; i9-14900KF, AVX2, in a run of 2026-03-24)",
      "validation_command": "cargo bench -p velesdb-core --bench simd_benchmark -- --noplot",
      "executable": false,
      "source": "commit dd0d2457 (2026-03-24, 'docs: final benchmark update - real measurements March 24, 2026 22:56'): 'All metrics re-measured sequentially on idle i9-14900KF, 64GB DDR5, Windows 11, Rust 1.92.0, Criterion.rs' - SIMD 768D: 'Cosine: 33.1 ns', 'Hamming: 35.8 ns', 'Jaccard: 35.1 ns'. measured_version is the workspace version of that commit. The line rounds its 33.1 ns to ~33 ns.",
      "measured_on": "2026-03-24",
      "measured_machine": "Intel Core i9-14900KF, 64GB DDR5, Windows 11, Rust 1.92.0, idle (commit dd0d2457)",
      "measured_version": "1.7.0",
      "stale_accepted": {
        "reason": "Dated by commit dd0d2457 (March 24, 2026, 1.7.0); re-measuring belongs on the idle i9-14900KF reference machine, not done in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "server_tour_simd_dot",
      "file": "docs/guides/SERVER_REST_TOUR.md",
      "must_contain": "~21.7 ns per operation (768D; same machine, in a run of 2026-03-27)",
      "validation_command": "cargo bench -p velesdb-core --bench simd_benchmark -- --noplot",
      "executable": false,
      "source": "commit df217860 (2026-03-27, 'docs: update README performance section with verified production metrics'): 'Core engine: Rust index-only for architecture reference (5K/768D: 55us HNSW search, 21.7ns dot product)'; 'Measured 2026-03-27, i9-14900KF, 64GB DDR5, Windows 11, Rust 1.92.' The April 3, 2026 run in docs/BENCHMARKS.md section 1 measured 21.8 ns for the same kernel. measured_version is the workspace version of df217860.",
      "measured_on": "2026-03-27",
      "measured_machine": "Intel Core i9-14900KF, 64GB DDR5, Windows 11, Rust 1.92 (commit df217860)",
      "measured_version": "1.7.2",
      "stale_accepted": {
        "reason": "Dated by commit df217860 (2026-03-27, 1.7.2); re-measuring belongs on the idle i9-14900KF reference machine, not done in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "rustdoc_memory_rebuild_205x",
      "file": "crates/velesdb-memory/src/migration/tests/performance.rs",
      "must_contain": "walk runs ~205x faster",
      "validation_command": "cargo test -p velesdb-memory --release -- --ignored batched_write_cost_stays_flat_as_the_volume_doubles",
      "executable": false,
      "source": "commit a4950038 (2026-08-03, 'requalify the rebuild throughput now that #1797 is merged'): 'Re-measured with the same code on the same machine, --release, batch 1 024: 16 000 facts go from 3345.8 us/fact to 16.3, i.e. 299 facts/s to 61 524 -- about 205x ... ratios 0.85 / 0.86 / 0.95 / 0.97'. measured_version is the workspace version of that commit (velesdb-memory 0.12.0).",
      "measured_on": "2026-08-03 (date of commit a4950038, which records the run)",
      "measured_machine": "unknown (commit a4950038 says 'the same machine' without naming it)",
      "measured_version": "4.2.0",
      "stale_accepted": {
        "reason": "The test's own recorded before/after (a4950038, 4.2.0 / velesdb-memory 0.12.0); re-running it writes 31 000 facts and, per its #[ignore], wants a machine at rest - not done in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "context_savings_stateless_compile_latency",
      "file": "crates/velesdb-memory/examples/context_savings/README.md",
      "must_contain": "compile latency (stateless: store_sources/record_events off): mean 0.7 ms, max 0.9 ms",
      "validation_command": "cd crates/velesdb-memory/examples/context_savings/real_measures && node agent_session.mjs",
      "executable": false,
      "source": "crates/velesdb-memory/examples/context_savings/README.md, agent-session output block (committed with its token figures by ebdbf015, 2026-07-26): 'compile latency (stateless: store_sources/record_events off): mean 0.7 ms, max 0.9 ms' - the same run as readme_economics_context_savings_corpus (82.5% saved).",
      "measured_on": "2026-07-26",
      "measured_machine": "unknown (latency depends on the machine, which the run does not record; its token figures are platform-independent)",
      "measured_version": "4.0.0",
      "stale_accepted": {
        "reason": "Printed by the same run as the 82.5% corpus figure (readme_economics_context_savings_corpus, 4.0.0) and published as that run's output; re-running agent_session.mjs needs the Node binding built from the current tree, not done in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "readme_long_session_window_growth",
      "file": "README.md",
      "must_contain": "compiled context grows **1.7× slower**",
      "validation_command": "cd examples/real-session-benchmark && node long-session.mjs",
      "executable": false,
      "source": "examples/real-session-benchmark/README.md, 'Long session (36 turns) - context-window headroom' (node long-session.mjs, recorded by 5b4df82a, 2026-07-18): the raw arm grows ~555 tokens/turn vs ~333 compiled over the full measured session, '1.7x slower window consumption'. measured_version is the workspace version of that commit.",
      "measured_on": "2026-07-18 (date of commit 5b4df82a, which records the run)",
      "measured_machine": "platform-independent (deterministic token counts; every turn compiled twice, byte-identical)",
      "measured_version": "3.12.0",
      "stale_accepted": {
        "reason": "Deterministic token counts published with their campaign in examples/real-session-benchmark; re-running long-session.mjs needs the Node binding built from the current tree, not done in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "examples_real_session_window_growth_1_7x",
      "file": "examples/real-session-benchmark/README.md",
      "must_contain": "(1.7× slower window consumption)",
      "validation_command": "cd examples/real-session-benchmark && node long-session.mjs",
      "executable": false,
      "source": "This README's long-session reading, written by commit 5b4df82a (2026-07-18) from node long-session.mjs: the raw arm grows '~555 tokens/turn versus ~333 compiled (1.7× slower window consumption)'. Same run as readme_long_session_window_growth. measured_version is the workspace version of that commit.",
      "measured_on": "2026-07-18 (date of commit 5b4df82a, which records the run)",
      "measured_machine": "platform-independent (deterministic token counts; every turn compiled twice, byte-identical)",
      "measured_version": "3.12.0",
      "stale_accepted": {
        "reason": "Deterministic token counts published with their campaign in examples/real-session-benchmark; re-running long-session.mjs needs the Node binding built from the current tree, not done in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "examples_real_session_tail_phase_6_6x",
      "file": "examples/real-session-benchmark/README.md",
      "must_contain": "the gap widens to ~234 vs ~35/turn (6.6×)",
      "validation_command": "cd examples/real-session-benchmark && node long-session.mjs",
      "executable": false,
      "source": "This README's long-session reading, written by commit 5b4df82a (2026-07-18) from node long-session.mjs: in the verification/wrap-up phase 'the gap widens to ~234 vs ~35/turn (6.6×)', a phase-specific rate never used for projections. Same run as context_compiler_tail_phase_6_6x. measured_version is the workspace version of that commit.",
      "measured_on": "2026-07-18 (date of commit 5b4df82a, which records the run)",
      "measured_machine": "platform-independent (deterministic token counts; every turn compiled twice, byte-identical)",
      "measured_version": "3.12.0",
      "stale_accepted": {
        "reason": "Deterministic token counts published with their campaign in examples/real-session-benchmark; re-running long-session.mjs needs the Node binding built from the current tree, not done in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "examples_real_session_no_media_saving_8pct",
      "file": "examples/real-session-benchmark/README.md",
      "must_contain": "saving drops to 8.0%",
      "validation_command": "cd examples/real-session-benchmark && BENCH_MEDIA=0 node offline-vibe.mjs",
      "executable": false,
      "source": "The README's vibe-coding scenario, written by commit fca21a88 (2026-07-19, 'feat(bench): vibe-coding billed scenario — media variants + metadata headroom'), which records the run's output: 'session totals [vibe-coding no-screenshots (BENCH_MEDIA=0)]: raw 51589 (text 51589 + image 0) -> compiled 47477 (...) = 8.0% saved'. measured_version is the workspace version of that commit.",
      "measured_on": "2026-07-19 (date of commit fca21a88, which records the run)",
      "measured_machine": "platform-independent (deterministic token counts)",
      "measured_version": "3.12.0",
      "stale_accepted": {
        "reason": "Deterministic counts published with their scenario in examples/real-session-benchmark; re-running offline-vibe.mjs needs the Node binding built from the current tree, not done in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "examples_real_session_metadata_headroom_14x",
      "file": "examples/real-session-benchmark/README.md",
      "must_contain": "roughly 14x headroom over the heaviest fragment measured",
      "validation_command": "cd examples/real-session-benchmark && node offline-vibe.mjs",
      "executable": false,
      "source": "The README's vibe-coding scenario, written by commit fca21a88 (2026-07-19): the heaviest fragment's metadata 'measures 4688 bytes — 7.15% of the 64 KiB cap', hence roughly 14x headroom. measured_version is the workspace version of that commit.",
      "measured_on": "2026-07-19 (date of commit fca21a88, which records the run)",
      "measured_machine": "platform-independent (deterministic byte counts)",
      "measured_version": "3.12.0",
      "stale_accepted": {
        "reason": "Deterministic counts published with their scenario in examples/real-session-benchmark; re-running offline-vibe.mjs needs the Node binding built from the current tree, not done in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "rustdoc_sparse_linear_scan_vs_maxscore",
      "file": "crates/velesdb-core/src/sparse_index/search/mod.rs",
      "must_contain": "measured 14–43× faster than `MaxScore` at the former 100k",
      "validation_command": "cargo bench -p velesdb-core --bench sparse_posting_scale",
      "executable": false,
      "source": "commit d4e29e17 (#2182, 2026-09-01): on the sparse_posting_scale bench (30k vocab, two passes per configuration) MaxScore query latency at 110k docs was 43x the linear arm under uniform queries and 14x under skewed SPLADE-shaped ones, flat at ~13x in a crossover sweep up to 400k docs; full evidence on #2177. measured_version is the workspace version of that commit.",
      "measured_on": "2026-09-01 (date of commit d4e29e17, which records the run)",
      "measured_machine": "unknown (commit d4e29e17 does not name it)",
      "measured_version": "5.2.0",
      "stale_accepted": {
        "reason": "Recorded with the change it justified (d4e29e17, on 5.2.0), and the comment names #2177 as its source; re-running sparse_posting_scale needs a quiet machine, not made in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "benchmarks_sift1m_recall_floor",
      "file": "docs/BENCHMARKS.md",
      "must_contain": "**Recall@10 < 0.90 at ef=128**",
      "validation_command": "cargo bench -p velesdb-core --bench sift1m_recall --features bench-sift1m",
      "executable": false,
      "source": ".github/workflows/bench-sift1m-nightly.yml gates SIFT1M_RECALL_MIN 0.90 at ef=128 ('SIFT1M Recall@10 >= 0.90 (ef=128)'); the line states the floor that gate enforces.",
      "notes": "A threshold, not a measurement: the figure is the floor or bound named in source.",
      "measured_on": "n/a",
      "measured_machine": "n/a",
      "measured_version": "n/a",
      "last_updated": "2026-09-13"
    },
    {
      "id": "benchmarks_sift1m_latency_review_threshold",
      "file": "docs/BENCHMARKS.md",
      "must_contain": "**p50 latency > 1 ms at ef=128 on reference hardware**",
      "validation_command": "cargo bench -p velesdb-core --bench sift1m_recall --features bench-sift1m",
      "executable": false,
      "source": "A review threshold stated in docs/BENCHMARKS.md section 11.5; section 11.6 records that latency is reported, not gated.",
      "notes": "A threshold, not a measurement: the figure is the floor or bound named in source.",
      "measured_on": "n/a",
      "measured_machine": "n/a",
      "measured_version": "n/a",
      "last_updated": "2026-09-13"
    },
    {
      "id": "core_tests_readme_search_bound_row",
      "file": "crates/velesdb-core/tests/README.md",
      "must_contain": "verify search <100ms",
      "validation_command": "cargo test -p velesdb-core --test integration_scenarios test_batch_product_indexing_performance",
      "executable": false,
      "source": "crates/velesdb-core/tests/integration_scenarios.rs asserts duration.as_millis() < 100 in test_batch_product_indexing_performance; the README states the bound the test asserts.",
      "notes": "A threshold, not a measurement: the figure is the floor or bound named in source.",
      "measured_on": "n/a",
      "measured_machine": "n/a",
      "measured_version": "n/a",
      "last_updated": "2026-09-13"
    },
    {
      "id": "core_tests_readme_search_bound_note",
      "file": "crates/velesdb-core/tests/README.md",
      "must_contain": "Search over 1000 products completes in <100ms",
      "validation_command": "cargo test -p velesdb-core --test integration_scenarios test_batch_product_indexing_performance",
      "executable": false,
      "source": "crates/velesdb-core/tests/integration_scenarios.rs asserts duration.as_millis() < 100 in test_batch_product_indexing_performance; the README states the bound the test asserts.",
      "notes": "A threshold, not a measurement: the figure is the floor or bound named in source.",
      "measured_on": "n/a",
      "measured_machine": "n/a",
      "measured_version": "n/a",
      "last_updated": "2026-09-13"
    },
    {
      "id": "planning_parity_recall_gate",
      "file": "docs/planning/CORE_PARITY_REMEDIATION.md",
      "must_contain": "QUALITY_BAR Gate 1 recall@10 ≥ 0.95",
      "validation_command": "cargo test -p velesdb-core --features persistence test_recall -- --test-threads=1",
      "executable": false,
      "source": "QUALITY_BAR.md Gate 1 - Recall@10 >= 0.95 on 10K vectors; the planning row cites that gate's floor.",
      "notes": "A threshold, not a measurement: the figure is the floor or bound named in source.",
      "measured_on": "n/a",
      "measured_machine": "n/a",
      "measured_version": "n/a",
      "last_updated": "2026-09-13"
    },
    {
      "id": "rustdoc_precision_contract_recall_floor",
      "file": "crates/velesdb-core/src/index/hnsw/native/precision_contract_tests.rs",
      "must_contain": "recall@10 >= 0.95 on 10K",
      "validation_command": "cargo test -p velesdb-core --lib precision_recall_above_threshold",
      "executable": false,
      "source": "The doc comment of check_recall_10k states the floor the function asserts (avg_recall >= 0.95); each codec's precision test runs it.",
      "notes": "A threshold, not a measurement: the figure is the floor or bound named in source.",
      "measured_on": "n/a",
      "measured_machine": "n/a",
      "measured_version": "n/a",
      "last_updated": "2026-09-13"
    },
    {
      "id": "rustdoc_rabitq_recall_floor",
      "file": "crates/velesdb-core/src/index/hnsw/native/rabitq_precision_tests.rs",
      "must_contain": "Verifies recall@10 >= 0.95 on 10K vectors with `RaBitQ` traversal",
      "validation_command": "cargo test -p velesdb-core --lib test_rabitq_precision_recall_above_threshold",
      "executable": false,
      "source": "test_rabitq_precision_recall_above_threshold runs suite::check_recall_10k::<RaBitQCodec>(), which asserts avg_recall >= 0.95.",
      "notes": "A threshold, not a measurement: the figure is the floor or bound named in source.",
      "measured_on": "n/a",
      "measured_machine": "n/a",
      "measured_version": "n/a",
      "last_updated": "2026-09-13"
    },
    {
      "id": "rustdoc_sq8_recall_floor",
      "file": "crates/velesdb-core/src/index/hnsw/native/sq8_precision_tests.rs",
      "must_contain": "Verifies recall@10 >= 0.95 on 10K vectors with int8 traversal",
      "validation_command": "cargo test -p velesdb-core --lib test_sq8_precision_recall_above_threshold",
      "executable": false,
      "source": "test_sq8_precision_recall_above_threshold runs suite::check_recall_10k::<Sq8Codec>(), which asserts avg_recall >= 0.95.",
      "notes": "A threshold, not a measurement: the figure is the floor or bound named in source.",
      "measured_on": "n/a",
      "measured_machine": "n/a",
      "measured_version": "n/a",
      "last_updated": "2026-09-13"
    },
    {
      "id": "rustdoc_search_pipeline_recall_floor",
      "file": "crates/velesdb-core/src/index/hnsw/native/search_pipeline_tests.rs",
      "must_contain": "Both should achieve >= 95% recall",
      "validation_command": "cargo test -p velesdb-core --lib test_pipelined_recall_above_threshold",
      "executable": false,
      "source": "test_pipelined_recall_above_threshold asserts avg_recall >= 0.95 against brute-force ground truth.",
      "notes": "A threshold, not a measurement: the figure is the floor or bound named in source.",
      "measured_on": "n/a",
      "measured_machine": "n/a",
      "measured_version": "n/a",
      "last_updated": "2026-09-13"
    },
    {
      "id": "readme_context_savings_compile_latency",
      "file": "README.md",
      "must_contain": "compiled in ~0.7 ms mean stateless (~24.5 ms with source persistence on)",
      "validation_command": "cd crates/velesdb-memory/examples/context_savings/real_measures && node agent_session.mjs",
      "executable": false,
      "source": "crates/velesdb-memory/examples/context_savings/README.md, agent-session output block (committed with its token figures by ebdbf015, 2026-07-26): 'compile latency (with source/event persistence, default): mean 24.5 ms, max 37.2 ms' and 'compile latency (stateless: ...): mean 0.7 ms' - the same run as readme_economics_context_savings_corpus (82.5% saved). The README carried ~0.5 ms / ~27 ms from the earlier dca8a002 run (2026-07-17) until this change aligned it with the output it links to.",
      "measured_on": "2026-07-26",
      "measured_machine": "unknown (latency depends on the machine, which the run does not record; its token figures are platform-independent)",
      "measured_version": "4.0.0",
      "stale_accepted": {
        "reason": "Printed by the same run as the 82.5% corpus figure (4.0.0) and published as that run's output; re-running agent_session.mjs needs the Node binding built from the current tree, not done in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "readme_production_search_recall",
      "file": "README.md",
      "must_contain": "(10K/384D, recall ≥ 96 %), measured 2026-03-27 on 1.7.2",
      "validation_command": "python benchmarks/velesdb_benchmark.py --recall",
      "executable": false,
      "source": "benchmarks/report_1.7.2_2026-03-27.json, written by benchmarks/velesdb_benchmark.py --recall (Python SDK, WAL ON, 10K x 384D): recall@10 mean 0.976 (min 0.8) on its clustered 10K x 384D recall set; CHANGELOG.md [1.8.0] (2026-03-27) publishes the run with 'recall@10 >= 96%'. The same run as the 450 us p50.",
      "measured_on": "2026-03-27",
      "measured_machine": "Intel Core i9-14900KF, 64 GB, Windows (benchmarks/report_1.7.2_2026-03-27.json)",
      "measured_version": "1.7.2",
      "stale_accepted": {
        "reason": "Published with its original date and version (2026-03-27, 1.7.2, Balanced then at ef 128). Re-measuring means re-running velesdb_benchmark.py on the idle i9-14900KF it was taken on.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "readme_recall_balanced_index_only",
      "file": "README.md",
      "must_contain": "Recall@10 balanced 99.8%",
      "validation_command": "cargo bench -p velesdb-core --bench recall_benchmark",
      "executable": false,
      "source": "cargo bench -p velesdb-core --bench recall_benchmark: its recall report (10K random 128-D vectors, HnswParams::max_recall index, 100 queries, k=10) at the current presets, 2026-09-10 on 6.0.0 (Balanced ef=160: 99.8%); mirrors readme_recall_balanced_mode.",
      "measured_on": "2026-09-10",
      "measured_machine": "Apple M5 Pro",
      "measured_version": "6.0.0",
      "last_updated": "2026-09-13"
    },
    {
      "id": "readme_columnstore_apple_json_scan_ratio",
      "file": "README.md",
      "must_contain": "the JSON scan itself runs ~2.8× faster",
      "validation_command": "cargo bench -p velesdb-core --bench column_filter_benchmark",
      "executable": false,
      "source": "commit 45cb8021 (2026-07-20, #1502): 'Numbers were re-verified today by a sequential quiet-machine re-bench; ... the ColumnStore 130x ratio is hardware-qualified (i9 reference; ~50-105x on Apple Silicon where the JSON scan itself is faster)'. The README sentence the commit wrote names the machine and the ~2.8x. measured_version is the workspace version of that commit.",
      "measured_on": "2026-07-20",
      "measured_machine": "Apple M5 Pro (named by the README sentence commit 45cb8021 wrote; the commit says 'Apple Silicon')",
      "measured_version": "3.12.0",
      "stale_accepted": {
        "reason": "Recorded by commit 45cb8021 (3.12.0) as a cross-check of the i9 ratio; re-measuring needs the same Apple M5 Pro at rest, not done in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "core_readme_e2e_recall",
      "file": "crates/velesdb-core/README.md",
      "must_contain": "384D, WAL on, recall ≥ 96%",
      "validation_command": "python benchmarks/velesdb_benchmark.py --recall",
      "executable": false,
      "source": "benchmarks/report_1.7.2_2026-03-27.json, written by benchmarks/velesdb_benchmark.py --recall (Python SDK, WAL ON, 10K x 384D): recall@10 mean 0.976 (min 0.8) on its clustered 10K x 384D recall set; CHANGELOG.md [1.8.0] (2026-03-27) publishes the run with 'recall@10 >= 96%'. The same run as the 450 us p50.",
      "measured_on": "2026-03-27",
      "measured_machine": "Intel Core i9-14900KF, 64 GB, Windows (benchmarks/report_1.7.2_2026-03-27.json)",
      "measured_version": "1.7.2",
      "stale_accepted": {
        "reason": "Published with its original date and version (2026-03-27, 1.7.2, Balanced then at ef 128). Re-measuring means re-running velesdb_benchmark.py on the idle i9-14900KF it was taken on.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "benchmarks_simd_dot_throughput",
      "file": "docs/BENCHMARKS.md",
      "must_contain": "### Throughput",
      "covers_section": true,
      "validation_command": "cargo bench -p velesdb-core --bench simd_benchmark -- --noplot",
      "executable": false,
      "source": "docs/BENCHMARKS.md section 1: 'Measured April 3, 2026 on Intel Core i9-14900KF (24C/32T, AVX2+FMA), 64GB DDR5, Rust 1.94.1, Windows 11 Pro, sequential run on idle machine', set with this table (768D 21.8 ns / 35.2 Gelem/s, 1536D 61.6 ns / 24.9, 3072D 94.8 ns / 32.4) by commit 81b371ba (2026-04-03). Gelem/s is the dimension divided by the latency. measured_version is the workspace version of that commit.",
      "measured_on": "2026-04-03",
      "measured_machine": "Intel Core i9-14900KF (24C/32T, AVX2+FMA), 64GB DDR5, Rust 1.94.1, Windows 11 Pro, idle",
      "measured_version": "1.11.0",
      "stale_accepted": {
        "reason": "Dated in docs/BENCHMARKS.md section 1 (April 3, 2026, commit 81b371ba); re-measuring belongs on the idle i9-14900KF reference machine.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "benchmarks_sparse_insert_sequential",
      "file": "docs/BENCHMARKS.md",
      "must_contain": "| `sequential_10k` | 93 ms | 9.3 µs/doc |",
      "validation_command": "cargo bench -p velesdb-core --bench sparse_benchmark -- sequential_10k",
      "executable": false,
      "source": "the v1.6.0 release commit 92cc35bd (2026-03-20): 'Re-run all 6 benchmark suites sequentially on idle machine (i9-14900KF, ...)', 'All metrics now match sequential benchmark run on idle i9-14900KF'. It set this row; docs/BENCHMARKS.md then read 'Last updated: March 20, 2026 (v1.6.0+simd-opt — sequential benchmarks on idle machine)'.",
      "measured_on": "2026-03-20 (release commit 92cc35bd; the row carries no date of its own)",
      "measured_machine": "Intel Core i9-14900KF, idle (release commit 92cc35bd)",
      "measured_version": "1.6.0",
      "stale_accepted": {
        "reason": "Published by the v1.6.0 release run; re-measuring belongs on the idle i9-14900KF reference machine, not done in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "benchmarks_sparse_concurrent_16_threads",
      "file": "docs/BENCHMARKS.md",
      "must_contain": "| `16_threads_8_insert_8_search` | 171 ms |",
      "validation_command": "cargo bench -p velesdb-core --bench sparse_benchmark -- 16_threads_8_insert_8_search",
      "executable": false,
      "source": "the v1.6.0 release commit 92cc35bd (2026-03-20): 'Re-run all 6 benchmark suites sequentially on idle machine (i9-14900KF, ...)', 'All metrics now match sequential benchmark run on idle i9-14900KF'. It set this row; docs/BENCHMARKS.md then read 'Last updated: March 20, 2026 (v1.6.0+simd-opt — sequential benchmarks on idle machine)'.",
      "measured_on": "2026-03-20 (release commit 92cc35bd; the row carries no date of its own)",
      "measured_machine": "Intel Core i9-14900KF, idle (release commit 92cc35bd)",
      "measured_version": "1.6.0",
      "stale_accepted": {
        "reason": "Published by the v1.6.0 release run; re-measuring belongs on the idle i9-14900KF reference machine, not done in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "benchmarks_hybrid_dense_component",
      "file": "docs/BENCHMARKS.md",
      "must_contain": "| Dense HNSW search (k=10, 768D) | ~47 µs |",
      "validation_command": "cargo bench -p velesdb-core --bench hnsw_benchmark -- hnsw_search_latency",
      "executable": false,
      "source": "commit 126d5a9a (2026-03-27) changed this row from ~55 to ~47 us with 'Updated metrics (measured March 27, 2026, i9-14900KF, AVX2, Rust 1.92): HNSW Search (10K/768D, k=10): 54.6us -> 47.0us'; mirrors benchmarks_hnsw_search_k10_qps.",
      "measured_on": "2026-03-27",
      "measured_machine": "Intel Core i9-14900KF, AVX2, Rust 1.92 (commit 126d5a9a)",
      "measured_version": "1.7.2",
      "stale_accepted": {
        "reason": "Dated by commit 126d5a9a (March 27, 2026); re-measuring belongs on the idle i9-14900KF reference machine.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "benchmarks_hybrid_sparse_component",
      "file": "docs/BENCHMARKS.md",
      "must_contain": "| Sparse search (top-10, 10K) | ~57.6 µs |",
      "validation_command": "cargo bench -p velesdb-core --bench sparse_benchmark -- top10_10k_corpus",
      "executable": false,
      "source": "commit f4999a74 (PR #621, 2026-04-21): 'Measured impact (i9-14900K, SPLADE 10K corpus): top-10: 956 us -> 57.6 us (-94%, 16.6x speedup); top-100: 927 us -> 75.1 us (-92%, 12.3x speedup)'. measured_version is the workspace version of that commit.",
      "measured_on": "2026-04-21 (date of commit f4999a74, which records the run)",
      "measured_machine": "i9-14900K, as commit f4999a74 names it (docs/BENCHMARKS.md's reference machine is the i9-14900KF)",
      "measured_version": "1.13.0",
      "stale_accepted": {
        "reason": "Recorded by PR #621 with its before and after on one machine (1.13.0); re-measuring belongs on the idle reference machine, not done in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "benchmarks_hnsw_search_k50",
      "file": "docs/BENCHMARKS.md",
      "must_contain": "| **Search k=50** | 63.5 µs |",
      "validation_command": "cargo bench -p velesdb-core --bench hnsw_benchmark -- hnsw_search_latency",
      "executable": false,
      "source": "the v1.6.0 release commit 92cc35bd (2026-03-20): 'Re-run all 6 benchmark suites sequentially on idle machine (i9-14900KF, ...)', 'All metrics now match sequential benchmark run on idle i9-14900KF'. It set this row; docs/BENCHMARKS.md then read 'Last updated: March 20, 2026 (v1.6.0+simd-opt — sequential benchmarks on idle machine)'. pr363_365_comparison.md section 3 measured it higher three days later.",
      "measured_on": "2026-03-20 (release commit 92cc35bd; the row carries no date of its own)",
      "measured_machine": "Intel Core i9-14900KF, idle (release commit 92cc35bd)",
      "measured_version": "1.6.0",
      "stale_accepted": {
        "reason": "Published by the v1.6.0 release run; re-measuring belongs on the idle i9-14900KF reference machine, not done in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "benchmarks_hnsw_search_k100",
      "file": "docs/BENCHMARKS.md",
      "must_contain": "| **Search k=100** | 151.5 µs |",
      "validation_command": "cargo bench -p velesdb-core --bench hnsw_benchmark -- hnsw_search_latency",
      "executable": false,
      "source": "the v1.6.0 release commit 92cc35bd (2026-03-20): 'Re-run all 6 benchmark suites sequentially on idle machine (i9-14900KF, ...)', 'All metrics now match sequential benchmark run on idle i9-14900KF'. It set this row; docs/BENCHMARKS.md then read 'Last updated: March 20, 2026 (v1.6.0+simd-opt — sequential benchmarks on idle machine)'. pr363_365_comparison.md section 3 measured it higher three days later, with high variance.",
      "measured_on": "2026-03-20 (release commit 92cc35bd; the row carries no date of its own)",
      "measured_machine": "Intel Core i9-14900KF, idle (release commit 92cc35bd)",
      "measured_version": "1.6.0",
      "stale_accepted": {
        "reason": "Published by the v1.6.0 release run; re-measuring belongs on the idle i9-14900KF reference machine, not done in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "benchmarks_hnsw_insert_sequential_1k",
      "file": "docs/BENCHMARKS.md",
      "must_contain": "| **Insert 1K x 768D** (sequential) | 299.79 ms | 3.3K vec/s |",
      "validation_command": "cargo bench -p velesdb-core --bench hnsw_benchmark -- hnsw_insert/",
      "executable": false,
      "source": "benchmarks/results/pr363_365_comparison.md, 'Current' column (commit 4eb1c1f0, after PR #363 + #365), clean re-run on a quiet machine, Criterion 100 samples: section 1 Sequential 1K×768D 299.79 ms (throughput summary: 3,336 vec/s). The v1.6.0 rows it replaces (263.7 ms, 156.5 ms, 2.26 s) measured the parallel path before #363's batch-insert pipeline; the report's Baseline (a0f5926e, a descendant of v1.6.0) is in the same regime (185.76 ms, 2,122 ms). measured_version is the workspace version of 4eb1c1f0.",
      "measured_on": "2026-03-23",
      "measured_machine": "Intel Core i9-14900KF, 64GB, Windows 11 Pro, quiet machine (benchmarks/results/pr363_365_comparison.md)",
      "measured_version": "1.6.0",
      "stale_accepted": {
        "reason": "Published with its date and version (2026-03-23, 1.6.0) and not re-measured since; commit fc1e3fe1 (#2271) reports build time up 31 % at 10 000 vectors after it. Re-measuring belongs on the idle i9-14900KF reference machine, not done in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "benchmarks_hnsw_insert_parallel_1k",
      "file": "docs/BENCHMARKS.md",
      "must_contain": "| **Parallel Insert 1K x 768D** | 25.01 ms | 40.0K vec/s |",
      "validation_command": "cargo bench -p velesdb-core --bench hnsw_benchmark -- hnsw_insert_parallel",
      "executable": false,
      "source": "benchmarks/results/pr363_365_comparison.md, 'Current' column (commit 4eb1c1f0, after PR #363 + #365), clean re-run on a quiet machine, Criterion 100 samples: section 2 Parallel 1K×768D 25.01 ms (throughput summary: 39,989 vec/s). The v1.6.0 rows it replaces (263.7 ms, 156.5 ms, 2.26 s) measured the parallel path before #363's batch-insert pipeline; the report's Baseline (a0f5926e, a descendant of v1.6.0) is in the same regime (185.76 ms, 2,122 ms). measured_version is the workspace version of 4eb1c1f0.",
      "measured_on": "2026-03-23",
      "measured_machine": "Intel Core i9-14900KF, 64GB, Windows 11 Pro, quiet machine (benchmarks/results/pr363_365_comparison.md)",
      "measured_version": "1.6.0",
      "stale_accepted": {
        "reason": "Published with its date and version (2026-03-23, 1.6.0) and not re-measured since; commit fc1e3fe1 (#2271) reports build time up 31 % at 10 000 vectors after it. Re-measuring belongs on the idle i9-14900KF reference machine, not done in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "benchmarks_hnsw_insert_parallel_10k",
      "file": "docs/BENCHMARKS.md",
      "must_contain": "| **Parallel Insert 10K x 768D** | 281.23 ms | 35.6K vec/s |",
      "validation_command": "cargo bench -p velesdb-core --bench hnsw_benchmark -- hnsw_insert_parallel",
      "executable": false,
      "source": "benchmarks/results/pr363_365_comparison.md, 'Current' column (commit 4eb1c1f0, after PR #363 + #365), clean re-run on a quiet machine, Criterion 100 samples: section 2 Parallel 10K×768D 281.23 ms (10 000 / 0.28123 s = 35.6K vec/s). The v1.6.0 rows it replaces (263.7 ms, 156.5 ms, 2.26 s) measured the parallel path before #363's batch-insert pipeline; the report's Baseline (a0f5926e, a descendant of v1.6.0) is in the same regime (185.76 ms, 2,122 ms). measured_version is the workspace version of 4eb1c1f0.",
      "measured_on": "2026-03-23",
      "measured_machine": "Intel Core i9-14900KF, 64GB, Windows 11 Pro, quiet machine (benchmarks/results/pr363_365_comparison.md)",
      "measured_version": "1.6.0",
      "stale_accepted": {
        "reason": "Published with its date and version (2026-03-23, 1.6.0) and not re-measured since; commit fc1e3fe1 (#2271) reports build time up 31 % at 10 000 vectors after it. Re-measuring belongs on the idle i9-14900KF reference machine, not done in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "benchmarks_balanced_recall_design_target_figure",
      "file": "docs/BENCHMARKS.md",
      "must_contain": "Recall@10 >= 95% is the design target for Balanced mode and above",
      "validation_command": "cargo bench -p velesdb-core --bench recall_benchmark",
      "executable": false,
      "source": "A design target stated in docs/BENCHMARKS.md section 5; the same line names the runs that measure Balanced (99.8%) and Accurate at 1M (0.98).",
      "notes": "A target, not a measurement: the figure is the goal named in source.",
      "measured_on": "n/a",
      "measured_machine": "n/a",
      "measured_version": "n/a",
      "last_updated": "2026-09-13"
    },
    {
      "id": "benchmarks_upsert_path_v172",
      "file": "docs/BENCHMARKS.md",
      "must_contain": "upsert throughput ~808 vec/s before, ~16,151 vec/s after",
      "validation_command": "cargo bench -p velesdb-core --bench bulk_insert_v2_benchmark",
      "executable": false,
      "source": "docs/BENCHMARKS.md 'Upsert Path Optimization (v1.7.2)': 'Local measurement (i9-14900KF, 10K vectors, 384D): upsert throughput ~808 vec/s before, ~16,151 vec/s after', written by commit 024e4a9b (2026-03-25, 'docs: complete v1.7.2 documentation with upsert lock contention fix'). measured_version is the workspace version of that commit.",
      "measured_on": "2026-03-25 (date of commit 024e4a9b, which records the run; the line states no date)",
      "measured_machine": "Intel Core i9-14900KF (as the line names it)",
      "measured_version": "1.7.2",
      "stale_accepted": {
        "reason": "Recorded with its version in the section heading (v1.7.2); re-measuring belongs on the idle i9-14900KF reference machine, not done in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "benchmarks_bulk_insert_v2",
      "file": "docs/BENCHMARKS.md",
      "must_contain": "### Bulk Insert V2 (Issue #488)",
      "covers_section": true,
      "validation_command": "cargo bench -p velesdb-core --bench bulk_insert_v2_benchmark",
      "executable": false,
      "source": "commit 8939a728 (2026-04-03, 'docs: update BENCHMARKS.md and CONCURRENCY_MODEL.md with perf optimization results (#487, #491, #488)') added the section with its figures. bulk_insert_v2_benchmark inserts 10,000 vectors of 16 dimensions (DIMENSION = 16 at that commit and today). measured_version is the workspace version of that commit.",
      "measured_on": "2026-04 (commit 8939a728, 2026-04-03; the section states no date)",
      "measured_machine": "unknown (neither the section nor commit 8939a728 names it; the Graph Traversal V2 section the same commit added says 'the same test environment', the i9-14900KF)",
      "measured_version": "1.11.0",
      "stale_accepted": {
        "reason": "Recorded by commit 8939a728 (1.11.0); re-measuring belongs on the idle i9-14900KF reference machine, not done in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "benchmarks_sift1m_nightly_gate_floor",
      "file": "docs/BENCHMARKS.md",
      "must_contain": "gates Recall@10 ≥ 0.90 at ef=128",
      "validation_command": "cargo bench -p velesdb-core --bench sift1m_recall --features bench-sift1m",
      "executable": false,
      "source": ".github/workflows/bench-sift1m-nightly.yml gates Recall@10 >= 0.90 at ef=128 (section 11.6); the figure is the gate, not a result.",
      "notes": "A threshold, not a measurement: the figure is the floor or bound named in source.",
      "measured_on": "n/a",
      "measured_machine": "n/a",
      "measured_version": "n/a",
      "last_updated": "2026-09-13"
    },
    {
      "id": "benchmarking_guide_arm64_faster_threshold",
      "file": "docs/contributing/BENCHMARKING_GUIDE.md",
      "must_contain": "| ARM64 >10% faster |",
      "validation_command": "pwsh scripts/compare-arch-benchmarks.ps1 -x86ResultsPath <x86> -arm64ResultsPath <arm64>",
      "executable": false,
      "source": "scripts/compare-arch-benchmarks.ps1 classifies a ratio below 0.9 as 'ARM64 FASTER'.",
      "notes": "A threshold, not a measurement: the figure is the floor or bound named in source.",
      "measured_on": "n/a",
      "measured_machine": "n/a",
      "measured_version": "n/a",
      "last_updated": "2026-09-13"
    },
    {
      "id": "benchmarking_guide_x86_faster_threshold",
      "file": "docs/contributing/BENCHMARKING_GUIDE.md",
      "must_contain": "| x86_64 >10% faster |",
      "validation_command": "pwsh scripts/compare-arch-benchmarks.ps1 -x86ResultsPath <x86> -arm64ResultsPath <arm64>",
      "executable": false,
      "source": "scripts/compare-arch-benchmarks.ps1 classifies a ratio above 1.1 as 'x86 FASTER'.",
      "notes": "A threshold, not a measurement: the figure is the floor or bound named in source.",
      "measured_on": "n/a",
      "measured_machine": "n/a",
      "measured_version": "n/a",
      "last_updated": "2026-09-13"
    },
    {
      "id": "core_perf_hnsw_index_only_55us",
      "file": "docs/guides/CORE_PERFORMANCE.md",
      "must_contain": "**55 µs** (k=10, Balanced mode; i9-14900KF, 2026-03-27",
      "validation_command": "cargo bench -p velesdb-core --bench hnsw_benchmark -- hnsw_search_latency",
      "executable": false,
      "source": "commit df217860 (2026-03-27, 'docs: update README performance section with verified production metrics'): 'Core engine: Rust index-only for architecture reference (5K/768D: 55us HNSW search, 21.7ns dot product)' and 'Full production path: Python SDK, WAL ON, recall@10 >= 95% (10K/384D: 571us p50, 15K vec/s insert)'; 'Measured 2026-03-27, i9-14900KF, 64GB DDR5, Windows 11, Rust 1.92.' hnsw_benchmark's hnsw_search_latency group inserted 10,000 vectors at that commit, so the '5K' label was wrong; commit 45cb8021 relabeled it 10K and re-measured 53.2 us on an Apple M5 Pro (2026-07-20). measured_version is the workspace version of df217860.",
      "measured_on": "2026-03-27",
      "measured_machine": "Intel Core i9-14900KF, 64GB DDR5, Windows 11, Rust 1.92 (commit df217860)",
      "measured_version": "1.7.2",
      "stale_accepted": {
        "reason": "Dated by commit df217860 (2026-03-27, 1.7.2, i9-14900KF). The same day commit 126d5a9a recorded 47.0 us for the same bench on the same machine, and a quiet Apple M5 Pro re-run on 2026-07-20 measured 53.2 us; re-measuring belongs on the idle i9-14900KF, not done in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "examples_ecommerce_hnsw_index_only_55us",
      "file": "examples/ecommerce_recommendation/README.md",
      "must_contain": "**~55µs** for 10K/768D vectors (k=10, Balanced mode; i9-14900KF, 2026-03-27)",
      "validation_command": "cargo bench -p velesdb-core --bench hnsw_benchmark -- hnsw_search_latency",
      "executable": false,
      "source": "commit df217860 (2026-03-27, 'docs: update README performance section with verified production metrics'): 'Core engine: Rust index-only for architecture reference (5K/768D: 55us HNSW search, 21.7ns dot product)' and 'Full production path: Python SDK, WAL ON, recall@10 >= 95% (10K/384D: 571us p50, 15K vec/s insert)'; 'Measured 2026-03-27, i9-14900KF, 64GB DDR5, Windows 11, Rust 1.92.' hnsw_benchmark's hnsw_search_latency group inserted 10,000 vectors at that commit, so the '5K' label was wrong; commit 45cb8021 relabeled it 10K and re-measured 53.2 us on an Apple M5 Pro (2026-07-20). measured_version is the workspace version of df217860.",
      "measured_on": "2026-03-27",
      "measured_machine": "Intel Core i9-14900KF, 64GB DDR5, Windows 11, Rust 1.92 (commit df217860)",
      "measured_version": "1.7.2",
      "stale_accepted": {
        "reason": "Dated by commit df217860 (2026-03-27, 1.7.2, i9-14900KF). The same day commit 126d5a9a recorded 47.0 us for the same bench on the same machine, and a quiet Apple M5 Pro re-run on 2026-07-20 measured 53.2 us; re-measuring belongs on the idle i9-14900KF, not done in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "core_perf_hnsw_index_only_m5_rerun",
      "file": "docs/guides/CORE_PERFORMANCE.md",
      "must_contain": "53.2 µs in a quiet Apple M5 Pro re-run, 2026-07-20",
      "validation_command": "cargo bench -p velesdb-core --bench hnsw_benchmark -- hnsw_search_latency",
      "executable": false,
      "source": "commit 45cb8021 (#1502, 2026-07-20, 're-verified numbers'): hnsw_benchmark's hnsw_search_latency group (10,000 x 768D, k=10) re-run on a quiet Apple M5 Pro at 53.2 us, recorded with the root README's index-only line (root_readme_hnsw_latency_microseconds). Another machine than the 55 us i9-14900KF figure it sits beside; neither replaces the other. measured_version is the workspace version of that commit.",
      "measured_on": "2026-07-20",
      "measured_machine": "Apple M5 Pro, quiet (commit 45cb8021)",
      "measured_version": "3.12.0",
      "stale_accepted": {
        "reason": "Recorded by commit 45cb8021 (3.12.0) on an Apple M5 Pro; re-running hnsw_benchmark on that machine at rest was not done in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "core_perf_hnsw_index_only_bullet",
      "file": "docs/guides/CORE_PERFORMANCE.md",
      "must_contain": "- HNSW index-only micro-benchmark: **~55 µs** (10K/768D, k=10, Balanced).",
      "validation_command": "cargo bench -p velesdb-core --bench hnsw_benchmark -- hnsw_search_latency",
      "executable": false,
      "source": "commit df217860 (2026-03-27, 'docs: update README performance section with verified production metrics'): 'Core engine: Rust index-only for architecture reference (5K/768D: 55us HNSW search, 21.7ns dot product)' and 'Full production path: Python SDK, WAL ON, recall@10 >= 95% (10K/384D: 571us p50, 15K vec/s insert)'; 'Measured 2026-03-27, i9-14900KF, 64GB DDR5, Windows 11, Rust 1.92.' hnsw_benchmark's hnsw_search_latency group inserted 10,000 vectors at that commit, so the '5K' label was wrong; commit 45cb8021 relabeled it 10K and re-measured 53.2 us on an Apple M5 Pro (2026-07-20). measured_version is the workspace version of df217860.",
      "measured_on": "2026-03-27",
      "measured_machine": "Intel Core i9-14900KF, 64GB DDR5, Windows 11, Rust 1.92 (commit df217860)",
      "measured_version": "1.7.2",
      "stale_accepted": {
        "reason": "Dated by commit df217860 (2026-03-27, 1.7.2, i9-14900KF). The same day commit 126d5a9a recorded 47.0 us for the same bench on the same machine, and a quiet Apple M5 Pro re-run on 2026-07-20 measured 53.2 us; re-measuring belongs on the idle i9-14900KF, not done in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "core_perf_hnsw_index_only_note",
      "file": "docs/guides/CORE_PERFORMANCE.md",
      "must_contain": "The 55 µs index-only micro-benchmark",
      "validation_command": "cargo bench -p velesdb-core --bench hnsw_benchmark -- hnsw_search_latency",
      "executable": false,
      "source": "commit df217860 (2026-03-27, 'docs: update README performance section with verified production metrics'): 'Core engine: Rust index-only for architecture reference (5K/768D: 55us HNSW search, 21.7ns dot product)' and 'Full production path: Python SDK, WAL ON, recall@10 >= 95% (10K/384D: 571us p50, 15K vec/s insert)'; 'Measured 2026-03-27, i9-14900KF, 64GB DDR5, Windows 11, Rust 1.92.' hnsw_benchmark's hnsw_search_latency group inserted 10,000 vectors at that commit, so the '5K' label was wrong; commit 45cb8021 relabeled it 10K and re-measured 53.2 us on an Apple M5 Pro (2026-07-20). measured_version is the workspace version of df217860.",
      "measured_on": "2026-03-27",
      "measured_machine": "Intel Core i9-14900KF, 64GB DDR5, Windows 11, Rust 1.92 (commit df217860)",
      "measured_version": "1.7.2",
      "stale_accepted": {
        "reason": "Dated by commit df217860 (2026-03-27, 1.7.2, i9-14900KF). The same day commit 126d5a9a recorded 47.0 us for the same bench on the same machine, and a quiet Apple M5 Pro re-run on 2026-07-20 measured 53.2 us; re-measuring belongs on the idle i9-14900KF, not done in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "core_perf_columnstore_130x",
      "file": "docs/guides/CORE_PERFORMANCE.md",
      "must_contain": "up to **130x** faster than JSON scanning at scale",
      "validation_command": "cargo bench -p velesdb-core --bench column_filter_benchmark",
      "executable": false,
      "source": "docs/BENCHMARKS.md section 6 '#### Integer Equality Filter (`filter_eq_int`, measured 2026-03-19)': JSON scan 3.84 ms vs ColumnStore 29.5 us at 100K rows = 130x, recorded by the v1.6.0 release commit 92cc35bd; mirrors architecture_columnstore_130x.",
      "measured_on": "2026-03-19",
      "measured_machine": "Intel Core i9-14900KF (x86_64, AVX2) - docs/BENCHMARKS.md reference machine",
      "measured_version": "1.6.0",
      "stale_accepted": {
        "reason": "Published with its date in docs/BENCHMARKS.md section 6; the ratio is hardware-dependent and re-measuring it belongs on the idle i9-14900KF reference machine.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "python_engine_hnsw_index_only_55us",
      "file": "docs/guides/PYTHON_ENGINE_BENCHMARKS.md",
      "must_contain": "**~55 µs** (k=10, Balanced mode; i9-14900KF, 2026-03-27)",
      "validation_command": "cargo bench -p velesdb-core --bench hnsw_benchmark -- hnsw_search_latency",
      "executable": false,
      "source": "commit df217860 (2026-03-27, 'docs: update README performance section with verified production metrics'): 'Core engine: Rust index-only for architecture reference (5K/768D: 55us HNSW search, 21.7ns dot product)' and 'Full production path: Python SDK, WAL ON, recall@10 >= 95% (10K/384D: 571us p50, 15K vec/s insert)'; 'Measured 2026-03-27, i9-14900KF, 64GB DDR5, Windows 11, Rust 1.92.' hnsw_benchmark's hnsw_search_latency group inserted 10,000 vectors at that commit, so the '5K' label was wrong; commit 45cb8021 relabeled it 10K and re-measured 53.2 us on an Apple M5 Pro (2026-07-20). measured_version is the workspace version of df217860.",
      "measured_on": "2026-03-27",
      "measured_machine": "Intel Core i9-14900KF, 64GB DDR5, Windows 11, Rust 1.92 (commit df217860)",
      "measured_version": "1.7.2",
      "stale_accepted": {
        "reason": "Dated by commit df217860 (2026-03-27, 1.7.2, i9-14900KF). The same day commit 126d5a9a recorded 47.0 us for the same bench on the same machine, and a quiet Apple M5 Pro re-run on 2026-07-20 measured 53.2 us; re-measuring belongs on the idle i9-14900KF, not done in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "python_perf_measured_upsert_rate",
      "file": "docs/guides/PYTHON_PERFORMANCE.md",
      "must_contain": "inserted 10K × 384D vectors at 18,595 vec/s from Python",
      "validation_command": "python benchmarks/velesdb_benchmark.py --recall",
      "executable": false,
      "source": "benchmarks/report_1.7.2_2026-03-27.json results[0]: num_vectors 10000, insert_rate 18595, insert_total_s 0.54; written by benchmarks/velesdb_benchmark.py, whose measured insert calls col.upsert(batch) with BATCH_SIZE = 5000, a text payload per point and WAL on.",
      "measured_on": "2026-03-27",
      "measured_machine": "Intel Core i9-14900KF, 64 GB, Windows (benchmarks/report_1.7.2_2026-03-27.json)",
      "measured_version": "1.7.2",
      "stale_accepted": {
        "reason": "Published with its original date and version (2026-03-27, 1.7.2, Balanced then at ef 128). Re-measuring means re-running velesdb_benchmark.py on the idle i9-14900KF it was taken on.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "tuning_rust_core_section",
      "file": "docs/guides/TUNING_GUIDE.md",
      "must_contain": "### Rust Core (Criterion)",
      "covers_section": true,
      "validation_command": "cargo bench -p velesdb-core --bench hnsw_benchmark -- hnsw_search_latency && cargo bench -p velesdb-core --bench simd_benchmark",
      "executable": false,
      "source": "commit df217860 (2026-03-27, 'docs: update README performance section with verified production metrics'): 'Core engine: Rust index-only for architecture reference (5K/768D: 55us HNSW search, 21.7ns dot product)' and 'Full production path: Python SDK, WAL ON, recall@10 >= 95% (10K/384D: 571us p50, 15K vec/s insert)'; 'Measured 2026-03-27, i9-14900KF, 64GB DDR5, Windows 11, Rust 1.92.' hnsw_benchmark's hnsw_search_latency group inserted 10,000 vectors at that commit, so the '5K' label was wrong; commit 45cb8021 relabeled it 10K and re-measured 53.2 us on an Apple M5 Pro (2026-07-20). measured_version is the workspace version of df217860. The parallel-insert row (20.7 ms) the section also showed is in no record and was removed.",
      "measured_on": "2026-03-27",
      "measured_machine": "Intel Core i9-14900KF, 64GB DDR5, Windows 11, Rust 1.92 (commit df217860)",
      "measured_version": "1.7.2",
      "stale_accepted": {
        "reason": "Dated by commit df217860 (2026-03-27, 1.7.2, i9-14900KF). The same day commit 126d5a9a recorded 47.0 us for the same bench on the same machine, and a quiet Apple M5 Pro re-run on 2026-07-20 measured 53.2 us; re-measuring belongs on the idle i9-14900KF, not done in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "tuning_python_bindings_section",
      "file": "docs/guides/TUNING_GUIDE.md",
      "must_contain": "### Python Bindings (PyO3 + NumPy)",
      "covers_section": true,
      "validation_command": "python benchmarks/velesdb_benchmark.py --recall",
      "executable": false,
      "source": "commit df217860 (2026-03-27, 'docs: update README performance section with verified production metrics'): 'Core engine: Rust index-only for architecture reference (5K/768D: 55us HNSW search, 21.7ns dot product)' and 'Full production path: Python SDK, WAL ON, recall@10 >= 95% (10K/384D: 571us p50, 15K vec/s insert)'; 'Measured 2026-03-27, i9-14900KF, 64GB DDR5, Windows 11, Rust 1.92.' hnsw_benchmark's hnsw_search_latency group inserted 10,000 vectors at that commit, so the '5K' label was wrong; commit 45cb8021 relabeled it 10K and re-measured 53.2 us on an Apple M5 Pro (2026-07-20). measured_version is the workspace version of df217860. The section showed ~15.4K vec/s and ~630 us (avg), which no record states; it now shows the commit's own 15K vec/s and 571 us p50.",
      "measured_on": "2026-03-27",
      "measured_machine": "Intel Core i9-14900KF, 64GB DDR5, Windows 11, Rust 1.92 (commit df217860)",
      "measured_version": "1.7.2",
      "stale_accepted": {
        "reason": "Dated by commit df217860 (2026-03-27, 1.7.2, i9-14900KF). The same day commit 126d5a9a recorded 47.0 us for the same bench on the same machine, and a quiet Apple M5 Pro re-run on 2026-07-20 measured 53.2 us; re-measuring belongs on the idle i9-14900KF, not done in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "use_cases_target_p50_5ms",
      "file": "docs/guides/USE_CASES.md",
      "must_contain": "| Latency (p50) | < 5ms |",
      "validation_command": "none: no benchmark or test in the repository measures this target",
      "executable": false,
      "source": "docs/guides/USE_CASES.md: a use case's latency target, in a table headed 'Target (not measured)'.",
      "notes": "A target, not a measurement: the figure is the goal named in source.",
      "measured_on": "n/a",
      "measured_machine": "n/a",
      "measured_version": "n/a",
      "last_updated": "2026-09-13"
    },
    {
      "id": "use_cases_target_p99_15ms",
      "file": "docs/guides/USE_CASES.md",
      "must_contain": "| Latency (p99) | < 15ms |",
      "validation_command": "none: no benchmark or test in the repository measures this target",
      "executable": false,
      "source": "docs/guides/USE_CASES.md: a use case's latency target, in a table headed 'Target (not measured)'.",
      "notes": "A target, not a measurement: the figure is the goal named in source.",
      "measured_on": "n/a",
      "measured_machine": "n/a",
      "measured_version": "n/a",
      "last_updated": "2026-09-13"
    },
    {
      "id": "use_cases_target_p50_10ms",
      "file": "docs/guides/USE_CASES.md",
      "must_contain": "| Latency (p50) | < 10ms |",
      "validation_command": "none: no benchmark or test in the repository measures this target",
      "executable": false,
      "source": "docs/guides/USE_CASES.md: a use case's latency target, in a table headed 'Target (not measured)'.",
      "notes": "A target, not a measurement: the figure is the goal named in source.",
      "measured_on": "n/a",
      "measured_machine": "n/a",
      "measured_version": "n/a",
      "last_updated": "2026-09-13"
    },
    {
      "id": "use_cases_target_p99_30ms",
      "file": "docs/guides/USE_CASES.md",
      "must_contain": "| Latency (p99) | < 30ms |",
      "validation_command": "none: no benchmark or test in the repository measures this target",
      "executable": false,
      "source": "docs/guides/USE_CASES.md: a use case's latency target, in a table headed 'Target (not measured)'.",
      "notes": "A target, not a measurement: the figure is the goal named in source.",
      "measured_on": "n/a",
      "measured_machine": "n/a",
      "measured_version": "n/a",
      "last_updated": "2026-09-13"
    },
    {
      "id": "use_cases_target_1_hop_5ms",
      "file": "docs/guides/USE_CASES.md",
      "must_contain": "| Latency (1 hop) | < 5ms |",
      "validation_command": "none: no benchmark or test in the repository measures this target",
      "executable": false,
      "source": "docs/guides/USE_CASES.md: a use case's latency target, in a table headed 'Target (not measured)'.",
      "notes": "A target, not a measurement: the figure is the goal named in source.",
      "measured_on": "n/a",
      "measured_machine": "n/a",
      "measured_version": "n/a",
      "last_updated": "2026-09-13"
    },
    {
      "id": "use_cases_target_3_hops_50ms",
      "file": "docs/guides/USE_CASES.md",
      "must_contain": "| Latency (3 hops) | < 50ms |",
      "validation_command": "none: no benchmark or test in the repository measures this target",
      "executable": false,
      "source": "docs/guides/USE_CASES.md: a use case's latency target, in a table headed 'Target (not measured)'.",
      "notes": "A target, not a measurement: the figure is the goal named in source.",
      "measured_on": "n/a",
      "measured_machine": "n/a",
      "measured_version": "n/a",
      "last_updated": "2026-09-13"
    },
    {
      "id": "use_cases_target_20ms",
      "file": "docs/guides/USE_CASES.md",
      "must_contain": "| Latency | < 20ms |",
      "validation_command": "none: no benchmark or test in the repository measures this target",
      "executable": false,
      "source": "docs/guides/USE_CASES.md: a use case's latency target, in a table headed 'Target (not measured)'.",
      "notes": "A target, not a measurement: the figure is the goal named in source.",
      "measured_on": "n/a",
      "measured_machine": "n/a",
      "measured_version": "n/a",
      "last_updated": "2026-09-13"
    },
    {
      "id": "use_cases_target_p50_3ms",
      "file": "docs/guides/USE_CASES.md",
      "must_contain": "| Latency (p50) | < 3ms |",
      "validation_command": "none: no benchmark or test in the repository measures this target",
      "executable": false,
      "source": "docs/guides/USE_CASES.md: a use case's latency target, in a table headed 'Target (not measured)'.",
      "notes": "A target, not a measurement: the figure is the goal named in source.",
      "measured_on": "n/a",
      "measured_machine": "n/a",
      "measured_version": "n/a",
      "last_updated": "2026-09-13"
    },
    {
      "id": "use_cases_target_p99_10ms",
      "file": "docs/guides/USE_CASES.md",
      "must_contain": "| Latency (p99) | < 10ms |",
      "validation_command": "none: no benchmark or test in the repository measures this target",
      "executable": false,
      "source": "docs/guides/USE_CASES.md: a use case's latency target, in a table headed 'Target (not measured)'.",
      "notes": "A target, not a measurement: the figure is the goal named in source.",
      "measured_on": "n/a",
      "measured_machine": "n/a",
      "measured_version": "n/a",
      "last_updated": "2026-09-13"
    },
    {
      "id": "use_cases_target_15ms",
      "file": "docs/guides/USE_CASES.md",
      "must_contain": "| Latency | < 15ms |",
      "validation_command": "none: no benchmark or test in the repository measures this target",
      "executable": false,
      "source": "docs/guides/USE_CASES.md: a use case's latency target, in a table headed 'Target (not measured)'.",
      "notes": "A target, not a measurement: the figure is the goal named in source.",
      "measured_on": "n/a",
      "measured_machine": "n/a",
      "measured_version": "n/a",
      "last_updated": "2026-09-13"
    },
    {
      "id": "use_cases_target_5ms_per_check",
      "file": "docs/guides/USE_CASES.md",
      "must_contain": "| Latency | < 5ms per check |",
      "validation_command": "none: no benchmark or test in the repository measures this target",
      "executable": false,
      "source": "docs/guides/USE_CASES.md: a use case's latency target, in a table headed 'Target (not measured)'.",
      "notes": "A target, not a measurement: the figure is the goal named in source.",
      "measured_on": "n/a",
      "measured_machine": "n/a",
      "measured_version": "n/a",
      "last_updated": "2026-09-13"
    },
    {
      "id": "use_cases_target_50ms",
      "file": "docs/guides/USE_CASES.md",
      "must_contain": "| Latency | < 50ms |",
      "validation_command": "none: no benchmark or test in the repository measures this target",
      "executable": false,
      "source": "docs/guides/USE_CASES.md: a use case's latency target, in a table headed 'Target (not measured)'.",
      "notes": "A target, not a measurement: the figure is the goal named in source.",
      "measured_on": "n/a",
      "measured_machine": "n/a",
      "measured_version": "n/a",
      "last_updated": "2026-09-13"
    },
    {
      "id": "use_cases_target_3_hops_30ms",
      "file": "docs/guides/USE_CASES.md",
      "must_contain": "| Latency (3 hops) | < 30ms |",
      "validation_command": "none: no benchmark or test in the repository measures this target",
      "executable": false,
      "source": "docs/guides/USE_CASES.md: a use case's latency target, in a table headed 'Target (not measured)'.",
      "notes": "A target, not a measurement: the figure is the goal named in source.",
      "measured_on": "n/a",
      "measured_machine": "n/a",
      "measured_version": "n/a",
      "last_updated": "2026-09-13"
    },
    {
      "id": "use_cases_target_5_hops_100ms",
      "file": "docs/guides/USE_CASES.md",
      "must_contain": "| Latency (5 hops) | < 100ms |",
      "validation_command": "none: no benchmark or test in the repository measures this target",
      "executable": false,
      "source": "docs/guides/USE_CASES.md: a use case's latency target, in a table headed 'Target (not measured)'.",
      "notes": "A target, not a measurement: the figure is the goal named in source.",
      "measured_on": "n/a",
      "measured_machine": "n/a",
      "measured_version": "n/a",
      "last_updated": "2026-09-13"
    },
    {
      "id": "use_cases_target_store_5ms",
      "file": "docs/guides/USE_CASES.md",
      "must_contain": "| Store latency | < 5ms |",
      "validation_command": "none: no benchmark or test in the repository measures this target",
      "executable": false,
      "source": "docs/guides/USE_CASES.md: a use case's latency target, in a table headed 'Target (not measured)'.",
      "notes": "A target, not a measurement: the figure is the goal named in source.",
      "measured_on": "n/a",
      "measured_machine": "n/a",
      "measured_version": "n/a",
      "last_updated": "2026-09-13"
    },
    {
      "id": "use_cases_target_retrieve_10ms",
      "file": "docs/guides/USE_CASES.md",
      "must_contain": "| Retrieve latency | < 10ms |",
      "validation_command": "none: no benchmark or test in the repository measures this target",
      "executable": false,
      "source": "docs/guides/USE_CASES.md: a use case's latency target, in a table headed 'Target (not measured)'.",
      "notes": "A target, not a measurement: the figure is the goal named in source.",
      "measured_on": "n/a",
      "measured_machine": "n/a",
      "measured_version": "n/a",
      "last_updated": "2026-09-13"
    },
    {
      "id": "wasm_persistence_export_rate",
      "file": "docs/guides/WASM_PERSISTENCE.md",
      "must_contain": "| `export_to_bytes` | ~7 ms | ~4479 MB/s |",
      "validation_command": "cargo bench -p velesdb-wasm",
      "executable": false,
      "source": "CHANGELOG.md [0.3.0] - 2025-12-22, IndexedDB Persistence: 'Performance (after optimization): Export 4479 MB/s, Import 2943 MB/s'. The entry names no machine and no corpus. The ~7 ms is that rate applied to 10 000 x 768-D vectors (~30.7 MB).",
      "measured_on": "2025-12-22 (the 0.3.0 release date; the entry gives no run date)",
      "measured_machine": "unknown (the CHANGELOG entry names none)",
      "measured_version": "0.3.0",
      "stale_accepted": {
        "reason": "Recorded only by the 0.3.0 release notes; no later run of export_to_bytes/import_from_bytes is recorded, and re-measuring needs a wasm-pack build and a browser or Node harness, not done in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "wasm_persistence_import_rate",
      "file": "docs/guides/WASM_PERSISTENCE.md",
      "must_contain": "| `import_from_bytes` | ~10 ms | ~2943 MB/s |",
      "validation_command": "cargo bench -p velesdb-wasm",
      "executable": false,
      "source": "CHANGELOG.md [0.3.0] - 2025-12-22, IndexedDB Persistence: 'Performance (after optimization): Export 4479 MB/s, Import 2943 MB/s'. The entry names no machine and no corpus. The ~10 ms is that rate applied to 10 000 x 768-D vectors (~30.7 MB).",
      "measured_on": "2025-12-22 (the 0.3.0 release date; the entry gives no run date)",
      "measured_machine": "unknown (the CHANGELOG entry names none)",
      "measured_version": "0.3.0",
      "stale_accepted": {
        "reason": "Recorded only by the 0.3.0 release notes; no later run of export_to_bytes/import_from_bytes is recorded, and re-measuring needs a wasm-pack build and a browser or Node harness, not done in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "architecture_velesql_parse_bullet",
      "file": "docs/reference/ARCHITECTURE.md",
      "must_contain": "Parses a simple query in ~1.26 µs",
      "validation_command": "cargo bench -p velesdb-core --features internal-bench --bench velesql_benchmark -- --noplot",
      "executable": false,
      "source": "docs/BENCHMARKS.md section 7 'Measured March 19, 2026, sequential run on idle machine', recorded by the v1.6.0 release commit 92cc35bd ('VelesQL parser (1.26us simple, ...)'), mirrored from benchmarks_velesql_simple_parse. It replaces '1.3M queries/sec', which CHANGELOG [0.1.0] (2025-12-19) and [0.2.0] recorded for the 0.1.0 parser on an unnamed machine.",
      "measured_on": "2026-03-19",
      "measured_machine": "Intel Core i9-14900KF, idle - docs/BENCHMARKS.md reference machine (release commit 92cc35bd)",
      "measured_version": "1.6.0",
      "stale_accepted": {
        "reason": "Published with its date in docs/BENCHMARKS.md section 7; re-measuring VelesQL parse latencies belongs on the idle i9-14900KF reference machine.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "architecture_velesql_parse_row",
      "file": "docs/reference/ARCHITECTURE.md",
      "must_contain": "1.26 µs, ~794K QPS",
      "validation_command": "cargo bench -p velesdb-core --features internal-bench --bench velesql_benchmark -- --noplot",
      "executable": false,
      "source": "docs/BENCHMARKS.md section 7 'Measured March 19, 2026, sequential run on idle machine', recorded by the v1.6.0 release commit 92cc35bd ('VelesQL parser (1.26us simple, ...)'), mirrored from benchmarks_velesql_simple_parse. It replaces '1.3M queries/sec', which CHANGELOG [0.1.0] (2025-12-19) and [0.2.0] recorded for the 0.1.0 parser on an unnamed machine.",
      "measured_on": "2026-03-19",
      "measured_machine": "Intel Core i9-14900KF, idle - docs/BENCHMARKS.md reference machine (release commit 92cc35bd)",
      "measured_version": "1.6.0",
      "stale_accepted": {
        "reason": "Published with its date in docs/BENCHMARKS.md section 7; re-measuring VelesQL parse latencies belongs on the idle i9-14900KF reference machine.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "architecture_search_100k_target",
      "file": "docs/reference/ARCHITECTURE.md",
      "must_contain": "< 5 ms (target,",
      "validation_command": "cargo bench -p velesdb-core --bench scale_benchmark",
      "executable": false,
      "source": "docs/BENCHMARKS.md section 10 'Performance Targets by Scale': 100K vectors, Search P99 < 5 ms.",
      "notes": "A target, not a measurement: the figure is the goal named in source.",
      "measured_on": "n/a",
      "measured_machine": "n/a",
      "measured_version": "n/a",
      "last_updated": "2026-09-13"
    },
    {
      "id": "architecture_wasm_export",
      "file": "docs/reference/ARCHITECTURE.md",
      "must_contain": "| Export (WASM) | 4,479 MB/s (0.3.0 release notes",
      "validation_command": "cargo bench -p velesdb-wasm",
      "executable": false,
      "source": "CHANGELOG.md [0.3.0] - 2025-12-22, IndexedDB Persistence: 'Performance (after optimization): Export 4479 MB/s, Import 2943 MB/s'. The entry names no machine and no corpus.",
      "measured_on": "2025-12-22 (the 0.3.0 release date; the entry gives no run date)",
      "measured_machine": "unknown (the CHANGELOG entry names none)",
      "measured_version": "0.3.0",
      "stale_accepted": {
        "reason": "Recorded only by the 0.3.0 release notes; no later run of export_to_bytes/import_from_bytes is recorded, and re-measuring needs a wasm-pack build and a browser or Node harness, not done in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "architecture_wasm_import",
      "file": "docs/reference/ARCHITECTURE.md",
      "must_contain": "| Import (WASM) | 2,943 MB/s (0.3.0 release notes",
      "validation_command": "cargo bench -p velesdb-wasm",
      "executable": false,
      "source": "CHANGELOG.md [0.3.0] - 2025-12-22, IndexedDB Persistence: 'Performance (after optimization): Export 4479 MB/s, Import 2943 MB/s'. The entry names no machine and no corpus.",
      "measured_on": "2025-12-22 (the 0.3.0 release date; the entry gives no run date)",
      "measured_machine": "unknown (the CHANGELOG entry names none)",
      "measured_version": "0.3.0",
      "stale_accepted": {
        "reason": "Recorded only by the 0.3.0 release notes; no later run of export_to_bytes/import_from_bytes is recorded, and re-measuring needs a wasm-pack build and a browser or Node harness, not done in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "simd_perf_dot_row",
      "file": "docs/reference/SIMD_PERFORMANCE.md",
      "must_contain": "| `dot_product_native` | **19.8ns** | 38.8 Gelem/s |",
      "validation_command": "cargo bench -p velesdb-core --bench simd_benchmark -- --noplot",
      "executable": false,
      "source": "commit 126d5a9a (2026-03-27): 'Updated metrics (measured March 27, 2026, i9-14900KF, AVX2, Rust 1.92): ... SIMD Dot Product (768D): 17.6ns -> 19.8ns (AVX2, was AVX-512)'. Throughput is 768 divided by the latency.",
      "measured_on": "2026-03-27",
      "measured_machine": "Intel Core i9-14900KF, AVX2, Rust 1.92 (commit 126d5a9a)",
      "measured_version": "1.7.2",
      "stale_accepted": {
        "reason": "Dated by commit 126d5a9a (March 27, 2026); re-measuring belongs on the idle i9-14900KF reference machine.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "simd_perf_euclidean_row",
      "file": "docs/reference/SIMD_PERFORMANCE.md",
      "must_contain": "| `euclidean_native` | **22.5ns** | 34.1 Gelem/s |",
      "validation_command": "cargo bench -p velesdb-core --bench simd_benchmark -- --noplot",
      "executable": false,
      "source": "commit dd0d2457 (2026-03-24, 'docs: final benchmark update — real measurements March 24, 2026 22:56'): 'All metrics re-measured sequentially on idle i9-14900KF, 64GB DDR5' - 'Euclidean: 22.5 ns (34.1 Gelem/s)'. Throughput is 768 divided by the latency. measured_version is the workspace version of that commit.",
      "measured_on": "2026-03-24",
      "measured_machine": "Intel Core i9-14900KF, 64GB DDR5, idle (commit dd0d2457)",
      "measured_version": "1.7.0",
      "stale_accepted": {
        "reason": "Dated by commit dd0d2457 (March 24, 2026, 1.7.0); re-measuring SIMD kernels belongs on the idle i9-14900KF reference machine.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "simd_perf_cosine_row",
      "file": "docs/reference/SIMD_PERFORMANCE.md",
      "must_contain": "| `cosine_similarity_native` | **33.1ns** | 23.2 Gelem/s |",
      "validation_command": "cargo bench -p velesdb-core --bench simd_benchmark -- --noplot",
      "executable": false,
      "source": "commit dd0d2457 (2026-03-24, 'docs: final benchmark update — real measurements March 24, 2026 22:56'): 'All metrics re-measured sequentially on idle i9-14900KF, 64GB DDR5' - 'Cosine: 33.1 ns (23.2 Gelem/s)'. Throughput is 768 divided by the latency. measured_version is the workspace version of that commit.",
      "measured_on": "2026-03-24",
      "measured_machine": "Intel Core i9-14900KF, 64GB DDR5, idle (commit dd0d2457)",
      "measured_version": "1.7.0",
      "stale_accepted": {
        "reason": "Dated by commit dd0d2457 (March 24, 2026, 1.7.0); re-measuring SIMD kernels belongs on the idle i9-14900KF reference machine.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "simd_perf_hamming_row",
      "file": "docs/reference/SIMD_PERFORMANCE.md",
      "must_contain": "| `hamming_distance_native` | **35.8ns** | 21.5 Gelem/s |",
      "validation_command": "cargo bench -p velesdb-core --bench simd_benchmark -- --noplot",
      "executable": false,
      "source": "commit dd0d2457 (2026-03-24, 'docs: final benchmark update — real measurements March 24, 2026 22:56'): 'All metrics re-measured sequentially on idle i9-14900KF, 64GB DDR5' - 'Hamming: 35.8 ns'. Throughput is 768 divided by the latency. measured_version is the workspace version of that commit.",
      "measured_on": "2026-03-24",
      "measured_machine": "Intel Core i9-14900KF, 64GB DDR5, idle (commit dd0d2457)",
      "measured_version": "1.7.0",
      "stale_accepted": {
        "reason": "Dated by commit dd0d2457 (March 24, 2026, 1.7.0); re-measuring SIMD kernels belongs on the idle i9-14900KF reference machine.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "simd_perf_jaccard_row",
      "file": "docs/reference/SIMD_PERFORMANCE.md",
      "must_contain": "| `jaccard_similarity_native` | **35.1ns** | 21.9 Gelem/s |",
      "validation_command": "cargo bench -p velesdb-core --bench simd_benchmark -- --noplot",
      "executable": false,
      "source": "commit dd0d2457 (2026-03-24, 'docs: final benchmark update — real measurements March 24, 2026 22:56'): 'All metrics re-measured sequentially on idle i9-14900KF, 64GB DDR5' - 'Jaccard: 35.1 ns'. Throughput is 768 divided by the latency. measured_version is the workspace version of that commit.",
      "measured_on": "2026-03-24",
      "measured_machine": "Intel Core i9-14900KF, 64GB DDR5, idle (commit dd0d2457)",
      "measured_version": "1.7.0",
      "stale_accepted": {
        "reason": "Dated by commit dd0d2457 (March 24, 2026, 1.7.0); re-measuring SIMD kernels belongs on the idle i9-14900KF reference machine.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "context_savings_persisted_compile_latency",
      "file": "crates/velesdb-memory/examples/context_savings/README.md",
      "must_contain": "compile latency (with source/event persistence, default): mean 24.5 ms, max 37.2 ms",
      "validation_command": "cd crates/velesdb-memory/examples/context_savings/real_measures && node agent_session.mjs",
      "executable": false,
      "source": "crates/velesdb-memory/examples/context_savings/README.md, agent-session output block (committed with its token figures by ebdbf015, 2026-07-26): 'compile latency (with source/event persistence, default): mean 24.5 ms, max 37.2 ms' and 'compile latency (stateless: ...): mean 0.7 ms' - the same run as readme_economics_context_savings_corpus (82.5% saved).",
      "measured_on": "2026-07-26",
      "measured_machine": "unknown (latency depends on the machine, which the run does not record; its token figures are platform-independent)",
      "measured_version": "4.0.0",
      "stale_accepted": {
        "reason": "Printed by the same run as the 82.5% corpus figure (4.0.0) and published as that run's output; re-running agent_session.mjs needs the Node binding built from the current tree, not done in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "context_savings_compile_scaling_throughput",
      "file": "crates/velesdb-memory/examples/context_savings/README.md",
      "must_contain": "throughput held at ~168 Kelem/s across all four sizes",
      "validation_command": "cargo bench -p velesdb-memory --no-default-features --features context",
      "executable": false,
      "source": "commit 4fa4fde8 (2026-07-17, EPIC-P-070) committed benches/context_compiler_benchmark.rs and states 'linear scaling: ~168 Kelem/s throughput held constant'; the README section says 'release profile, Apple Silicon'. measured_version is the workspace version of that commit.",
      "measured_on": "2026-07-17 (date of commit 4fa4fde8, which records the run)",
      "measured_machine": "Apple Silicon (the README's words; commit 4fa4fde8 names no model)",
      "measured_version": "3.12.0",
      "stale_accepted": {
        "reason": "Recorded by commit 4fa4fde8 (3.12.0); re-running the criterion bench needs a machine at rest, not done in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "rustdoc_memory_per_point_fsync_cap",
      "file": "crates/velesdb-memory/src/migration/tests/performance.rs",
      "must_contain": "at ~290 points/s where ~66 000",
      "validation_command": "cargo test -p velesdb-memory --release -- --ignored the_per_point_write_cost_is_attributed_to_payload_or_vector",
      "executable": false,
      "source": "the ignored diagnostic test's doc comment: '~3.4 ms is exactly one fsync' per text-bearing point, which caps inserts at ~290 points/s against ~66 000 without it. Commit 83784a92 (2026-08-03, #1762) committed it and says its 'Diagnostic benchmarks are #[ignore]d and carry their measurements in their doc comments, including the write-path attribution that located a per-document fsync in the bulk text-index path'. measured_version is the workspace version of that commit.",
      "measured_on": "2026-08-03 (date of commit 83784a92, which records the run)",
      "measured_machine": "unknown (neither the doc comment nor commit 83784a92 names it)",
      "measured_version": "4.2.0",
      "stale_accepted": {
        "reason": "The test's own recorded measurement (83784a92, 4.2.0); re-running it writes 2 000 facts several times and, per its #[ignore], wants a machine at rest - not done in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "migrate_readme_quickstart_throughput",
      "file": "crates/velesdb-migrate/README.md",
      "must_contain": "Throughput: 24 vectors/sec",
      "validation_command": "velesdb-migrate run --config migration.yaml",
      "executable": false,
      "source": "the quickstart log pasted in crates/velesdb-migrate/README.md by commit 871d21af (2026-07-25): 'Migration complete: 3 extracted, 3 loaded, 0 failed in 0.12s (24 pts/sec)' - a 3-point smoke run, which the README says is not a rate to plan with.",
      "measured_on": "2026-07-25 (the log's timestamps)",
      "measured_machine": "unknown (the log names none; it reports avx512=false avx2=false)",
      "measured_version": "4.0.0 (workspace version of commit 871d21af, which pasted the log the day it was taken)",
      "stale_accepted": {
        "reason": "A 3-point smoke run kept as sample output, not a rate; re-running it changes only the timestamps.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "cli_import_target_rate",
      "file": "crates/velesdb-cli/src/import.rs",
      "must_contain": "Target: ~3-5K vectors/sec at 768D with batch_size=1000",
      "validation_command": "none: no benchmark or test in the repository measures this target",
      "executable": false,
      "source": "A design target stated in the rustdoc of import_jsonl and import_csv.",
      "notes": "A target, not a measurement: the figure is the goal named in source.",
      "measured_on": "n/a",
      "measured_machine": "n/a",
      "measured_version": "n/a",
      "last_updated": "2026-09-13"
    },
    {
      "id": "agent_memory_throughput_section",
      "file": "docs/guides/AGENT_MEMORY.md",
      "must_contain": "### Throughput",
      "covers_section": true,
      "validation_command": "cargo bench -p velesdb-core --bench agent_memory_benchmark",
      "executable": false,
      "source": "docs/guides/AGENT_MEMORY.md 'Performance & Limits > Throughput': 'Measured with `cargo bench -p velesdb-core --bench agent_memory_benchmark` (criterion, release profile) on an Apple M5 Pro (64 GB), 2026-06-12', written with its table by commit 194af969 (2026-06-12, 'feat(bench): agent memory criterion benchmark + measured figures in the guide'). measured_version is the workspace version of that commit.",
      "measured_on": "2026-06-12",
      "measured_machine": "Apple M5 Pro, 64 GB (commit 194af969)",
      "measured_version": "1.18.0",
      "stale_accepted": {
        "reason": "Dated in the section itself (2026-06-12, commit 194af969, 1.18.0); re-running agent_memory_benchmark needs the same Apple M5 Pro at rest, not done in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "readme_columnstore_apple_ratio_range",
      "file": "README.md",
      "must_contain": "the same bench measures ~50–105x",
      "validation_command": "cargo bench -p velesdb-core --bench column_filter_benchmark",
      "executable": false,
      "source": "commit 45cb8021 (2026-07-20, #1502): 'Numbers were re-verified today by a sequential quiet-machine re-bench; ... the ColumnStore 130x ratio is hardware-qualified (i9 reference; ~50-105x on Apple Silicon where the JSON scan itself is faster)'. The README sentence the commit wrote names the machine. measured_version is the workspace version of that commit.",
      "measured_on": "2026-07-20",
      "measured_machine": "Apple M5 Pro (named by the README sentence commit 45cb8021 wrote; the commit says 'Apple Silicon')",
      "measured_version": "3.12.0",
      "stale_accepted": {
        "reason": "Recorded by commit 45cb8021 (3.12.0) as a cross-check of the i9 ratio; re-measuring needs the same Apple M5 Pro at rest, not done in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "readme_columnstore_apple_absolute_time",
      "file": "README.md",
      "must_contain": "the ColumnStore's absolute time holds (~27 µs)",
      "validation_command": "cargo bench -p velesdb-core --bench column_filter_benchmark",
      "executable": false,
      "source": "commit 45cb8021 (2026-07-20, #1502): 'Numbers were re-verified today by a sequential quiet-machine re-bench; ... the ColumnStore 130x ratio is hardware-qualified (i9 reference; ~50-105x on Apple Silicon where the JSON scan itself is faster)'. The README sentence the commit wrote names the machine. measured_version is the workspace version of that commit. The same sentence gives the ColumnStore's absolute time on that machine, ~27 us.",
      "measured_on": "2026-07-20",
      "measured_machine": "Apple M5 Pro (named by the README sentence commit 45cb8021 wrote; the commit says 'Apple Silicon')",
      "measured_version": "3.12.0",
      "stale_accepted": {
        "reason": "Recorded by commit 45cb8021 (3.12.0) as a cross-check of the i9 ratio; re-measuring needs the same Apple M5 Pro at rest, not done in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "readme_columnstore_i9_absolute_times",
      "file": "README.md",
      "must_contain": "`JSON scan 3.84 ms → ColumnStore 29.5 us`",
      "validation_command": "cargo bench -p velesdb-core --bench column_filter_benchmark",
      "executable": false,
      "source": "docs/BENCHMARKS.md section 6 '#### Integer Equality Filter (`filter_eq_int`, measured 2026-03-19)', row '| 100K rows | 29.5 µs | 3.84 ms | 130x |', recorded by the v1.6.0 release commit 92cc35bd ('Re-run all 6 benchmark suites sequentially on idle machine (i9-14900KF, 64GB DDR5, Windows 11, Rust 1.92.0) — March 19, 2026'). These are the two times whose ratio readme_columnstore_130x_claim pins.",
      "measured_on": "2026-03-19",
      "measured_machine": "Intel Core i9-14900KF (x86_64, AVX2) - docs/BENCHMARKS.md reference machine",
      "measured_version": "1.6.0",
      "stale_accepted": {
        "reason": "Published with its date in docs/BENCHMARKS.md section 6; the ratio is hardware-dependent and re-measuring it belongs on the idle i9-14900KF reference machine.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "benchmarks_simd_kernel_latency_section",
      "file": "docs/BENCHMARKS.md",
      "must_contain": "### SIMD Kernel Latency",
      "covers_section": true,
      "validation_command": "cargo bench -p velesdb-core --bench simd_benchmark -- --noplot",
      "executable": false,
      "source": "docs/BENCHMARKS.md section 1: 'Measured April 3, 2026 on Intel Core i9-14900KF (24C/32T, AVX2+FMA), 64GB DDR5, Rust 1.94.1, Windows 11 Pro, sequential run on idle machine'; commit 81b371ba (2026-04-03) set this table's figures with that note. measured_version is the workspace version of that commit.",
      "measured_on": "2026-04-03",
      "measured_machine": "Intel Core i9-14900KF (24C/32T, AVX2+FMA), 64GB DDR5, Windows 11 Pro, Rust 1.94.1, idle (docs/BENCHMARKS.md section 1)",
      "measured_version": "1.11.0",
      "stale_accepted": {
        "reason": "Dated in docs/BENCHMARKS.md section 1 (2026-04-03, 1.11.0); re-measuring belongs on the idle i9-14900KF reference machine, not done in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "benchmarks_cosine_engine_dispatch_section",
      "file": "docs/BENCHMARKS.md",
      "must_contain": "### Cosine Engine Dispatch Overhead (March 19, 2026)",
      "covers_section": true,
      "validation_command": "cargo bench -p velesdb-core --bench simd_benchmark -- --noplot",
      "executable": false,
      "source": "the v1.6.0 release commit 92cc35bd (2026-03-20): 'Re-run all 6 benchmark suites sequentially on idle machine (i9-14900KF, 64GB DDR5, Windows 11, Rust 1.92.0) - March 19, 2026'; it set this table (384D native 21.1 ns, engine 28.0 ns). measured_version is the workspace version of that commit.",
      "measured_on": "2026-03-19",
      "measured_machine": "Intel Core i9-14900KF, 64GB DDR5, Windows 11, Rust 1.92.0, idle (release commit 92cc35bd)",
      "measured_version": "1.6.0",
      "stale_accepted": {
        "reason": "Recorded by the v1.6.0 release run (2026-03-19, 1.6.0); re-measuring belongs on the idle i9-14900KF reference machine, not done in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "benchmarks_batch_distance_section",
      "file": "docs/BENCHMARKS.md",
      "must_contain": "### Batch Distance Computation",
      "covers_section": true,
      "validation_command": "cargo bench -p velesdb-core --bench simd_benchmark -- --noplot",
      "executable": false,
      "source": "docs/BENCHMARKS.md section 1: 'Measured April 3, 2026 on Intel Core i9-14900KF (24C/32T, AVX2+FMA), 64GB DDR5, Rust 1.94.1, Windows 11 Pro, sequential run on idle machine'; commit 81b371ba (2026-04-03) set this table's figures with that note. measured_version is the workspace version of that commit.",
      "measured_on": "2026-04-03",
      "measured_machine": "Intel Core i9-14900KF (24C/32T, AVX2+FMA), 64GB DDR5, Windows 11 Pro, Rust 1.94.1, idle (docs/BENCHMARKS.md section 1)",
      "measured_version": "1.11.0",
      "stale_accepted": {
        "reason": "Dated in docs/BENCHMARKS.md section 1 (2026-04-03, 1.11.0); re-measuring belongs on the idle i9-14900KF reference machine, not done in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "benchmarks_pq_recall_full_precision_latency",
      "file": "docs/BENCHMARKS.md",
      "must_contain": "19.1 ms ² | 382 us ²",
      "validation_command": "cargo bench -p velesdb-core --bench pq_recall_benchmark -- --noplot",
      "executable": false,
      "source": "commit 7f05d0f0 (2026-03-07, 'update BENCHMARKS.md with v1.5 benchmark results'): 'Add PQ recall@k and latency tables from pq_recall_benchmark and pq_hnsw_benchmark' and 'Add Test Environment section from machine-config.json' (Intel Core i9-14900KF, 64 GB, Windows 11, rustc 1.92.0, captured 2026-03-06). It set the Full Precision latencies 19.1 ms / 382 us on the clustered data the benchmark generated then; the row's recall is registered separately (benchmarks_pq_full_precision_recall). measured_version is the workspace version of that commit.",
      "measured_on": "2026-03-07 (date of commit 7f05d0f0, which records the run)",
      "measured_machine": "Intel Core i9-14900KF, 64 GB, Windows 11, Rust 1.92.0 (benchmarks/machine-config.json, added as the Test Environment by the same commit)",
      "measured_version": "1.4.5",
      "stale_accepted": {
        "reason": "Recorded by commit 7f05d0f0 (2026-03-07, 1.4.5) and footnoted as not re-measured since; re-running it belongs on the idle i9-14900KF, not done in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "benchmarks_pq_recall_pq_latency",
      "file": "docs/BENCHMARKS.md",
      "must_contain": "30.6 ms ² | 612 us ²",
      "validation_command": "cargo bench -p velesdb-core --bench pq_recall_benchmark -- --noplot",
      "executable": false,
      "source": "commit 7f05d0f0 (2026-03-07, 'update BENCHMARKS.md with v1.5 benchmark results'): 'Add PQ recall@k and latency tables from pq_recall_benchmark and pq_hnsw_benchmark' and 'Add Test Environment section from machine-config.json' (Intel Core i9-14900KF, 64 GB, Windows 11, rustc 1.92.0, captured 2026-03-06). It set the PQ latencies 30.6 ms / 612 us, when the benchmark ran 'PQ (m=auto, k=auto, rescore)' on clustered data; commit c4d8acbe (2026-06-21) relabeled the row m=8, k=256. measured_version is the workspace version of that commit.",
      "measured_on": "2026-03-07 (date of commit 7f05d0f0, which records the run)",
      "measured_machine": "Intel Core i9-14900KF, 64 GB, Windows 11, Rust 1.92.0 (benchmarks/machine-config.json, added as the Test Environment by the same commit)",
      "measured_version": "1.4.5",
      "stale_accepted": {
        "reason": "Recorded by commit 7f05d0f0 (2026-03-07, 1.4.5) and footnoted as not re-measured since; re-running it belongs on the idle i9-14900KF, not done in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "benchmarks_pq_sq8_full_latency_section",
      "file": "docs/BENCHMARKS.md",
      "must_contain": "### PQ vs SQ8 vs Full HNSW Latency (pq_hnsw_benchmark)",
      "covers_section": true,
      "validation_command": "cargo bench -p velesdb-core --bench pq_hnsw_benchmark -- --noplot",
      "executable": false,
      "source": "commit 7f05d0f0 (2026-03-07, 'update BENCHMARKS.md with v1.5 benchmark results'): 'Add PQ recall@k and latency tables from pq_recall_benchmark and pq_hnsw_benchmark' and 'Add Test Environment section from machine-config.json' (Intel Core i9-14900KF, 64 GB, Windows 11, rustc 1.92.0, captured 2026-03-06). It set this table (Full 24.9 us, SQ8 25.2 us, PQ 257.6 us, 2,000 x 64D, top-20). measured_version is the workspace version of that commit.",
      "measured_on": "2026-03-07 (date of commit 7f05d0f0, which records the run)",
      "measured_machine": "Intel Core i9-14900KF, 64 GB, Windows 11, Rust 1.92.0 (benchmarks/machine-config.json, added as the Test Environment by the same commit)",
      "measured_version": "1.4.5",
      "stale_accepted": {
        "reason": "Recorded by commit 7f05d0f0 (2026-03-07, 1.4.5) and footnoted as not re-measured since; re-running it belongs on the idle i9-14900KF, not done in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "benchmarks_sparse_insert_parallel",
      "file": "docs/BENCHMARKS.md",
      "must_contain": "`parallel_10k_manual_4x2500_doc_granular` | 155 ms",
      "validation_command": "cargo bench -p velesdb-core --bench sparse_benchmark -- --noplot",
      "executable": false,
      "source": "the v1.6.0 release commit 92cc35bd (2026-03-20): 'Re-run all 6 benchmark suites sequentially on idle machine (i9-14900KF, 64GB DDR5, Windows 11, Rust 1.92.0) - March 19, 2026'; it set this row (Insert 10K parallel, 155 ms). measured_version is the workspace version of that commit.",
      "measured_on": "2026-03-19",
      "measured_machine": "Intel Core i9-14900KF, 64GB DDR5, Windows 11, Rust 1.92.0, idle (release commit 92cc35bd)",
      "measured_version": "1.6.0",
      "stale_accepted": {
        "reason": "Recorded by the v1.6.0 release run (2026-03-19, 1.6.0); re-measuring belongs on the idle i9-14900KF reference machine, not done in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "benchmarks_sparse_search_top10",
      "file": "docs/BENCHMARKS.md",
      "must_contain": "| `top10_10k_corpus` | **57.6 µs** |",
      "validation_command": "cargo bench -p velesdb-core --bench sparse_benchmark -- --noplot",
      "executable": false,
      "source": "commit f4999a74 (PR #621, 2026-04-21): 'Measured impact (i9-14900K, SPLADE 10K corpus): top-10: 956 us -> 57.6 us (-94%, 16.6x speedup); top-100: 927 us -> 75.1 us (-92%, 12.3x speedup)'; this row is the top-10 figure. measured_version is the workspace version of that commit.",
      "measured_on": "2026-04-21 (date of commit f4999a74, which records the run)",
      "measured_machine": "i9-14900K, as commit f4999a74 names it (docs/BENCHMARKS.md's reference machine is the i9-14900KF)",
      "measured_version": "1.13.0",
      "stale_accepted": {
        "reason": "Recorded by commit f4999a74 (2026-04-21, 1.13.0); re-measuring belongs on the idle i9-14900KF reference machine, not done in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "benchmarks_sparse_search_top100",
      "file": "docs/BENCHMARKS.md",
      "must_contain": "| `top100_10k_corpus` | **75.1 µs** |",
      "validation_command": "cargo bench -p velesdb-core --bench sparse_benchmark -- --noplot",
      "executable": false,
      "source": "commit f4999a74 (PR #621, 2026-04-21): 'Measured impact (i9-14900K, SPLADE 10K corpus): top-10: 956 us -> 57.6 us (-94%, 16.6x speedup); top-100: 927 us -> 75.1 us (-92%, 12.3x speedup)'; this row is the top-100 figure. measured_version is the workspace version of that commit.",
      "measured_on": "2026-04-21 (date of commit f4999a74, which records the run)",
      "measured_machine": "i9-14900K, as commit f4999a74 names it (docs/BENCHMARKS.md's reference machine is the i9-14900KF)",
      "measured_version": "1.13.0",
      "stale_accepted": {
        "reason": "Recorded by commit f4999a74 (2026-04-21, 1.13.0); re-measuring belongs on the idle i9-14900KF reference machine, not done in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "benchmarks_upsert_bulk_ratio_v172",
      "file": "docs/BENCHMARKS.md",
      "must_contain": "The upsert/bulk ratio dropped from ~19x to ~1x.",
      "validation_command": "cargo bench -p velesdb-core --bench bulk_insert_v2_benchmark",
      "executable": false,
      "source": "commit 024e4a9b (2026-03-25, 'docs: complete v1.7.2 documentation with upsert lock contention fix') and CHANGELOG.md [1.7.2]: 'Measured on i9-14900KF (10K vectors, 384D): upsert throughput rose from ~808 vec/s to ~16,151 vec/s, closing the gap with upsert_bulk() from 19x to ~1x'. measured_version is the workspace version of that commit.",
      "measured_on": "2026-03-25 (date of commit 024e4a9b, which records the run; the line states no date)",
      "measured_machine": "Intel Core i9-14900KF (as the CHANGELOG entry names it)",
      "measured_version": "1.7.2",
      "stale_accepted": {
        "reason": "Recorded with the v1.7.2 change it measured (2026-03-25); re-running it belongs on the i9-14900KF, not done in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "benchmarks_columnstore_string_eq_section",
      "file": "docs/BENCHMARKS.md",
      "must_contain": "#### String Equality Filter (`filter_eq_string`, measured 2026-03-19)",
      "covers_section": true,
      "validation_command": "cargo bench -p velesdb-core --bench column_filter_benchmark",
      "executable": false,
      "source": "the v1.6.0 release commit 92cc35bd (2026-03-20): 'Re-run all 6 benchmark suites sequentially on idle machine (i9-14900KF, 64GB DDR5, Windows 11, Rust 1.92.0) - March 19, 2026'; it set this table (100K rows: ColumnStore 46.5 us, JSON scan 3.50 ms, 75x). measured_version is the workspace version of that commit.",
      "measured_on": "2026-03-19",
      "measured_machine": "Intel Core i9-14900KF, 64GB DDR5, Windows 11, Rust 1.92.0, idle (release commit 92cc35bd)",
      "measured_version": "1.6.0",
      "stale_accepted": {
        "reason": "Recorded by the v1.6.0 release run (2026-03-19, 1.6.0); re-measuring belongs on the idle i9-14900KF reference machine, not done in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "benchmarks_columnstore_int_eq_section",
      "file": "docs/BENCHMARKS.md",
      "must_contain": "#### Integer Equality Filter (`filter_eq_int`, measured 2026-03-19)",
      "covers_section": true,
      "validation_command": "cargo bench -p velesdb-core --bench column_filter_benchmark",
      "executable": false,
      "source": "the v1.6.0 release commit 92cc35bd (2026-03-20): 'Re-run all 6 benchmark suites sequentially on idle machine (i9-14900KF, 64GB DDR5, Windows 11, Rust 1.92.0) - March 19, 2026'; it set this table (100K rows: ColumnStore 29.5 us, JSON scan 3.84 ms, 130x). measured_version is the workspace version of that commit.",
      "measured_on": "2026-03-19",
      "measured_machine": "Intel Core i9-14900KF, 64GB DDR5, Windows 11, Rust 1.92.0, idle (release commit 92cc35bd)",
      "measured_version": "1.6.0",
      "stale_accepted": {
        "reason": "Recorded by the v1.6.0 release run (2026-03-19, 1.6.0); re-measuring belongs on the idle i9-14900KF reference machine, not done in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "benchmarks_graph_edgestore_section",
      "file": "docs/BENCHMARKS.md",
      "must_contain": "## 8. Graph (EdgeStore)",
      "covers_section": true,
      "validation_command": "cargo bench -p velesdb-core --bench edge_store_benchmark -- --noplot",
      "executable": false,
      "source": "commit 69638b38 (2026-03-11, 'update all benchmark metrics after full re-benchmark (2026-03-11)'): 'All benchmarks re-run on i9-14900KF, Rust 1.92.0' - 'EdgeStore get_neighbors: 165ns -> 95ns'; it set every figure of this table (95 ns, 485 ns, 265 ns, 3.32 us, 292 us), which later release runs left unchanged. measured_version is the workspace version of that commit.",
      "measured_on": "2026-03-11 (date of commit 69638b38, which records the run)",
      "measured_machine": "Intel Core i9-14900KF, Rust 1.92.0 (commit 69638b38)",
      "measured_version": "1.5.1",
      "stale_accepted": {
        "reason": "Recorded by commit 69638b38 (2026-03-11, 1.5.1); re-measuring belongs on the idle i9-14900KF reference machine, not done in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "benchmarks_filtered_search_bitmap_section",
      "file": "docs/BENCHMARKS.md",
      "must_contain": "### Filtered Search — Bitmap Pre-filter V2 (Issue #487)",
      "covers_section": true,
      "validation_command": "cargo bench -p velesdb-core --bench bitmap_prefilter_benchmark -- --noplot",
      "executable": false,
      "source": "commit 8939a728 (2026-04-03, 'docs: update BENCHMARKS.md and CONCURRENCY_MODEL.md with perf optimization results (#487, #491, #488)') added the section with its figures (1% 32 us, 10% 65 us, 50% 302 us, unfiltered 83 us). measured_version is the workspace version of that commit.",
      "measured_on": "2026-04 (commit 8939a728, 2026-04-03; the section states no date)",
      "measured_machine": "unknown (neither the section nor commit 8939a728 names it)",
      "measured_version": "1.11.0",
      "stale_accepted": {
        "reason": "Recorded by commit 8939a728 (1.11.0), which names no machine; re-running it needs a quiet reference machine, not done in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "benchmarks_targets_by_scale_section",
      "file": "docs/BENCHMARKS.md",
      "must_contain": "## 10. Performance Targets by Scale",
      "covers_section": true,
      "validation_command": "none: no benchmark or test in the repository measures this target",
      "executable": false,
      "source": "Targets stated in docs/BENCHMARKS.md section 10; no recorded run measures search P99 at these scales. The 'Achieved (96.1%)' the 100K row carried came from commit 32bfd2df's benchmarks/DOCKER_BENCHMARK_RESULTS.md (100K x 768D, HighRecall ef 1024: recall 96.1%, P99 90.6 ms), which met the recall floor but not the 5 ms P99 target; the status column was removed.",
      "notes": "A target, not a measurement: the figure is the goal named in source.",
      "measured_on": "n/a",
      "measured_machine": "n/a",
      "measured_version": "n/a",
      "last_updated": "2026-09-13"
    },
    {
      "id": "benchmarks_sift1m_plain_ef64",
      "file": "docs/BENCHMARKS.md",
      "must_contain": "| 64 | 0.9012 | 71.8 µs |",
      "validation_command": "cargo bench -p velesdb-core --bench sift1m_recall --features bench-sift1m",
      "executable": false,
      "source": "commit 4b62338f (#1225, 2026-06-24): 'Re-measured on an idle machine, single bench process (Criterion 95% CIs, stable within +/-1% run-to-run): ef 64/128/256/512 p50 = 71.8 / 130.5 / 235.8 / 433.3 us'; section 11.3 records the run as 'First reproducible run, VelesDB v3.3.0' on Apple Silicon, with recall@10 0.9012 / 0.9435 / 0.9659 / 0.9759. measured_version is the workspace version of that commit.",
      "measured_on": "2026-06-24 (date of commit 4b62338f, which records the run)",
      "measured_machine": "Apple Silicon (M-series, ARM64/NEON), otherwise idle, single bench process (docs/BENCHMARKS.md section 11.3)",
      "measured_version": "3.3.0",
      "stale_accepted": {
        "reason": "Published with its version and host in section 11.3 (v3.3.0, Apple Silicon); re-running SIFT1M builds a 1M index and needs the same idle host, not done in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "benchmarks_sift1m_plain_ef128",
      "file": "docs/BENCHMARKS.md",
      "must_contain": "| 128 | 0.9435 | 130.5 µs |",
      "validation_command": "cargo bench -p velesdb-core --bench sift1m_recall --features bench-sift1m",
      "executable": false,
      "source": "commit 4b62338f (#1225, 2026-06-24): 'Re-measured on an idle machine, single bench process (Criterion 95% CIs, stable within +/-1% run-to-run): ef 64/128/256/512 p50 = 71.8 / 130.5 / 235.8 / 433.3 us'; section 11.3 records the run as 'First reproducible run, VelesDB v3.3.0' on Apple Silicon, with recall@10 0.9012 / 0.9435 / 0.9659 / 0.9759. measured_version is the workspace version of that commit.",
      "measured_on": "2026-06-24 (date of commit 4b62338f, which records the run)",
      "measured_machine": "Apple Silicon (M-series, ARM64/NEON), otherwise idle, single bench process (docs/BENCHMARKS.md section 11.3)",
      "measured_version": "3.3.0",
      "stale_accepted": {
        "reason": "Published with its version and host in section 11.3 (v3.3.0, Apple Silicon); re-running SIFT1M builds a 1M index and needs the same idle host, not done in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "benchmarks_sift1m_plain_ef256",
      "file": "docs/BENCHMARKS.md",
      "must_contain": "| 256 | 0.9659 | 235.8 µs |",
      "validation_command": "cargo bench -p velesdb-core --bench sift1m_recall --features bench-sift1m",
      "executable": false,
      "source": "commit 4b62338f (#1225, 2026-06-24): 'Re-measured on an idle machine, single bench process (Criterion 95% CIs, stable within +/-1% run-to-run): ef 64/128/256/512 p50 = 71.8 / 130.5 / 235.8 / 433.3 us'; section 11.3 records the run as 'First reproducible run, VelesDB v3.3.0' on Apple Silicon, with recall@10 0.9012 / 0.9435 / 0.9659 / 0.9759. At ef 256 the commit message's first bullet gives recall 0.9656, but the table the same commit wrote gives 0.9659; the committed table value is the one the line states and this claim pins. measured_version is the workspace version of that commit.",
      "measured_on": "2026-06-24 (date of commit 4b62338f, which records the run)",
      "measured_machine": "Apple Silicon (M-series, ARM64/NEON), otherwise idle, single bench process (docs/BENCHMARKS.md section 11.3)",
      "measured_version": "3.3.0",
      "stale_accepted": {
        "reason": "Published with its version and host in section 11.3 (v3.3.0, Apple Silicon); re-running SIFT1M builds a 1M index and needs the same idle host, not done in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "benchmarks_sift1m_plain_ef512",
      "file": "docs/BENCHMARKS.md",
      "must_contain": "| 512 | 0.9759 | 433.3 µs |",
      "validation_command": "cargo bench -p velesdb-core --bench sift1m_recall --features bench-sift1m",
      "executable": false,
      "source": "commit 4b62338f (#1225, 2026-06-24): 'Re-measured on an idle machine, single bench process (Criterion 95% CIs, stable within +/-1% run-to-run): ef 64/128/256/512 p50 = 71.8 / 130.5 / 235.8 / 433.3 us'; section 11.3 records the run as 'First reproducible run, VelesDB v3.3.0' on Apple Silicon, with recall@10 0.9012 / 0.9435 / 0.9659 / 0.9759. measured_version is the workspace version of that commit.",
      "measured_on": "2026-06-24 (date of commit 4b62338f, which records the run)",
      "measured_machine": "Apple Silicon (M-series, ARM64/NEON), otherwise idle, single bench process (docs/BENCHMARKS.md section 11.3)",
      "measured_version": "3.3.0",
      "stale_accepted": {
        "reason": "Published with its version and host in section 11.3 (v3.3.0, Apple Silicon); re-running SIFT1M builds a 1M index and needs the same idle host, not done in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "velesql_spec_explain_divergence_threshold",
      "file": "docs/VELESQL_SPEC.md",
      "must_contain": "diverges from actual time by more than a fixed 10×",
      "validation_command": "grep -n 'ratio > 10.0' crates/velesdb-cli/src/repl_query_cmds.rs",
      "executable": false,
      "source": "The EXPLAIN ANALYZE warning threshold: divergence_exceeds_threshold in crates/velesdb-cli/src/repl_query_cmds.rs returns ratio > 10.0, the ratio being the larger of estimated and actual time over the smaller.",
      "notes": "A threshold, not a measurement: the figure is the floor or bound named in source.",
      "measured_on": "n/a",
      "measured_machine": "n/a",
      "measured_version": "n/a",
      "last_updated": "2026-09-13"
    },
    {
      "id": "velesql_spec_sq8_anon_rss_falls_61",
      "file": "docs/VELESQL_SPEC.md",
      "must_contain": "anonymous RSS falls 61%",
      "validation_command": "cargo run -p velesdb-core --release --example resident_set --features persistence -- sq8",
      "executable": false,
      "source": "commit abc2eefc (2026-08-24, 'measure the resident set instead of claiming it'): examples/resident_set.rs, 100 000 x 768-d, one mode per process: Full RssAnon 385.0 MiB, total 517.9 MiB; SQ8 RssAnon 150.4 MiB, total 576.2 MiB - 'Anonymous RSS falls 61%. Total RSS *rises* 11%'. measured_version is the workspace version of that commit.",
      "measured_on": "2026-08-24",
      "measured_machine": "4-vCPU Linux container, 16 GiB RAM (docs/guides/QUANTIZATION.md 'Measured resident set')",
      "measured_version": "5.2.0",
      "stale_accepted": {
        "reason": "Taken 2026-08-24 on 5.2.0 and dated in docs/guides/QUANTIZATION.md; re-running resident_set wants the same 4-vCPU Linux container at rest, not done in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "velesql_spec_sq8_total_rss_rises_11",
      "file": "docs/VELESQL_SPEC.md",
      "must_contain": "while total RSS rises 11%",
      "validation_command": "cargo run -p velesdb-core --release --example resident_set --features persistence -- sq8",
      "executable": false,
      "source": "commit abc2eefc (2026-08-24, 'measure the resident set instead of claiming it'): examples/resident_set.rs, 100 000 x 768-d, one mode per process: Full RssAnon 385.0 MiB, total 517.9 MiB; SQ8 RssAnon 150.4 MiB, total 576.2 MiB - 'Anonymous RSS falls 61%. Total RSS *rises* 11%'. measured_version is the workspace version of that commit.",
      "measured_on": "2026-08-24",
      "measured_machine": "4-vCPU Linux container, 16 GiB RAM (docs/guides/QUANTIZATION.md 'Measured resident set')",
      "measured_version": "5.2.0",
      "stale_accepted": {
        "reason": "Taken 2026-08-24 on 5.2.0 and dated in docs/guides/QUANTIZATION.md; re-running resident_set wants the same 4-vCPU Linux container at rest, not done in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "business_scenarios_simd_cosine",
      "file": "docs/guides/BUSINESS_SCENARIOS.md",
      "must_contain": "| **Cosine** | 33.1 ns | AVX2 |",
      "validation_command": "cargo bench -p velesdb-core --bench simd_benchmark -- --noplot",
      "executable": false,
      "source": "commit dd0d2457 (2026-03-24, 'docs: final benchmark update - real measurements March 24, 2026 22:56'): 'All metrics re-measured sequentially on idle i9-14900KF, 64GB DDR5, Windows 11, Rust 1.92.0, Criterion.rs' - SIMD 768D: 'Cosine: 33.1 ns', 'Hamming: 35.8 ns', 'Jaccard: 35.1 ns'. measured_version is the workspace version of that commit.",
      "measured_on": "2026-03-24",
      "measured_machine": "Intel Core i9-14900KF, 64GB DDR5, Windows 11, Rust 1.92.0, idle (commit dd0d2457)",
      "measured_version": "1.7.0",
      "stale_accepted": {
        "reason": "Dated by commit dd0d2457 (March 24, 2026, 1.7.0); re-measuring belongs on the idle i9-14900KF reference machine, not done in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "business_scenarios_simd_euclidean",
      "file": "docs/guides/BUSINESS_SCENARIOS.md",
      "must_contain": "| **Euclidean** | 26.0 ns | AVX2 |",
      "validation_command": "cargo bench -p velesdb-core --bench simd_benchmark -- --noplot",
      "executable": false,
      "source": "docs/BENCHMARKS.md section 1, SIMD Kernel Latency: Euclidean 26.0 ns at 768D, set by commit 81b371ba (2026-04-03) under 'Measured April 3, 2026 on Intel Core i9-14900KF (24C/32T, AVX2+FMA), 64GB DDR5, Rust 1.94.1, Windows 11 Pro, sequential run on idle machine'. This row copies that figure. measured_version is the workspace version of 81b371ba.",
      "measured_on": "2026-04-03",
      "measured_machine": "Intel Core i9-14900KF (24C/32T, AVX2+FMA), 64GB DDR5, Windows 11 Pro, Rust 1.94.1, idle (docs/BENCHMARKS.md section 1)",
      "measured_version": "1.11.0",
      "stale_accepted": {
        "reason": "Dated in docs/BENCHMARKS.md section 1 (2026-04-03, 1.11.0); re-measuring belongs on the idle i9-14900KF reference machine, not done in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "business_scenarios_simd_dot",
      "file": "docs/guides/BUSINESS_SCENARIOS.md",
      "must_contain": "| **DotProduct** | 21.7 ns | AVX2 |",
      "validation_command": "cargo bench -p velesdb-core --bench simd_benchmark -- --noplot",
      "executable": false,
      "source": "commit df217860 (2026-03-27, 'docs: update README performance section with verified production metrics'): 'Core engine: Rust index-only for architecture reference (5K/768D: 55us HNSW search, 21.7ns dot product)'; 'Measured 2026-03-27, i9-14900KF, 64GB DDR5, Windows 11, Rust 1.92.' The April 3, 2026 run in docs/BENCHMARKS.md section 1 measured 21.8 ns for the same kernel. measured_version is the workspace version of df217860.",
      "measured_on": "2026-03-27",
      "measured_machine": "Intel Core i9-14900KF, 64GB DDR5, Windows 11, Rust 1.92 (commit df217860)",
      "measured_version": "1.7.2",
      "stale_accepted": {
        "reason": "Dated by commit df217860 (2026-03-27, 1.7.2); re-measuring belongs on the idle i9-14900KF reference machine, not done in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "business_scenarios_simd_hamming",
      "file": "docs/guides/BUSINESS_SCENARIOS.md",
      "must_contain": "| **Hamming** | **35.8 ns** | POPCNT |",
      "validation_command": "cargo bench -p velesdb-core --bench simd_benchmark -- --noplot",
      "executable": false,
      "source": "commit dd0d2457 (2026-03-24, 'docs: final benchmark update - real measurements March 24, 2026 22:56'): 'All metrics re-measured sequentially on idle i9-14900KF, 64GB DDR5, Windows 11, Rust 1.92.0, Criterion.rs' - SIMD 768D: 'Cosine: 33.1 ns', 'Hamming: 35.8 ns', 'Jaccard: 35.1 ns'. measured_version is the workspace version of that commit.",
      "measured_on": "2026-03-24",
      "measured_machine": "Intel Core i9-14900KF, 64GB DDR5, Windows 11, Rust 1.92.0, idle (commit dd0d2457)",
      "measured_version": "1.7.0",
      "stale_accepted": {
        "reason": "Dated by commit dd0d2457 (March 24, 2026, 1.7.0); re-measuring belongs on the idle i9-14900KF reference machine, not done in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "business_scenarios_simd_jaccard",
      "file": "docs/guides/BUSINESS_SCENARIOS.md",
      "must_contain": "| **Jaccard** | 35.1 ns | AVX2 |",
      "validation_command": "cargo bench -p velesdb-core --bench simd_benchmark -- --noplot",
      "executable": false,
      "source": "commit dd0d2457 (2026-03-24, 'docs: final benchmark update - real measurements March 24, 2026 22:56'): 'All metrics re-measured sequentially on idle i9-14900KF, 64GB DDR5, Windows 11, Rust 1.92.0, Criterion.rs' - SIMD 768D: 'Cosine: 33.1 ns', 'Hamming: 35.8 ns', 'Jaccard: 35.1 ns'. measured_version is the workspace version of that commit.",
      "measured_on": "2026-03-24",
      "measured_machine": "Intel Core i9-14900KF, 64GB DDR5, Windows 11, Rust 1.92.0, idle (commit dd0d2457)",
      "measured_version": "1.7.0",
      "stale_accepted": {
        "reason": "Dated by commit dd0d2457 (March 24, 2026, 1.7.0); re-measuring belongs on the idle i9-14900KF reference machine, not done in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "core_perf_simd_dot",
      "file": "docs/guides/CORE_PERFORMANCE.md",
      "must_contain": "| Dot product | **21.7 ns** |",
      "validation_command": "cargo bench -p velesdb-core --bench simd_benchmark -- --noplot",
      "executable": false,
      "source": "commit df217860 (2026-03-27, 'docs: update README performance section with verified production metrics'): 'Core engine: Rust index-only for architecture reference (5K/768D: 55us HNSW search, 21.7ns dot product)'; 'Measured 2026-03-27, i9-14900KF, 64GB DDR5, Windows 11, Rust 1.92.' The April 3, 2026 run in docs/BENCHMARKS.md section 1 measured 21.8 ns for the same kernel. measured_version is the workspace version of df217860.",
      "measured_on": "2026-03-27",
      "measured_machine": "Intel Core i9-14900KF, 64GB DDR5, Windows 11, Rust 1.92 (commit df217860)",
      "measured_version": "1.7.2",
      "stale_accepted": {
        "reason": "Dated by commit df217860 (2026-03-27, 1.7.2); re-measuring belongs on the idle i9-14900KF reference machine, not done in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "core_perf_simd_euclidean",
      "file": "docs/guides/CORE_PERFORMANCE.md",
      "must_contain": "| Euclidean distance | **26.0 ns** |",
      "validation_command": "cargo bench -p velesdb-core --bench simd_benchmark -- --noplot",
      "executable": false,
      "source": "docs/BENCHMARKS.md section 1, SIMD Kernel Latency: Euclidean 26.0 ns at 768D, set by commit 81b371ba (2026-04-03) under 'Measured April 3, 2026 on Intel Core i9-14900KF (24C/32T, AVX2+FMA), 64GB DDR5, Rust 1.94.1, Windows 11 Pro, sequential run on idle machine'. This row copies that figure. measured_version is the workspace version of 81b371ba.",
      "measured_on": "2026-04-03",
      "measured_machine": "Intel Core i9-14900KF (24C/32T, AVX2+FMA), 64GB DDR5, Windows 11 Pro, Rust 1.94.1, idle (docs/BENCHMARKS.md section 1)",
      "measured_version": "1.11.0",
      "stale_accepted": {
        "reason": "Dated in docs/BENCHMARKS.md section 1 (2026-04-03, 1.11.0); re-measuring belongs on the idle i9-14900KF reference machine, not done in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "core_perf_simd_cosine",
      "file": "docs/guides/CORE_PERFORMANCE.md",
      "must_contain": "| Cosine similarity | **33.1 ns** |",
      "validation_command": "cargo bench -p velesdb-core --bench simd_benchmark -- --noplot",
      "executable": false,
      "source": "commit dd0d2457 (2026-03-24, 'docs: final benchmark update - real measurements March 24, 2026 22:56'): 'All metrics re-measured sequentially on idle i9-14900KF, 64GB DDR5, Windows 11, Rust 1.92.0, Criterion.rs' - SIMD 768D: 'Cosine: 33.1 ns', 'Hamming: 35.8 ns', 'Jaccard: 35.1 ns'. measured_version is the workspace version of that commit.",
      "measured_on": "2026-03-24",
      "measured_machine": "Intel Core i9-14900KF, 64GB DDR5, Windows 11, Rust 1.92.0, idle (commit dd0d2457)",
      "measured_version": "1.7.0",
      "stale_accepted": {
        "reason": "Dated by commit dd0d2457 (March 24, 2026, 1.7.0); re-measuring belongs on the idle i9-14900KF reference machine, not done in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "core_perf_simd_hamming",
      "file": "docs/guides/CORE_PERFORMANCE.md",
      "must_contain": "| Hamming distance | **35.8 ns** |",
      "validation_command": "cargo bench -p velesdb-core --bench simd_benchmark -- --noplot",
      "executable": false,
      "source": "commit dd0d2457 (2026-03-24, 'docs: final benchmark update - real measurements March 24, 2026 22:56'): 'All metrics re-measured sequentially on idle i9-14900KF, 64GB DDR5, Windows 11, Rust 1.92.0, Criterion.rs' - SIMD 768D: 'Cosine: 33.1 ns', 'Hamming: 35.8 ns', 'Jaccard: 35.1 ns'. measured_version is the workspace version of that commit.",
      "measured_on": "2026-03-24",
      "measured_machine": "Intel Core i9-14900KF, 64GB DDR5, Windows 11, Rust 1.92.0, idle (commit dd0d2457)",
      "measured_version": "1.7.0",
      "stale_accepted": {
        "reason": "Dated by commit dd0d2457 (March 24, 2026, 1.7.0); re-measuring belongs on the idle i9-14900KF reference machine, not done in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "core_perf_simd_jaccard",
      "file": "docs/guides/CORE_PERFORMANCE.md",
      "must_contain": "| Jaccard similarity | **35.1 ns** |",
      "validation_command": "cargo bench -p velesdb-core --bench simd_benchmark -- --noplot",
      "executable": false,
      "source": "commit dd0d2457 (2026-03-24, 'docs: final benchmark update - real measurements March 24, 2026 22:56'): 'All metrics re-measured sequentially on idle i9-14900KF, 64GB DDR5, Windows 11, Rust 1.92.0, Criterion.rs' - SIMD 768D: 'Cosine: 33.1 ns', 'Hamming: 35.8 ns', 'Jaccard: 35.1 ns'. measured_version is the workspace version of that commit.",
      "measured_on": "2026-03-24",
      "measured_machine": "Intel Core i9-14900KF, 64GB DDR5, Windows 11, Rust 1.92.0, idle (commit dd0d2457)",
      "measured_version": "1.7.0",
      "stale_accepted": {
        "reason": "Dated by commit dd0d2457 (March 24, 2026, 1.7.0); re-measuring belongs on the idle i9-14900KF reference machine, not done in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "python_engine_simd_cosine",
      "file": "docs/guides/PYTHON_ENGINE_BENCHMARKS.md",
      "must_contain": "| Cosine | ~33.1 ns |",
      "validation_command": "cargo bench -p velesdb-core --bench simd_benchmark -- --noplot",
      "executable": false,
      "source": "commit dd0d2457 (2026-03-24, 'docs: final benchmark update - real measurements March 24, 2026 22:56'): 'All metrics re-measured sequentially on idle i9-14900KF, 64GB DDR5, Windows 11, Rust 1.92.0, Criterion.rs' - SIMD 768D: 'Cosine: 33.1 ns', 'Hamming: 35.8 ns', 'Jaccard: 35.1 ns'. measured_version is the workspace version of that commit.",
      "measured_on": "2026-03-24",
      "measured_machine": "Intel Core i9-14900KF, 64GB DDR5, Windows 11, Rust 1.92.0, idle (commit dd0d2457)",
      "measured_version": "1.7.0",
      "stale_accepted": {
        "reason": "Dated by commit dd0d2457 (March 24, 2026, 1.7.0); re-measuring belongs on the idle i9-14900KF reference machine, not done in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "python_engine_simd_euclidean",
      "file": "docs/guides/PYTHON_ENGINE_BENCHMARKS.md",
      "must_contain": "| Euclidean | ~26.0 ns |",
      "validation_command": "cargo bench -p velesdb-core --bench simd_benchmark -- --noplot",
      "executable": false,
      "source": "docs/BENCHMARKS.md section 1, SIMD Kernel Latency: Euclidean 26.0 ns at 768D, set by commit 81b371ba (2026-04-03) under 'Measured April 3, 2026 on Intel Core i9-14900KF (24C/32T, AVX2+FMA), 64GB DDR5, Rust 1.94.1, Windows 11 Pro, sequential run on idle machine'. This row copies that figure. measured_version is the workspace version of 81b371ba.",
      "measured_on": "2026-04-03",
      "measured_machine": "Intel Core i9-14900KF (24C/32T, AVX2+FMA), 64GB DDR5, Windows 11 Pro, Rust 1.94.1, idle (docs/BENCHMARKS.md section 1)",
      "measured_version": "1.11.0",
      "stale_accepted": {
        "reason": "Dated in docs/BENCHMARKS.md section 1 (2026-04-03, 1.11.0); re-measuring belongs on the idle i9-14900KF reference machine, not done in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "python_engine_simd_dot",
      "file": "docs/guides/PYTHON_ENGINE_BENCHMARKS.md",
      "must_contain": "| Dot Product | ~21.7 ns |",
      "validation_command": "cargo bench -p velesdb-core --bench simd_benchmark -- --noplot",
      "executable": false,
      "source": "commit df217860 (2026-03-27, 'docs: update README performance section with verified production metrics'): 'Core engine: Rust index-only for architecture reference (5K/768D: 55us HNSW search, 21.7ns dot product)'; 'Measured 2026-03-27, i9-14900KF, 64GB DDR5, Windows 11, Rust 1.92.' The April 3, 2026 run in docs/BENCHMARKS.md section 1 measured 21.8 ns for the same kernel. measured_version is the workspace version of df217860.",
      "measured_on": "2026-03-27",
      "measured_machine": "Intel Core i9-14900KF, 64GB DDR5, Windows 11, Rust 1.92 (commit df217860)",
      "measured_version": "1.7.2",
      "stale_accepted": {
        "reason": "Dated by commit df217860 (2026-03-27, 1.7.2); re-measuring belongs on the idle i9-14900KF reference machine, not done in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "python_engine_simd_hamming",
      "file": "docs/guides/PYTHON_ENGINE_BENCHMARKS.md",
      "must_contain": "| Hamming | ~35.8 ns |",
      "validation_command": "cargo bench -p velesdb-core --bench simd_benchmark -- --noplot",
      "executable": false,
      "source": "commit dd0d2457 (2026-03-24, 'docs: final benchmark update - real measurements March 24, 2026 22:56'): 'All metrics re-measured sequentially on idle i9-14900KF, 64GB DDR5, Windows 11, Rust 1.92.0, Criterion.rs' - SIMD 768D: 'Cosine: 33.1 ns', 'Hamming: 35.8 ns', 'Jaccard: 35.1 ns'. measured_version is the workspace version of that commit.",
      "measured_on": "2026-03-24",
      "measured_machine": "Intel Core i9-14900KF, 64GB DDR5, Windows 11, Rust 1.92.0, idle (commit dd0d2457)",
      "measured_version": "1.7.0",
      "stale_accepted": {
        "reason": "Dated by commit dd0d2457 (March 24, 2026, 1.7.0); re-measuring belongs on the idle i9-14900KF reference machine, not done in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "architecture_simd_dot",
      "file": "docs/reference/ARCHITECTURE.md",
      "must_contain": "| Dot Product | AVX2 FMA | **21.7 ns** |",
      "validation_command": "cargo bench -p velesdb-core --bench simd_benchmark -- --noplot",
      "executable": false,
      "source": "commit df217860 (2026-03-27, 'docs: update README performance section with verified production metrics'): 'Core engine: Rust index-only for architecture reference (5K/768D: 55us HNSW search, 21.7ns dot product)'; 'Measured 2026-03-27, i9-14900KF, 64GB DDR5, Windows 11, Rust 1.92.' The April 3, 2026 run in docs/BENCHMARKS.md section 1 measured 21.8 ns for the same kernel. measured_version is the workspace version of df217860.",
      "measured_on": "2026-03-27",
      "measured_machine": "Intel Core i9-14900KF, 64GB DDR5, Windows 11, Rust 1.92 (commit df217860)",
      "measured_version": "1.7.2",
      "stale_accepted": {
        "reason": "Dated by commit df217860 (2026-03-27, 1.7.2); re-measuring belongs on the idle i9-14900KF reference machine, not done in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "architecture_simd_euclidean",
      "file": "docs/reference/ARCHITECTURE.md",
      "must_contain": "| Euclidean | AVX2 FMA | **26.0 ns** |",
      "validation_command": "cargo bench -p velesdb-core --bench simd_benchmark -- --noplot",
      "executable": false,
      "source": "docs/BENCHMARKS.md section 1, SIMD Kernel Latency: Euclidean 26.0 ns at 768D, set by commit 81b371ba (2026-04-03) under 'Measured April 3, 2026 on Intel Core i9-14900KF (24C/32T, AVX2+FMA), 64GB DDR5, Rust 1.94.1, Windows 11 Pro, sequential run on idle machine'. This row copies that figure. measured_version is the workspace version of 81b371ba.",
      "measured_on": "2026-04-03",
      "measured_machine": "Intel Core i9-14900KF (24C/32T, AVX2+FMA), 64GB DDR5, Windows 11 Pro, Rust 1.94.1, idle (docs/BENCHMARKS.md section 1)",
      "measured_version": "1.11.0",
      "stale_accepted": {
        "reason": "Dated in docs/BENCHMARKS.md section 1 (2026-04-03, 1.11.0); re-measuring belongs on the idle i9-14900KF reference machine, not done in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "architecture_simd_cosine",
      "file": "docs/reference/ARCHITECTURE.md",
      "must_contain": "| Cosine | AVX2 4-acc, single-sqrt finish | **33.1 ns** |",
      "validation_command": "cargo bench -p velesdb-core --bench simd_benchmark -- --noplot",
      "executable": false,
      "source": "commit dd0d2457 (2026-03-24, 'docs: final benchmark update - real measurements March 24, 2026 22:56'): 'All metrics re-measured sequentially on idle i9-14900KF, 64GB DDR5, Windows 11, Rust 1.92.0, Criterion.rs' - SIMD 768D: 'Cosine: 33.1 ns', 'Hamming: 35.8 ns', 'Jaccard: 35.1 ns'. measured_version is the workspace version of that commit.",
      "measured_on": "2026-03-24",
      "measured_machine": "Intel Core i9-14900KF, 64GB DDR5, Windows 11, Rust 1.92.0, idle (commit dd0d2457)",
      "measured_version": "1.7.0",
      "stale_accepted": {
        "reason": "Dated by commit dd0d2457 (March 24, 2026, 1.7.0); re-measuring belongs on the idle i9-14900KF reference machine, not done in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "architecture_simd_hamming",
      "file": "docs/reference/ARCHITECTURE.md",
      "must_contain": "| Hamming | AVX2 FP-domain 4-acc | **35.8 ns** |",
      "validation_command": "cargo bench -p velesdb-core --bench simd_benchmark -- --noplot",
      "executable": false,
      "source": "commit dd0d2457 (2026-03-24, 'docs: final benchmark update - real measurements March 24, 2026 22:56'): 'All metrics re-measured sequentially on idle i9-14900KF, 64GB DDR5, Windows 11, Rust 1.92.0, Criterion.rs' - SIMD 768D: 'Cosine: 33.1 ns', 'Hamming: 35.8 ns', 'Jaccard: 35.1 ns'. measured_version is the workspace version of that commit.",
      "measured_on": "2026-03-24",
      "measured_machine": "Intel Core i9-14900KF, 64GB DDR5, Windows 11, Rust 1.92.0, idle (commit dd0d2457)",
      "measured_version": "1.7.0",
      "stale_accepted": {
        "reason": "Dated by commit dd0d2457 (March 24, 2026, 1.7.0); re-measuring belongs on the idle i9-14900KF reference machine, not done in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "architecture_simd_jaccard",
      "file": "docs/reference/ARCHITECTURE.md",
      "must_contain": "| Jaccard | AVX2 4-acc | **35.1 ns** |",
      "validation_command": "cargo bench -p velesdb-core --bench simd_benchmark -- --noplot",
      "executable": false,
      "source": "commit dd0d2457 (2026-03-24, 'docs: final benchmark update - real measurements March 24, 2026 22:56'): 'All metrics re-measured sequentially on idle i9-14900KF, 64GB DDR5, Windows 11, Rust 1.92.0, Criterion.rs' - SIMD 768D: 'Cosine: 33.1 ns', 'Hamming: 35.8 ns', 'Jaccard: 35.1 ns'. measured_version is the workspace version of that commit. The i9-14900KF has no AVX-512: at 768D, `jaccard_simd` (crates/velesdb-core/src/simd_native/dispatch/hamming.rs at dd0d2457) takes its `SimdLevel::Avx512 | SimdLevel::Avx2` arm to `jaccard_avx2`, documented 'AVX2 Jaccard with 4 accumulators', so the row names that kernel.",
      "measured_on": "2026-03-24",
      "measured_machine": "Intel Core i9-14900KF, 64GB DDR5, Windows 11, Rust 1.92.0, idle (commit dd0d2457)",
      "measured_version": "1.7.0",
      "stale_accepted": {
        "reason": "Dated by commit dd0d2457 (March 24, 2026, 1.7.0); re-measuring belongs on the idle i9-14900KF reference machine, not done in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "architecture_hnsw_index_only_55us",
      "file": "docs/reference/ARCHITECTURE.md",
      "must_contain": "| Search k=10 (10K vectors, 768D, HNSW index-only) | ~55 µs |",
      "validation_command": "cargo bench -p velesdb-core --bench hnsw_benchmark -- hnsw_search_latency",
      "executable": false,
      "source": "commit df217860 (2026-03-27, 'docs: update README performance section with verified production metrics'): 'Core engine: Rust index-only for architecture reference (5K/768D: 55us HNSW search, 21.7ns dot product)'; 'Measured 2026-03-27, i9-14900KF, 64GB DDR5, Windows 11, Rust 1.92.' hnsw_benchmark's hnsw_search_latency group inserted 10,000 vectors at that commit; mirrors core_perf_hnsw_index_only_55us. measured_version is the workspace version of df217860.",
      "measured_on": "2026-03-27",
      "measured_machine": "Intel Core i9-14900KF, 64GB DDR5, Windows 11, Rust 1.92 (commit df217860)",
      "measured_version": "1.7.2",
      "stale_accepted": {
        "reason": "Dated by commit df217860 (2026-03-27, 1.7.2, i9-14900KF). The same day commit 126d5a9a recorded 47.0 us for the same bench on the same machine; re-measuring belongs on the idle i9-14900KF reference machine, not done in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "concurrency_locking_search_scaling_967",
      "file": "docs/guides/CONCURRENCY_LOCKING.md",
      "must_contain": "8 concurrent searches reached ~5× the throughput",
      "validation_command": "cargo bench -p velesdb-core --bench scalability_benchmark -- concurrent_queries",
      "executable": false,
      "source": "commit 185660d3 (2026-06-01, 'document concurrent-search scalability analysis for issue #967'): 'The scalability_benchmark on M5 Pro shows ~5x throughput at 8 threads instead of the ideal 8x'; the rustdoc names the bench (scalability_benchmark::concurrent_queries, M5 Pro, 100 K x 768-dim). measured_version is the workspace version of that commit.",
      "measured_on": "2026-06-01 (date of commit 185660d3, which records the run; the rustdoc dates the closed investigation 2026-06-03)",
      "measured_machine": "Apple M5 Pro (as commit 185660d3 and the rustdoc name it)",
      "measured_version": "1.16.0",
      "stale_accepted": {
        "reason": "Recorded by commit 185660d3 (1.16.0) as the closing measurement of #967; re-running it needs the same M5 Pro at rest, not done in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "context_compiler_tail_phase_6_6x",
      "file": "docs/guides/CONTEXT_COMPILER.md",
      "must_contain": "**up to 6.6×** in the verification/wrap-up phase",
      "validation_command": "cd examples/real-session-benchmark && node long-session.mjs",
      "executable": false,
      "source": "examples/real-session-benchmark/README.md, long-session reading (node long-session.mjs, written by commit e34dd2cb, 2026-07-18): 'in the verification/wrap-up phase (turns 27-36, mostly re-reads) the gap widens to ~234 vs ~35/turn (6.6x) - that tail rate is phase-specific and is never used for the projections'. Same campaign as readme_long_session_window_growth. measured_version is the workspace version of that commit.",
      "measured_on": "2026-07-18 (date of commit e34dd2cb, which records the run)",
      "measured_machine": "platform-independent (deterministic token counts; every turn compiled twice, byte-identical)",
      "measured_version": "3.12.0",
      "stale_accepted": {
        "reason": "Deterministic token counts published with their campaign in examples/real-session-benchmark; re-running long-session.mjs needs the Node binding built from the current tree, not done in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "context_compiler_headroom_1_7x",
      "file": "docs/guides/CONTEXT_COMPILER.md",
      "must_contain": "**1.7× more headroom**",
      "validation_command": "cd examples/real-session-benchmark && node long-session.mjs",
      "executable": false,
      "source": "examples/real-session-benchmark/README.md, long-session reading (node long-session.mjs, recorded by 5b4df82a, 2026-07-18): 'over the FULL measured session the raw arm grows ~555 tokens/turn versus ~333 compiled (1.7× slower window consumption)'. Same run as readme_long_session_window_growth. measured_version is the workspace version of that commit.",
      "measured_on": "2026-07-18 (date of commit 5b4df82a, which records the run)",
      "measured_machine": "platform-independent (deterministic token counts; every turn compiled twice, byte-identical)",
      "measured_version": "3.12.0",
      "stale_accepted": {
        "reason": "Deterministic token counts published with their campaign in examples/real-session-benchmark; re-running long-session.mjs needs the Node binding built from the current tree, not done in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "core_perf_columnstore_string_75x",
      "file": "docs/guides/CORE_PERFORMANCE.md",
      "must_contain": "string equality up to **75x**",
      "validation_command": "cargo bench -p velesdb-core --bench column_filter_benchmark",
      "executable": false,
      "source": "the v1.6.0 release commit 92cc35bd (2026-03-20): 'Re-run all 6 benchmark suites sequentially on idle machine (i9-14900KF, 64GB DDR5, Windows 11, Rust 1.92.0) - March 19, 2026'; it set docs/BENCHMARKS.md section 6's String Equality table (100K rows: ColumnStore 46.5 us vs JSON scan 3.50 ms, 75x). measured_version is the workspace version of that commit.",
      "measured_on": "2026-03-19",
      "measured_machine": "Intel Core i9-14900KF, 64GB DDR5, Windows 11, Rust 1.92.0, idle (release commit 92cc35bd)",
      "measured_version": "1.6.0",
      "stale_accepted": {
        "reason": "Recorded by the v1.6.0 release run (2026-03-19, 1.6.0); re-measuring belongs on the idle i9-14900KF reference machine, not done in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "migrate_embeddings_reembed_23x",
      "file": "docs/guides/MIGRATE_EMBEDDINGS.md",
      "must_contain": "**23×** a re-insertion",
      "validation_command": "cargo test -p velesdb-memory -- --ignored the_embedder_dominates_only_when_the_model_changes",
      "executable": false,
      "source": "commit 8ef88b6a (#1816, 2026-08-05): 'Measured on aarch64-apple-darwin, debug, machine at rest, bge-m3 at 1024 dimensions, 32 facts, one call per fact: re-embedding (model CHANGED) 88461.8 us/fact, reinsertion (vectors REUSED) 3899.9 us/fact, ratio: x23'; restated in crates/velesdb-memory/src/migration/diagnosis/capabilities.rs by c779494d (#1818). measured_version is the workspace version of that commit.",
      "measured_on": "2026-08-05 (date of commit 8ef88b6a, which records the run)",
      "measured_machine": "aarch64-apple-darwin, debug build, machine at rest (commit 8ef88b6a names no model)",
      "measured_version": "4.2.0",
      "stale_accepted": {
        "reason": "One configuration recorded with the change it informed (8ef88b6a, 4.2.0); re-running it needs bge-m3 served locally and a machine at rest, not done in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "migrate_embeddings_guard_direct_count",
      "file": "docs/guides/MIGRATE_EMBEDDINGS.md",
      "must_contain": "`NativeStore::count` measured 5.45 ns",
      "validation_command": "cargo bench -p velesdb-memory --bench generation_gate_benchmark -- --noplot",
      "executable": false,
      "source": "commit 8f130bf8 (#1936, 2026-08-15) wrote this paragraph with Criterion's generation_gate_benchmark, run 'on the same machine' as the section's online-path evidence (arm64 macOS 26.5.2, optimized build): direct NativeStore::count 5.45 ns, guarded MemoryService::fact_count 6.18 ns; the 0.73 ns increment is their difference. measured_version is the workspace version of that commit.",
      "measured_on": "2026-08-15 (date of commit 8f130bf8, which records the run)",
      "measured_machine": "arm64 macOS 26.5.2, optimized build (docs/guides/MIGRATE_EMBEDDINGS.md 'Online-path evidence')",
      "measured_version": "5.0.0",
      "stale_accepted": {
        "reason": "Recorded with the change it measured (8f130bf8, 5.0.0); re-running the bench on that machine was not done in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "migrate_embeddings_guard_fact_count",
      "file": "docs/guides/MIGRATE_EMBEDDINGS.md",
      "must_contain": "`MemoryService::fact_count` 6.18 ns, an absolute 0.73 ns increment",
      "validation_command": "cargo bench -p velesdb-memory --bench generation_gate_benchmark -- --noplot",
      "executable": false,
      "source": "commit 8f130bf8 (#1936, 2026-08-15) wrote this paragraph with Criterion's generation_gate_benchmark, run 'on the same machine' as the section's online-path evidence (arm64 macOS 26.5.2, optimized build): direct NativeStore::count 5.45 ns, guarded MemoryService::fact_count 6.18 ns; the 0.73 ns increment is their difference. measured_version is the workspace version of that commit.",
      "measured_on": "2026-08-15 (date of commit 8f130bf8, which records the run)",
      "measured_machine": "arm64 macOS 26.5.2, optimized build (docs/guides/MIGRATE_EMBEDDINGS.md 'Online-path evidence')",
      "measured_version": "5.0.0",
      "stale_accepted": {
        "reason": "Recorded with the change it measured (8f130bf8, 5.0.0); re-running the bench on that machine was not done in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "migrate_embeddings_capture_ratio",
      "file": "docs/guides/MIGRATE_EMBEDDINGS.md",
      "must_contain": "1.88× baseline",
      "validation_command": "cargo test -p velesdb-memory --release --test online_migration_process reports_steady_state_capture_and_replay_overhead -- --ignored --test-threads=1 --nocapture",
      "executable": false,
      "source": "docs/guides/MIGRATE_EMBEDDINGS.md 'Online-path evidence', written by commit 8f130bf8 (#1936, 2026-08-15): 'Measured 2026-08-15 on arm64 macOS 26.5.2, optimized build, offline hash embedder at 384 dimensions, 64 seeded facts and 16 measured writes per arm': baseline_us_per_write=10231.3, capture_us_per_write=19214.4, capture_ratio=1.88. measured_version is the workspace version of that commit.",
      "measured_on": "2026-08-15",
      "measured_machine": "arm64 macOS 26.5.2, optimized build",
      "measured_version": "5.0.0",
      "stale_accepted": {
        "reason": "Recorded with the change it measured (8f130bf8, 5.0.0); re-running the ignored test on that machine was not done in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "migrate_embeddings_reinsert_16us",
      "file": "docs/guides/MIGRATE_EMBEDDINGS.md",
      "must_contain": "measured at ~16 µs/fact",
      "validation_command": "cargo test -p velesdb-memory --release -- --ignored batched_write_cost_stays_flat_as_the_volume_doubles",
      "executable": false,
      "source": "crates/velesdb-memory/src/migration/tests/performance.rs, batched_write_cost_stays_flat_as_the_volume_doubles: its table, recorded by commit a4950038 (2026-08-03) after the BM25 fix #1797 ('Measured twice, --release, batch = 1 024, upsert_bulk, distinct vectors, dimension 4, on an idle machine'), falls from 24.1 us/fact at 1 000 facts to 16.3 at 16 000; '~16 us/fact' is the 16 000-fact figure. measured_version is the workspace version of that commit.",
      "measured_on": "2026-08-03 (date of commit a4950038, which records the run)",
      "measured_machine": "unknown (the doc comment says 'an idle machine'; commit a4950038 names none)",
      "measured_version": "4.2.0",
      "stale_accepted": {
        "reason": "The test's own recorded measurement (a4950038, 4.2.0); re-running it writes 31 000 facts on a machine at rest, not done in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "quantization_sq8_total_rss_rises_11",
      "file": "docs/guides/QUANTIZATION.md",
      "must_contain": "Total RSS *rises* 11% over `Full`",
      "validation_command": "cargo run -p velesdb-core --release --example resident_set --features persistence -- sq8",
      "executable": false,
      "source": "commit abc2eefc (2026-08-24, 'measure the resident set instead of claiming it'): examples/resident_set.rs, 100 000 x 768-d, one mode per process: Full RssAnon 385.0 MiB, total 517.9 MiB; SQ8 RssAnon 150.4 MiB, total 576.2 MiB - 'Anonymous RSS falls 61%. Total RSS *rises* 11%'. measured_version is the workspace version of that commit.",
      "measured_on": "2026-08-24",
      "measured_machine": "4-vCPU Linux container, 16 GiB RAM (docs/guides/QUANTIZATION.md 'Measured resident set')",
      "measured_version": "5.2.0",
      "stale_accepted": {
        "reason": "Taken 2026-08-24 on 5.2.0 and dated in docs/guides/QUANTIZATION.md; re-running resident_set wants the same 4-vCPU Linux container at rest, not done in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "quantization_measured_resident_set_section",
      "file": "docs/guides/QUANTIZATION.md",
      "must_contain": "## Measured resident set",
      "covers_section": true,
      "validation_command": "cargo run -p velesdb-core --release --example resident_set --features persistence -- sq8",
      "executable": false,
      "source": "commit abc2eefc (2026-08-24, 'measure the resident set instead of claiming it'): examples/resident_set.rs, 100 000 x 768-d, one mode per process: Full RssAnon 385.0 MiB, total 517.9 MiB; SQ8 RssAnon 150.4 MiB, total 576.2 MiB - 'Anonymous RSS falls 61%. Total RSS *rises* 11%'. The fill-time table adc85f5b (#2128) added the same day is another run, registered under its own heading by quantization_arena_fill_time_section. measured_version is the workspace version of that commit.",
      "measured_on": "2026-08-24",
      "measured_machine": "4-vCPU Linux container, 16 GiB RAM (docs/guides/QUANTIZATION.md 'Measured resident set')",
      "measured_version": "5.2.0",
      "stale_accepted": {
        "reason": "Taken 2026-08-24 on 5.2.0 and dated in docs/guides/QUANTIZATION.md; re-running resident_set wants the same 4-vCPU Linux container at rest, not done in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "quantization_arena_fill_time_section",
      "file": "docs/guides/QUANTIZATION.md",
      "must_contain": "### Arena fill time",
      "covers_section": true,
      "validation_command": "cargo run --release --example resident_set --features persistence -- write",
      "executable": false,
      "source": "commit adc85f5b (#2128, 2026-08-24, 'decide the arena is not configurable, and say why') added the `write` mode of examples/resident_set.rs and this table: filling 100 000 x 768-d into each backing, no graph and no quantizer - heap 58-62 ms, stable across runs; file-mapped 0.30-1.35 s, rising with each consecutive run (five runs). The commit names no machine. measured_version is the workspace version of that commit.",
      "measured_on": "2026-08-24 (date of commit adc85f5b, which records the run)",
      "measured_machine": "unknown (commit adc85f5b names none)",
      "measured_version": "5.2.0",
      "stale_accepted": {
        "reason": "Recorded with the decision it informed (adc85f5b, 5.2.0); re-running the `write` mode five times was not done in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "quantization_cold_page_rerank_section",
      "file": "docs/guides/QUANTIZATION.md",
      "must_contain": "### Cold-page re-rank cost",
      "covers_section": true,
      "validation_command": "cargo run -p velesdb-core --release --example resident_set --features persistence -- sq8",
      "executable": false,
      "source": "commit abc2eefc (2026-08-24, 'measure the resident set instead of claiming it'): 'Cold-page cost, 100 scattered vectors with every dimension read: warm 0.14 ms, pages dropped with page cache warm 0.55 ms, pages dropped with page cache dropped 8-10 ms'. measured_version is the workspace version of that commit.",
      "measured_on": "2026-08-24",
      "measured_machine": "4-vCPU Linux container, 16 GiB RAM (docs/guides/QUANTIZATION.md 'Measured resident set')",
      "measured_version": "5.2.0",
      "stale_accepted": {
        "reason": "Taken 2026-08-24 on 5.2.0 and dated in docs/guides/QUANTIZATION.md; re-running resident_set wants the same 4-vCPU Linux container at rest, not done in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "use_cases_target_aggregation_overhead_5ms",
      "file": "docs/guides/USE_CASES.md",
      "must_contain": "| Aggregation overhead | < 5ms |",
      "validation_command": "none: no benchmark or test in the repository measures this target",
      "executable": false,
      "source": "docs/guides/USE_CASES.md: a use case's latency target, in a table headed 'Target (not measured)'.",
      "notes": "A target, not a measurement: the figure is the goal named in source.",
      "measured_on": "n/a",
      "measured_machine": "n/a",
      "measured_version": "n/a",
      "last_updated": "2026-09-13"
    },
    {
      "id": "timing_results_tldr_section",
      "file": "docs/quickstart/timing-results.md",
      "must_contain": "## TL;DR",
      "covers_section": true,
      "validation_command": "bash scripts/dx-timing/run_all.sh",
      "executable": false,
      "source": "benchmarks/dx-timing/results-2026-04-29T07-06-34Z.json (3 runs per scenario; medians python 4.95 s, rust 25.4 s, node 0.48 s, server 45.84 s; worst runs 5.66, 30.25, 0.74, 46.29 s), committed with this document by f06468bd (2026-04-29), whose harness pinned velesdb-core@1.13.7 and velesdb-server@1.13.7. d5e82db9 (2026-05-01) later relabeled those pins 1.14.2 without re-running; the JSON is unchanged since.",
      "measured_on": "2026-04-29",
      "measured_machine": "MINGW64_NT-10.0-26200 (Windows / Docker Desktop), x86_64 (the JSON's host field)",
      "measured_version": "1.13.7",
      "stale_accepted": {
        "reason": "A frozen onboarding measurement kept with the release it installed (1.13.7), as the page says; re-running the harness installs a current release and would be a new measurement, not done in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "timing_results_worst_python",
      "file": "docs/quickstart/timing-results.md",
      "must_contain": "| Python | 5.66 s |",
      "validation_command": "bash scripts/dx-timing/run_all.sh",
      "executable": false,
      "source": "benchmarks/dx-timing/results-2026-04-29T07-06-34Z.json (3 runs per scenario; medians python 4.95 s, rust 25.4 s, node 0.48 s, server 45.84 s; worst runs 5.66, 30.25, 0.74, 46.29 s), committed with this document by f06468bd (2026-04-29), whose harness pinned velesdb-core@1.13.7 and velesdb-server@1.13.7. d5e82db9 (2026-05-01) later relabeled those pins 1.14.2 without re-running; the JSON is unchanged since.",
      "measured_on": "2026-04-29",
      "measured_machine": "MINGW64_NT-10.0-26200 (Windows / Docker Desktop), x86_64 (the JSON's host field)",
      "measured_version": "1.13.7",
      "stale_accepted": {
        "reason": "A frozen onboarding measurement kept with the release it installed (1.13.7), as the page says; re-running the harness installs a current release and would be a new measurement, not done in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "timing_results_worst_rust",
      "file": "docs/quickstart/timing-results.md",
      "must_contain": "| Rust | 30.25 s |",
      "validation_command": "bash scripts/dx-timing/run_all.sh",
      "executable": false,
      "source": "benchmarks/dx-timing/results-2026-04-29T07-06-34Z.json (3 runs per scenario; medians python 4.95 s, rust 25.4 s, node 0.48 s, server 45.84 s; worst runs 5.66, 30.25, 0.74, 46.29 s), committed with this document by f06468bd (2026-04-29), whose harness pinned velesdb-core@1.13.7 and velesdb-server@1.13.7. d5e82db9 (2026-05-01) later relabeled those pins 1.14.2 without re-running; the JSON is unchanged since.",
      "measured_on": "2026-04-29",
      "measured_machine": "MINGW64_NT-10.0-26200 (Windows / Docker Desktop), x86_64 (the JSON's host field)",
      "measured_version": "1.13.7",
      "stale_accepted": {
        "reason": "A frozen onboarding measurement kept with the release it installed (1.13.7), as the page says; re-running the harness installs a current release and would be a new measurement, not done in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "timing_results_worst_typescript",
      "file": "docs/quickstart/timing-results.md",
      "must_contain": "| TypeScript | 0.74 s |",
      "validation_command": "bash scripts/dx-timing/run_all.sh",
      "executable": false,
      "source": "benchmarks/dx-timing/results-2026-04-29T07-06-34Z.json (3 runs per scenario; medians python 4.95 s, rust 25.4 s, node 0.48 s, server 45.84 s; worst runs 5.66, 30.25, 0.74, 46.29 s), committed with this document by f06468bd (2026-04-29), whose harness pinned velesdb-core@1.13.7 and velesdb-server@1.13.7. d5e82db9 (2026-05-01) later relabeled those pins 1.14.2 without re-running; the JSON is unchanged since.",
      "measured_on": "2026-04-29",
      "measured_machine": "MINGW64_NT-10.0-26200 (Windows / Docker Desktop), x86_64 (the JSON's host field)",
      "measured_version": "1.13.7",
      "stale_accepted": {
        "reason": "A frozen onboarding measurement kept with the release it installed (1.13.7), as the page says; re-running the harness installs a current release and would be a new measurement, not done in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "timing_results_worst_server",
      "file": "docs/quickstart/timing-results.md",
      "must_contain": "| Server | 46.29 s |",
      "validation_command": "bash scripts/dx-timing/run_all.sh",
      "executable": false,
      "source": "benchmarks/dx-timing/results-2026-04-29T07-06-34Z.json (3 runs per scenario; medians python 4.95 s, rust 25.4 s, node 0.48 s, server 45.84 s; worst runs 5.66, 30.25, 0.74, 46.29 s), committed with this document by f06468bd (2026-04-29), whose harness pinned velesdb-core@1.13.7 and velesdb-server@1.13.7. d5e82db9 (2026-05-01) later relabeled those pins 1.14.2 without re-running; the JSON is unchanged since.",
      "measured_on": "2026-04-29",
      "measured_machine": "MINGW64_NT-10.0-26200 (Windows / Docker Desktop), x86_64 (the JSON's host field)",
      "measured_version": "1.13.7",
      "stale_accepted": {
        "reason": "A frozen onboarding measurement kept with the release it installed (1.13.7), as the page says; re-running the harness installs a current release and would be a new measurement, not done in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "performance_slo_upsert_bulk_ratio_v172",
      "file": "docs/reference/PERFORMANCE_SLO.md",
      "must_contain": "dropped from ~19x to ~1x on local benchmarks (i9-14900KF, 10K/384D)",
      "validation_command": "cargo bench -p velesdb-core --bench bulk_insert_v2_benchmark",
      "executable": false,
      "source": "commit 024e4a9b (2026-03-25, 'docs: complete v1.7.2 documentation with upsert lock contention fix') and CHANGELOG.md [1.7.2]: 'Measured on i9-14900KF (10K vectors, 384D): upsert throughput rose from ~808 vec/s to ~16,151 vec/s, closing the gap with upsert_bulk() from 19x to ~1x'. measured_version is the workspace version of that commit.",
      "measured_on": "2026-03-25 (date of commit 024e4a9b, which records the run; the line states no date)",
      "measured_machine": "Intel Core i9-14900KF (as the CHANGELOG entry names it)",
      "measured_version": "1.7.2",
      "stale_accepted": {
        "reason": "Recorded with the v1.7.2 change it measured (2026-03-25); re-running it belongs on the i9-14900KF, not done in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "performance_slo_smoke_search_v172",
      "file": "docs/reference/PERFORMANCE_SLO.md",
      "must_contain": "`smoke_search/10k_128d_k10`: 191.73 µs (local machine",
      "validation_command": "cargo bench -p velesdb-core --bench smoke_test",
      "executable": false,
      "source": "release commit 399fc2fc (2026-03-25, 'chore: release v1.7.2') wrote this line with its local validation against the then-current CI baseline, on the i9-14900KF the paragraph names; the section keeps it as a historical record of the v1.7.2 pass. measured_version is the workspace version of that commit.",
      "measured_on": "2026-03-25 (date of release commit 399fc2fc, which records the run)",
      "measured_machine": "Intel Core i9-14900KF (as the paragraph names it)",
      "measured_version": "1.7.2",
      "stale_accepted": {
        "reason": "A historical record of the v1.7.2 pass, labelled so by its section; re-running smoke_test on the i9-14900KF was not done in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "tuning_upsert_gap_19x",
      "file": "docs/guides/TUNING_GUIDE.md",
      "must_contain": "closed the throughput gap between `upsert()` and `upsert_bulk()` from ~19x to ~1x",
      "validation_command": "cargo bench -p velesdb-core --bench bulk_insert_v2_benchmark",
      "executable": false,
      "source": "commit 024e4a9b (2026-03-25, 'docs: complete v1.7.2 documentation with upsert lock contention fix') and CHANGELOG.md [1.7.2]: 'Measured on i9-14900KF (10K vectors, 384D): upsert throughput rose from ~808 vec/s to ~16,151 vec/s, closing the gap with upsert_bulk() from 19x to ~1x'. measured_version is the workspace version of that commit.",
      "measured_on": "2026-03-25 (date of commit 024e4a9b, which records the run; the line states no date)",
      "measured_machine": "Intel Core i9-14900KF (as the CHANGELOG entry names it)",
      "measured_version": "1.7.2",
      "stale_accepted": {
        "reason": "Recorded with the v1.7.2 change it measured (2026-03-25); re-running it belongs on the i9-14900KF, not done in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "simd_perf_scaling_128",
      "file": "docs/reference/SIMD_PERFORMANCE.md",
      "must_contain": "| 128 | 8.1ns | 5.4ns |",
      "validation_command": "cargo bench -p velesdb-core --bench simd_benchmark -- --noplot",
      "executable": false,
      "source": "the v1.6.0 release commit 92cc35bd (2026-03-20): 'Re-run all 6 benchmark suites sequentially on idle machine (i9-14900KF, 64GB DDR5, Windows 11, Rust 1.92.0) - March 19, 2026'; it set this row of the dimension-scaling table (cosine and dot product). measured_version is the workspace version of that commit.",
      "measured_on": "2026-03-19",
      "measured_machine": "Intel Core i9-14900KF, 64GB DDR5, Windows 11, Rust 1.92.0, idle (release commit 92cc35bd)",
      "measured_version": "1.6.0",
      "stale_accepted": {
        "reason": "Recorded by the v1.6.0 release run (2026-03-19, 1.6.0); re-measuring belongs on the idle i9-14900KF reference machine, not done in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "simd_perf_scaling_384",
      "file": "docs/reference/SIMD_PERFORMANCE.md",
      "must_contain": "| 384 | 20.1ns | 12.0ns |",
      "validation_command": "cargo bench -p velesdb-core --bench simd_benchmark -- --noplot",
      "executable": false,
      "source": "the v1.6.0 release commit 92cc35bd (2026-03-20): 'Re-run all 6 benchmark suites sequentially on idle machine (i9-14900KF, 64GB DDR5, Windows 11, Rust 1.92.0) - March 19, 2026'; it set this row of the dimension-scaling table (cosine and dot product). measured_version is the workspace version of that commit.",
      "measured_on": "2026-03-19",
      "measured_machine": "Intel Core i9-14900KF, 64GB DDR5, Windows 11, Rust 1.92.0, idle (release commit 92cc35bd)",
      "measured_version": "1.6.0",
      "stale_accepted": {
        "reason": "Recorded by the v1.6.0 release run (2026-03-19, 1.6.0); re-measuring belongs on the idle i9-14900KF reference machine, not done in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "simd_perf_scaling_1536",
      "file": "docs/reference/SIMD_PERFORMANCE.md",
      "must_contain": "| 1536 | 69.0ns | 43.8ns |",
      "validation_command": "cargo bench -p velesdb-core --bench simd_benchmark -- --noplot",
      "executable": false,
      "source": "the v1.6.0 release commit 92cc35bd (2026-03-20): 'Re-run all 6 benchmark suites sequentially on idle machine (i9-14900KF, 64GB DDR5, Windows 11, Rust 1.92.0) - March 19, 2026'; it set this row of the dimension-scaling table (cosine and dot product). measured_version is the workspace version of that commit.",
      "measured_on": "2026-03-19",
      "measured_machine": "Intel Core i9-14900KF, 64GB DDR5, Windows 11, Rust 1.92.0, idle (release commit 92cc35bd)",
      "measured_version": "1.6.0",
      "stale_accepted": {
        "reason": "Recorded by the v1.6.0 release run (2026-03-19, 1.6.0); re-measuring belongs on the idle i9-14900KF reference machine, not done in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "simd_perf_scaling_3072",
      "file": "docs/reference/SIMD_PERFORMANCE.md",
      "must_contain": "| 3072 | 112.2ns | 91.2ns |",
      "validation_command": "cargo bench -p velesdb-core --bench simd_benchmark -- --noplot",
      "executable": false,
      "source": "the v1.6.0 release commit 92cc35bd (2026-03-20): 'Re-run all 6 benchmark suites sequentially on idle machine (i9-14900KF, 64GB DDR5, Windows 11, Rust 1.92.0) - March 19, 2026'; it set this row of the dimension-scaling table (cosine and dot product). measured_version is the workspace version of that commit.",
      "measured_on": "2026-03-19",
      "measured_machine": "Intel Core i9-14900KF, 64GB DDR5, Windows 11, Rust 1.92.0, idle (release commit 92cc35bd)",
      "measured_version": "1.6.0",
      "stale_accepted": {
        "reason": "Recorded by the v1.6.0 release run (2026-03-19, 1.6.0); re-measuring belongs on the idle i9-14900KF reference machine, not done in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "simd_perf_scaling_768",
      "file": "docs/reference/SIMD_PERFORMANCE.md",
      "must_contain": "| 768 | 33.1ns | 19.8ns |",
      "validation_command": "cargo bench -p velesdb-core --bench simd_benchmark -- --noplot",
      "executable": false,
      "source": "two runs on the same machine: the cosine cell is commit dd0d2457's (2026-03-24, 'All metrics re-measured sequentially on idle i9-14900KF, 64GB DDR5, Windows 11, Rust 1.92.0' - 'Cosine: 33.1 ns'), the dot-product cell commit 126d5a9a's (2026-03-27, 'measured March 27, 2026, i9-14900KF, AVX2, Rust 1.92' - 'SIMD Dot Product (768D): 17.6ns -> 19.8ns'), which set this row. measured_version is the workspace version of 126d5a9a (dd0d2457 was 1.7.0).",
      "measured_on": "2026-03-24 (cosine) and 2026-03-27 (dot product)",
      "measured_machine": "Intel Core i9-14900KF, 64GB DDR5, Windows 11, Rust 1.92, idle (commits dd0d2457 and 126d5a9a)",
      "measured_version": "1.7.2",
      "stale_accepted": {
        "reason": "Dated by commits dd0d2457 and 126d5a9a (March 2026, 1.7.x); re-measuring belongs on the idle i9-14900KF reference machine, not done in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "context_savings_compile_10_fragments",
      "file": "crates/velesdb-memory/examples/context_savings/README.md",
      "must_contain": "| 10 plain fragments | ~60 µs |",
      "validation_command": "cargo bench -p velesdb-memory --no-default-features --features context",
      "executable": false,
      "source": "commit 4fa4fde8 (2026-07-17, EPIC-P-070) committed benches/context_compiler_benchmark.rs with this table (10 plain fragments ~60 us); the README section says 'release profile, Apple Silicon'. measured_version is the workspace version of that commit.",
      "measured_on": "2026-07-17 (date of commit 4fa4fde8, which records the run)",
      "measured_machine": "Apple Silicon (the README's words; commit 4fa4fde8 names no model)",
      "measured_version": "3.12.0",
      "stale_accepted": {
        "reason": "Recorded by commit 4fa4fde8 (3.12.0); re-running the criterion bench needs a machine at rest, not done in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "context_savings_compile_100_fragments",
      "file": "crates/velesdb-memory/examples/context_savings/README.md",
      "must_contain": "| 100 plain fragments | ~590 µs |",
      "validation_command": "cargo bench -p velesdb-memory --no-default-features --features context",
      "executable": false,
      "source": "commit 4fa4fde8 (2026-07-17, EPIC-P-070) committed benches/context_compiler_benchmark.rs with this table (100 plain fragments ~590 us); the README section says 'release profile, Apple Silicon'. measured_version is the workspace version of that commit.",
      "measured_on": "2026-07-17 (date of commit 4fa4fde8, which records the run)",
      "measured_machine": "Apple Silicon (the README's words; commit 4fa4fde8 names no model)",
      "measured_version": "3.12.0",
      "stale_accepted": {
        "reason": "Recorded by commit 4fa4fde8 (3.12.0); re-running the criterion bench needs a machine at rest, not done in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "context_savings_compile_500_fragments",
      "file": "crates/velesdb-memory/examples/context_savings/README.md",
      "must_contain": "| 500 plain fragments | ~2.98 ms |",
      "validation_command": "cargo bench -p velesdb-memory --no-default-features --features context",
      "executable": false,
      "source": "commit 4fa4fde8 (2026-07-17, EPIC-P-070) committed benches/context_compiler_benchmark.rs with this table (500 plain fragments ~2.98 ms); the README section says 'release profile, Apple Silicon'. measured_version is the workspace version of that commit.",
      "measured_on": "2026-07-17 (date of commit 4fa4fde8, which records the run)",
      "measured_machine": "Apple Silicon (the README's words; commit 4fa4fde8 names no model)",
      "measured_version": "3.12.0",
      "stale_accepted": {
        "reason": "Recorded by commit 4fa4fde8 (3.12.0); re-running the criterion bench needs a machine at rest, not done in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "context_savings_compile_1000_fragments",
      "file": "crates/velesdb-memory/examples/context_savings/README.md",
      "must_contain": "| 1,000 plain fragments | ~5.97 ms |",
      "validation_command": "cargo bench -p velesdb-memory --no-default-features --features context",
      "executable": false,
      "source": "commit 4fa4fde8 (2026-07-17, EPIC-P-070) committed benches/context_compiler_benchmark.rs with this table (1,000 plain fragments ~5.97 ms); the README section says 'release profile, Apple Silicon'. measured_version is the workspace version of that commit.",
      "measured_on": "2026-07-17 (date of commit 4fa4fde8, which records the run)",
      "measured_machine": "Apple Silicon (the README's words; commit 4fa4fde8 names no model)",
      "measured_version": "3.12.0",
      "stale_accepted": {
        "reason": "Recorded by commit 4fa4fde8 (3.12.0); re-running the criterion bench needs a machine at rest, not done in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "context_savings_compile_duplicate_heavy",
      "file": "crates/velesdb-memory/examples/context_savings/README.md",
      "must_contain": "| 1,000 fragments, 1/3 exact dupes | ~5.32 ms |",
      "validation_command": "cargo bench -p velesdb-memory --no-default-features --features context",
      "executable": false,
      "source": "commit 4fa4fde8 (2026-07-17, EPIC-P-070) committed benches/context_compiler_benchmark.rs with this table (1,000 fragments with 1/3 exact duplicates ~5.32 ms); the README section says 'release profile, Apple Silicon'. measured_version is the workspace version of that commit.",
      "measured_on": "2026-07-17 (date of commit 4fa4fde8, which records the run)",
      "measured_machine": "Apple Silicon (the README's words; commit 4fa4fde8 names no model)",
      "measured_version": "3.12.0",
      "stale_accepted": {
        "reason": "Recorded by commit 4fa4fde8 (3.12.0); re-running the criterion bench needs a machine at rest, not done in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "context_savings_compile_budget_tight",
      "file": "crates/velesdb-memory/examples/context_savings/README.md",
      "must_contain": "| Budget pressure | 200 fragments, budget 200 tok (tight) | ~1.07 ms |",
      "validation_command": "cargo bench -p velesdb-memory --no-default-features --features context",
      "executable": false,
      "source": "commit 4fa4fde8 (2026-07-17, EPIC-P-070) committed benches/context_compiler_benchmark.rs with this table (200 fragments, budget 200 tok ~1.07 ms); the README section says 'release profile, Apple Silicon'. measured_version is the workspace version of that commit.",
      "measured_on": "2026-07-17 (date of commit 4fa4fde8, which records the run)",
      "measured_machine": "Apple Silicon (the README's words; commit 4fa4fde8 names no model)",
      "measured_version": "3.12.0",
      "stale_accepted": {
        "reason": "Recorded by commit 4fa4fde8 (3.12.0); re-running the criterion bench needs a machine at rest, not done in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "context_savings_compile_budget_generous",
      "file": "crates/velesdb-memory/examples/context_savings/README.md",
      "must_contain": "| 200 fragments, budget 1,000,000 tok (generous) | ~1.19 ms |",
      "validation_command": "cargo bench -p velesdb-memory --no-default-features --features context",
      "executable": false,
      "source": "commit 4fa4fde8 (2026-07-17, EPIC-P-070) committed benches/context_compiler_benchmark.rs with this table (200 fragments, budget 1,000,000 tok ~1.19 ms); the README section says 'release profile, Apple Silicon'. measured_version is the workspace version of that commit.",
      "measured_on": "2026-07-17 (date of commit 4fa4fde8, which records the run)",
      "measured_machine": "Apple Silicon (the README's words; commit 4fa4fde8 names no model)",
      "measured_version": "3.12.0",
      "stale_accepted": {
        "reason": "Recorded by commit 4fa4fde8 (3.12.0); re-running the criterion bench needs a machine at rest, not done in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "context_savings_compile_oversized_tight_budget",
      "file": "crates/velesdb-memory/examples/context_savings/README.md",
      "must_contain": "| Oversized fragment | one ~100 KB fragment, budget 2,000 tok | ~4.65 ms |",
      "validation_command": "cargo bench -p velesdb-memory --no-default-features --features context",
      "executable": false,
      "source": "commit 4fa4fde8 (2026-07-17, EPIC-P-070) committed benches/context_compiler_benchmark.rs with this table (one ~100 KB fragment, budget 2,000 tok ~4.65 ms); the README section says 'release profile, Apple Silicon'. measured_version is the workspace version of that commit.",
      "measured_on": "2026-07-17 (date of commit 4fa4fde8, which records the run)",
      "measured_machine": "Apple Silicon (the README's words; commit 4fa4fde8 names no model)",
      "measured_version": "3.12.0",
      "stale_accepted": {
        "reason": "Recorded by commit 4fa4fde8 (3.12.0); re-running the criterion bench needs a machine at rest, not done in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "context_savings_compile_oversized_generous_budget",
      "file": "crates/velesdb-memory/examples/context_savings/README.md",
      "must_contain": "| one ~100 KB fragment, budget 1,000,000 tok | ~5.80 ms |",
      "validation_command": "cargo bench -p velesdb-memory --no-default-features --features context",
      "executable": false,
      "source": "commit 4fa4fde8 (2026-07-17, EPIC-P-070) committed benches/context_compiler_benchmark.rs with this table (one ~100 KB fragment, budget 1,000,000 tok ~5.80 ms); the README section says 'release profile, Apple Silicon'. measured_version is the workspace version of that commit.",
      "measured_on": "2026-07-17 (date of commit 4fa4fde8, which records the run)",
      "measured_machine": "Apple Silicon (the README's words; commit 4fa4fde8 names no model)",
      "measured_version": "3.12.0",
      "stale_accepted": {
        "reason": "Recorded by commit 4fa4fde8 (3.12.0); re-running the criterion bench needs a machine at rest, not done in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "node_readme_compile_cost",
      "file": "crates/velesdb-node/README.md",
      "must_contain": "**0.7 ms** stateless, 24.5 ms with source/event persistence on",
      "validation_command": "cd crates/velesdb-memory/examples/context_savings/real_measures && node agent_session.mjs",
      "executable": false,
      "source": "crates/velesdb-memory/examples/context_savings/README.md, agent-session output block (committed with its token figures by ebdbf015, 2026-07-26): 'compile latency (stateless: store_sources/record_events off): mean 0.7 ms' and 'compile latency (with source/event persistence, default): mean 24.5 ms'; mirrors context_savings_stateless_compile_latency.",
      "measured_on": "2026-07-26",
      "measured_machine": "unknown (latency depends on the machine, which the run does not record; its token figures are platform-independent)",
      "measured_version": "4.0.0",
      "stale_accepted": {
        "reason": "Printed by the same run as the 82.5% corpus figure (4.0.0) and published as that run's output; re-running agent_session.mjs needs the Node binding built from the current tree, not done in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "cli_explain_divergence_threshold",
      "file": "crates/velesdb-cli/src/repl_query_cmds.rs",
      "must_contain": "by more than the 10× threshold",
      "validation_command": "grep -n 'ratio > 10.0' crates/velesdb-cli/src/repl_query_cmds.rs",
      "executable": false,
      "source": "The function's own threshold: divergence_exceeds_threshold returns ratio > 10.0, the ratio being the larger of estimated and actual time over the smaller.",
      "notes": "A threshold, not a measurement: the figure is the floor or bound named in source.",
      "measured_on": "n/a",
      "measured_machine": "n/a",
      "measured_version": "n/a",
      "last_updated": "2026-09-13"
    },
    {
      "id": "rustdoc_file_arena_flush_fadvise",
      "file": "crates/velesdb-core/src/contiguous_file_arena.rs",
      "must_contain": "Measured both ways at 100 000 ×",
      "validation_command": "cargo run -p velesdb-core --release --example resident_set --features persistence -- sq8",
      "executable": false,
      "source": "commit abc2eefc (2026-08-24, 'measure the resident set instead of claiming it'): 'The flush stays because it makes a later POSIX_FADV_DONTNEED deterministic (8.2 ms with, 6.9 ms without, the difference being pages writeback had not reached)'; the doc comment gives the same figures at 100 000 x 768. measured_version is the workspace version of that commit.",
      "measured_on": "2026-08-24",
      "measured_machine": "4-vCPU Linux container, 16 GiB RAM (docs/guides/QUANTIZATION.md 'Measured resident set')",
      "measured_version": "5.2.0",
      "stale_accepted": {
        "reason": "Taken 2026-08-24 on 5.2.0 and dated in docs/guides/QUANTIZATION.md; re-running resident_set wants the same 4-vCPU Linux container at rest, not done in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "rustdoc_hnsw_concurrent_search_5x",
      "file": "crates/velesdb-core/src/index/hnsw/native/graph/mod.rs",
      "must_contain": "8 parallel searches achieve ~5× throughput instead of the ideal ~8×",
      "validation_command": "cargo bench -p velesdb-core --bench scalability_benchmark -- concurrent_queries",
      "executable": false,
      "source": "commit 185660d3 (2026-06-01, 'document concurrent-search scalability analysis for issue #967'): 'The scalability_benchmark on M5 Pro shows ~5x throughput at 8 threads instead of the ideal 8x'; the rustdoc names the bench (scalability_benchmark::concurrent_queries, M5 Pro, 100 K x 768-dim). measured_version is the workspace version of that commit.",
      "measured_on": "2026-06-01 (date of commit 185660d3, which records the run; the rustdoc dates the closed investigation 2026-06-03)",
      "measured_machine": "Apple M5 Pro (as commit 185660d3 and the rustdoc name it)",
      "measured_version": "1.16.0",
      "stale_accepted": {
        "reason": "Recorded by commit 185660d3 (1.16.0) as the closing measurement of #967; re-running it needs the same M5 Pro at rest, not done in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "rustdoc_hnsw_concurrent_search_5x_observed",
      "file": "crates/velesdb-core/src/index/hnsw/native/graph/mod.rs",
      "must_contain": "observed ~5× rather than ~8× throughput scaling",
      "validation_command": "cargo bench -p velesdb-core --bench scalability_benchmark -- concurrent_queries",
      "executable": false,
      "source": "commit 185660d3 (2026-06-01, 'document concurrent-search scalability analysis for issue #967'): 'The scalability_benchmark on M5 Pro shows ~5x throughput at 8 threads instead of the ideal 8x'; the rustdoc names the bench (scalability_benchmark::concurrent_queries, M5 Pro, 100 K x 768-dim). measured_version is the workspace version of that commit.",
      "measured_on": "2026-06-01 (date of commit 185660d3, which records the run; the rustdoc dates the closed investigation 2026-06-03)",
      "measured_machine": "Apple M5 Pro (as commit 185660d3 and the rustdoc name it)",
      "measured_version": "1.16.0",
      "stale_accepted": {
        "reason": "Recorded by commit 185660d3 (1.16.0) as the closing measurement of #967; re-running it needs the same M5 Pro at rest, not done in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "rustdoc_storage_mode_rss_61_11",
      "file": "crates/velesdb-core/src/quantization/mod.rs",
      "must_contain": "falls 61% (385 -> 150 MiB) while total RSS rises 11%",
      "validation_command": "cargo run -p velesdb-core --release --example resident_set --features persistence -- sq8",
      "executable": false,
      "source": "commit abc2eefc (2026-08-24, 'measure the resident set instead of claiming it'): examples/resident_set.rs, 100 000 x 768-d, one mode per process: Full RssAnon 385.0 MiB, total 517.9 MiB; SQ8 RssAnon 150.4 MiB, total 576.2 MiB - 'Anonymous RSS falls 61%. Total RSS *rises* 11%'. measured_version is the workspace version of that commit.",
      "measured_on": "2026-08-24",
      "measured_machine": "4-vCPU Linux container, 16 GiB RAM (docs/guides/QUANTIZATION.md 'Measured resident set')",
      "measured_version": "5.2.0",
      "stale_accepted": {
        "reason": "Taken 2026-08-24 on 5.2.0 and dated in docs/guides/QUANTIZATION.md; re-running resident_set wants the same 4-vCPU Linux container at rest, not done in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "rustdoc_sparse_43x_uniform",
      "file": "crates/velesdb-core/src/sparse_index/search/mod.rs",
      "must_contain": "crossing 100k docs cost **43×** per query on",
      "validation_command": "cargo bench -p velesdb-core --bench sparse_posting_scale",
      "executable": false,
      "source": "commit d4e29e17 (#2182, 2026-09-01): on the sparse_posting_scale bench (30k vocab, two passes per configuration) MaxScore query latency at 110k docs was 43x the linear arm under uniform queries and 14x under skewed SPLADE-shaped ones, flat at ~13x in a crossover sweep up to 400k docs; full evidence on #2177. measured_version is the workspace version of that commit.",
      "measured_on": "2026-09-01 (date of commit d4e29e17, which records the run)",
      "measured_machine": "unknown (commit d4e29e17 does not name it)",
      "measured_version": "5.2.0",
      "stale_accepted": {
        "reason": "Recorded with the change it justified (d4e29e17, on 5.2.0), and the comment names #2177 as its source; re-running sparse_posting_scale needs a quiet machine, not made in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "rustdoc_sparse_14x_skewed",
      "file": "crates/velesdb-core/src/sparse_index/search/mod.rs",
      "must_contain": "uniform queries and **14×** on skewed SPLADE-shaped ones",
      "validation_command": "cargo bench -p velesdb-core --bench sparse_posting_scale",
      "executable": false,
      "source": "commit d4e29e17 (#2182, 2026-09-01): on the sparse_posting_scale bench (30k vocab, two passes per configuration) MaxScore query latency at 110k docs was 43x the linear arm under uniform queries and 14x under skewed SPLADE-shaped ones, flat at ~13x in a crossover sweep up to 400k docs; full evidence on #2177. measured_version is the workspace version of that commit.",
      "measured_on": "2026-09-01 (date of commit d4e29e17, which records the run)",
      "measured_machine": "unknown (commit d4e29e17 does not name it)",
      "measured_version": "5.2.0",
      "stale_accepted": {
        "reason": "Recorded with the change it justified (d4e29e17, on 5.2.0), and the comment names #2177 as its source; re-running sparse_posting_scale needs a quiet machine, not made in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "rustdoc_sparse_13x_flat_sweep",
      "file": "crates/velesdb-core/src/sparse_index/search/mod.rs",
      "must_contain": "ratio stayed flat (~13×) in a crossover sweep up to 400k docs",
      "validation_command": "cargo bench -p velesdb-core --bench sparse_posting_scale",
      "executable": false,
      "source": "commit d4e29e17 (#2182, 2026-09-01): on the sparse_posting_scale bench (30k vocab, two passes per configuration) MaxScore query latency at 110k docs was 43x the linear arm under uniform queries and 14x under skewed SPLADE-shaped ones, flat at ~13x in a crossover sweep up to 400k docs; full evidence on #2177. measured_version is the workspace version of that commit.",
      "measured_on": "2026-09-01 (date of commit d4e29e17, which records the run)",
      "measured_machine": "unknown (commit d4e29e17 does not name it)",
      "measured_version": "5.2.0",
      "stale_accepted": {
        "reason": "Recorded with the change it justified (d4e29e17, on 5.2.0), and the comment names #2177 as its source; re-running sparse_posting_scale needs a quiet machine, not made in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "rustdoc_sparse_13x_at_400k",
      "file": "crates/velesdb-core/src/sparse_index/search/mod.rs",
      "must_contain": "still ~13× ahead at 400k (#2177)",
      "validation_command": "cargo bench -p velesdb-core --bench sparse_posting_scale",
      "executable": false,
      "source": "commit d4e29e17 (#2182, 2026-09-01): on the sparse_posting_scale bench (30k vocab, two passes per configuration) MaxScore query latency at 110k docs was 43x the linear arm under uniform queries and 14x under skewed SPLADE-shaped ones, flat at ~13x in a crossover sweep up to 400k docs; full evidence on #2177. measured_version is the workspace version of that commit.",
      "measured_on": "2026-09-01 (date of commit d4e29e17, which records the run)",
      "measured_machine": "unknown (commit d4e29e17 does not name it)",
      "measured_version": "5.2.0",
      "stale_accepted": {
        "reason": "Recorded with the change it justified (d4e29e17, on 5.2.0), and the comment names #2177 as its source; re-running sparse_posting_scale needs a quiet machine, not made in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "rustdoc_memory_release_gain_6pct",
      "file": "crates/velesdb-memory/src/migration/tests/performance.rs",
      "must_contain": "(`--release` gains ~6%)",
      "validation_command": "cargo test -p velesdb-memory --release -- --ignored the_per_point_write_cost_is_attributed_to_payload_or_vector",
      "executable": false,
      "source": "the ignored diagnostic test's doc comment, committed by 83784a92 (2026-08-03, #1762), whose message says its 'Diagnostic benchmarks are #[ignore]d and carry their measurements in their doc comments': three hypotheses refuted by measurement before the test was written, the last being the debug profile ('--release gains ~6%'). measured_version is the workspace version of that commit.",
      "measured_on": "2026-08-03 (date of commit 83784a92, which records the run)",
      "measured_machine": "unknown (neither the doc comment nor commit 83784a92 names it)",
      "measured_version": "4.2.0",
      "stale_accepted": {
        "reason": "The test's own recorded measurement (83784a92, 4.2.0); re-running it writes 2 000 facts and, per its #[ignore], wants a machine at rest - not done in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "rustdoc_memory_write_vs_read_950x",
      "file": "crates/velesdb-memory/src/migration/tests/performance.rs",
      "must_contain": "is ~950x the cost of reading",
      "validation_command": "cargo test -p velesdb-memory --release -- --ignored past_the_ceiling_the_offset_walk_truncates_and_the_cursor_does_not",
      "executable": false,
      "source": "the ignored diagnostic test's doc comment, committed by 83784a92 (2026-08-03, #1762), whose message says its 'Diagnostic benchmarks are #[ignore]d and carry their measurements in their doc comments': 'seeding took 1242 s for 100 001 facts - 12.4 ms per fact, against 13.1 US per fact to read one back'; ~950x is 12.4 ms over 13.1 us. measured_version is the workspace version of that commit.",
      "measured_on": "2026-08-03 (date of commit 83784a92, which records the run)",
      "measured_machine": "unknown (neither the doc comment nor commit 83784a92 names it)",
      "measured_version": "4.2.0",
      "stale_accepted": {
        "reason": "The test's own recorded measurement (83784a92, 4.2.0); re-running it writes 100 001 facts and, per its #[ignore], wants a machine at rest - not done in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "rustdoc_memory_payload_write_5us",
      "file": "crates/velesdb-memory/src/migration/tests/performance.rs",
      "must_contain": "costs ~5 us per point",
      "validation_command": "cargo test -p velesdb-memory --release -- --ignored the_per_point_write_cost_is_attributed_to_payload_or_vector",
      "executable": false,
      "source": "the_per_point_write_cost_is_attributed_to_payload_or_vector's doc comment, committed by 83784a92 (2026-08-03, #1762): 'Measured 2026-08-03 (macOS, --release, n = 2 000, batch = 1 024)': numeric-only payload 15.1 us/fact against 10.2 with no payload, so ~5 us is their difference; the BM25 WAL's ~3 403 is 3418.7 - 15.1. measured_version is the workspace version of that commit.",
      "measured_on": "2026-08-03 (date of commit 83784a92, which records the run)",
      "measured_machine": "macOS, --release (the doc comment names no model)",
      "measured_version": "4.2.0",
      "stale_accepted": {
        "reason": "The test's own recorded measurement (83784a92, 4.2.0); re-running it writes 2 000 facts three times on a machine at rest, not done in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "rustdoc_memory_text_wal_3457us",
      "file": "crates/velesdb-memory/src/migration/tests/performance.rs",
      "must_contain": "(3 457.8 µs/fact with text against 15.1",
      "validation_command": "cargo test -p velesdb-memory --release -- --ignored the_per_point_write_cost_is_attributed_to_payload_or_vector",
      "executable": false,
      "source": "the_embedder_dominates_only_when_the_model_changes's doc comment quotes the table of the_per_point_write_cost_is_attributed_to_payload_or_vector, committed by 83784a92 (2026-08-03, #1762): 'Measured 2026-08-03 (macOS, --release, n = 2 000, batch = 1 024)': full payload 3457.8 us/fact, numeric only 15.1. measured_version is the workspace version of that commit.",
      "measured_on": "2026-08-03 (date of commit 83784a92, which records the run)",
      "measured_machine": "macOS, --release (the doc comment names no model)",
      "measured_version": "4.2.0",
      "stale_accepted": {
        "reason": "The test's own recorded measurement (83784a92, 4.2.0); re-running it writes 2 000 facts three times on a machine at rest, not done in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "rustdoc_memory_rebuild_before_1797",
      "file": "crates/velesdb-memory/src/migration/tests/performance.rs",
      "must_contain": "| 1 000  | 3 555.8 us",
      "validation_command": "cargo test -p velesdb-memory --release -- --ignored batched_write_cost_stays_flat_as_the_volume_doubles",
      "executable": false,
      "source": "batched_write_cost_stays_flat_as_the_volume_doubles's doc comment: its 'before #1797' column, recorded by commit a4950038 (2026-08-03) beside the 'after #1797' one ('Measured twice, --release, batch = 1 024, upsert_bulk, distinct vectors, dimension 4, on an idle machine'), the same code either side of the BM25 fix that removed one fsync per document: 3 555.8 us/fact at 1 000 facts. measured_version is the workspace version of that commit.",
      "measured_on": "2026-08-03 (date of commit a4950038, which records the run)",
      "measured_machine": "unknown (the doc comment says 'an idle machine'; commit a4950038 names none)",
      "measured_version": "4.2.0",
      "stale_accepted": {
        "reason": "The test's own recorded measurement (a4950038, 4.2.0); re-running it writes 31 000 facts on a machine at rest, not done in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "rustdoc_memory_rebuild_after_1797",
      "file": "crates/velesdb-memory/src/migration/tests/performance.rs",
      "must_contain": "**24.1 us** | 41 475",
      "validation_command": "cargo test -p velesdb-memory --release -- --ignored batched_write_cost_stays_flat_as_the_volume_doubles",
      "executable": false,
      "source": "batched_write_cost_stays_flat_as_the_volume_doubles's doc comment: its 'after #1797' column, recorded by commit a4950038 (2026-08-03) ('Measured twice, --release, batch = 1 024, upsert_bulk, distinct vectors, dimension 4, on an idle machine'): 24.1 us/fact at 1 000 facts, 41 475 facts/s. measured_version is the workspace version of that commit.",
      "measured_on": "2026-08-03 (date of commit a4950038, which records the run)",
      "measured_machine": "unknown (the doc comment says 'an idle machine'; commit a4950038 names none)",
      "measured_version": "4.2.0",
      "stale_accepted": {
        "reason": "The test's own recorded measurement (a4950038, 4.2.0); re-running it writes 31 000 facts on a machine at rest, not done in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "rustdoc_memory_reembed_per_fact",
      "file": "crates/velesdb-memory/src/migration/tests/performance.rs",
      "must_contain": "re-embedding (model CHANGED)    | 88 462 µs",
      "validation_command": "cargo test -p velesdb-memory -- --ignored the_embedder_dominates_only_when_the_model_changes",
      "executable": false,
      "source": "commit 8ef88b6a (#1816, 2026-08-05): 'Measured on aarch64-apple-darwin, debug, machine at rest, bge-m3 at 1024 dimensions, 32 facts, one call per fact: re-embedding (model CHANGED) 88461.8 us/fact, reinsertion (vectors REUSED) 3899.9 us/fact, ratio: x23'; restated in crates/velesdb-memory/src/migration/diagnosis/capabilities.rs by c779494d (#1818). measured_version is the workspace version of that commit.",
      "measured_on": "2026-08-05 (date of commit 8ef88b6a, which records the run)",
      "measured_machine": "aarch64-apple-darwin, debug build, machine at rest (commit 8ef88b6a names no model)",
      "measured_version": "4.2.0",
      "stale_accepted": {
        "reason": "One configuration recorded with the change it informed (8ef88b6a, 4.2.0); re-running it needs bge-m3 served locally and a machine at rest, not done in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "rustdoc_memory_reinsert_per_fact",
      "file": "crates/velesdb-memory/src/migration/tests/performance.rs",
      "must_contain": "reinsertion (vectors REUSED)    |  3 900 µs",
      "validation_command": "cargo test -p velesdb-memory -- --ignored the_embedder_dominates_only_when_the_model_changes",
      "executable": false,
      "source": "commit 8ef88b6a (#1816, 2026-08-05): 'Measured on aarch64-apple-darwin, debug, machine at rest, bge-m3 at 1024 dimensions, 32 facts, one call per fact: re-embedding (model CHANGED) 88461.8 us/fact, reinsertion (vectors REUSED) 3899.9 us/fact, ratio: x23'; restated in crates/velesdb-memory/src/migration/diagnosis/capabilities.rs by c779494d (#1818). measured_version is the workspace version of that commit.",
      "measured_on": "2026-08-05 (date of commit 8ef88b6a, which records the run)",
      "measured_machine": "aarch64-apple-darwin, debug build, machine at rest (commit 8ef88b6a names no model)",
      "measured_version": "4.2.0",
      "stale_accepted": {
        "reason": "One configuration recorded with the change it informed (8ef88b6a, 4.2.0); re-running it needs bge-m3 served locally and a machine at rest, not done in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "rustdoc_memory_former_blocker_reinsert_16us",
      "file": "crates/velesdb-memory/src/migration/diagnosis/capabilities.rs",
      "must_contain": "used to quote `16.3 us/fact`",
      "validation_command": "none: commit ac5c3d6e states the figure but names neither the test nor the machine that produced it",
      "executable": false,
      "source": "commit ac5c3d6e (#1762 PR A, 2026-08-03): its message keeps 'what was measured (16.3 us/fact of store cost)' apart from the embedder cost, and the blocker text it wrote read '16.3 us/fact to re-insert, once #1797 removed the per-document fsync'. The blocker now states the #1816 measurement instead (rustdoc_memory_reinsert_per_fact); the doc comment quotes the old figure as history. measured_version is the workspace version of that commit.",
      "measured_on": "2026-08-03 (date of commit ac5c3d6e, which records the figure)",
      "measured_machine": "unknown (commit ac5c3d6e names none)",
      "measured_version": "4.2.0",
      "stale_accepted": {
        "reason": "A superseded figure the doc comment quotes as history; no test in the tree reproduces it, and the #1816 measurement that replaced it is registered on its own.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "rustdoc_memory_former_blocker_reinsert_16us_compared",
      "file": "crates/velesdb-memory/src/migration/tests/performance.rs",
      "must_contain": "`16.3 us/fact",
      "validation_command": "none: commit ac5c3d6e states the figure but names neither the test nor the machine that produced it",
      "executable": false,
      "source": "commit ac5c3d6e (#1762 PR A, 2026-08-03): its message keeps 'what was measured (16.3 us/fact of store cost)' apart from the embedder cost, and the blocker text it wrote read '16.3 us/fact to re-insert, once #1797 removed the per-document fsync'. The blocker now states the #1816 measurement instead (rustdoc_memory_reinsert_per_fact); the doc comment quotes the old figure as history. measured_version is the workspace version of that commit.",
      "notes": "must_contain is part of both lines that quote the figure: the blocker's former wording, and the comparison an embedding time was wrongly held against.",
      "measured_on": "2026-08-03 (date of commit ac5c3d6e, which records the figure)",
      "measured_machine": "unknown (commit ac5c3d6e names none)",
      "measured_version": "4.2.0",
      "stale_accepted": {
        "reason": "A superseded figure the doc comment quotes as history; no test in the tree reproduces it, and the #1816 measurement that replaced it is registered on its own.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "rustdoc_memory_reinsert_no_text_460us",
      "file": "crates/velesdb-memory/src/migration/tests/performance.rs",
      "must_contain": "reinsertion drops to 460 µs/fact and the ratio rises to ×194",
      "validation_command": "cargo test -p velesdb-memory -- --ignored the_embedder_dominates_only_when_the_model_changes",
      "executable": false,
      "source": "the_embedder_dominates_only_when_the_model_changes's doc comment, committed by 8ef88b6a (#1816, 2026-08-05): 'Probed the same day, same 1024 dimensions, payload carrying NO text: reinsertion drops to 460 us/fact and the ratio rises to x194', on the machine of the table above it (aarch64-apple-darwin, debug, machine at rest). measured_version is the workspace version of that commit.",
      "measured_on": "2026-08-05 (date of commit 8ef88b6a, which records the run)",
      "measured_machine": "aarch64-apple-darwin, debug build, machine at rest (commit 8ef88b6a names no model)",
      "measured_version": "4.2.0",
      "stale_accepted": {
        "reason": "One configuration recorded with the change it informed (8ef88b6a, 4.2.0); re-running it needs bge-m3 served locally and a machine at rest, not done in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "rustdoc_memory_cursor_walk_13us",
      "file": "crates/velesdb-memory/src/migration/tests/performance.rs",
      "must_contain": "100 001 rows in 1.31 s (13.1 us/fact)",
      "validation_command": "cargo test -p velesdb-memory --release -- --ignored past_the_ceiling_the_offset_walk_truncates_and_the_cursor_does_not",
      "executable": false,
      "source": "past_the_ceiling_the_offset_walk_truncates_and_the_cursor_does_not's doc comment, committed by 83784a92 (2026-08-03, #1762): 'Measured 2026-08-03 (macOS, 100 001 facts seeded in 1242 s)': the cursor walk read 100 001 rows in 1.31 s, 13.1 us per fact. measured_version is the workspace version of that commit.",
      "measured_on": "2026-08-03 (date of commit 83784a92, which records the run)",
      "measured_machine": "macOS (the doc comment names no model)",
      "measured_version": "4.2.0",
      "stale_accepted": {
        "reason": "The test's own recorded measurement (83784a92, 4.2.0); re-running it writes 100 001 facts and, per its #[ignore], wants a machine at rest - not done in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "rustdoc_trigram_target_10k",
      "file": "crates/velesdb-core/src/index/trigram/mod.rs",
      "must_contain": "| 10K    | < 5ms",
      "validation_command": "none: no benchmark or test in the repository measures this target",
      "executable": false,
      "source": "A design target stated under '# Performance Targets' in the trigram module's rustdoc since commit cfaccf21 (2026-01-10, EPIC-CORE-003); no run records it.",
      "notes": "A target, not a measurement: the figure is the goal named in source.",
      "measured_on": "n/a",
      "measured_machine": "n/a",
      "measured_version": "n/a",
      "last_updated": "2026-09-13"
    },
    {
      "id": "rustdoc_trigram_target_100k",
      "file": "crates/velesdb-core/src/index/trigram/mod.rs",
      "must_contain": "| 100K   | < 20ms",
      "validation_command": "none: no benchmark or test in the repository measures this target",
      "executable": false,
      "source": "A design target stated under '# Performance Targets' in the trigram module's rustdoc since commit cfaccf21 (2026-01-10, EPIC-CORE-003); no run records it.",
      "notes": "A target, not a measurement: the figure is the goal named in source.",
      "measured_on": "n/a",
      "measured_machine": "n/a",
      "measured_version": "n/a",
      "last_updated": "2026-09-13"
    },
    {
      "id": "rustdoc_trigram_target_1m",
      "file": "crates/velesdb-core/src/index/trigram/mod.rs",
      "must_contain": "| 1M     | < 100ms",
      "validation_command": "none: no benchmark or test in the repository measures this target",
      "executable": false,
      "source": "A design target stated under '# Performance Targets' in the trigram module's rustdoc since commit cfaccf21 (2026-01-10, EPIC-CORE-003); no run records it.",
      "notes": "A target, not a measurement: the figure is the goal named in source.",
      "measured_on": "n/a",
      "measured_machine": "n/a",
      "measured_version": "n/a",
      "last_updated": "2026-09-13"
    },
    {
      "id": "benchmarks_recall_profile_fast",
      "file": "docs/BENCHMARKS.md",
      "must_contain": "| Fast | 96 | 97.4% |",
      "validation_command": "cargo bench -p velesdb-core --bench recall_benchmark",
      "executable": false,
      "source": "cargo bench -p velesdb-core --bench recall_benchmark: its recall report (10K random 128-D vectors, HnswParams::max_recall index, 100 queries, k=10) at the current presets, 2026-09-10 on 6.0.0 (Fast ef=96: 97.4%, Balanced ef=160: 99.8%, Accurate ef=512: 100.0%) - the run behind readme_recall_fast_mode, readme_recall_balanced_mode, readme_recall_accurate_mode and docs/BENCHMARKS.md section 5 'HNSW Recall Profiles (10K/128D)'.",
      "measured_on": "2026-09-10",
      "measured_machine": "Apple M5 Pro",
      "measured_version": "6.0.0",
      "last_updated": "2026-09-13"
    },
    {
      "id": "benchmarks_recall_profile_balanced",
      "file": "docs/BENCHMARKS.md",
      "must_contain": "| Balanced | 160 | 99.8% |",
      "validation_command": "cargo bench -p velesdb-core --bench recall_benchmark",
      "executable": false,
      "source": "cargo bench -p velesdb-core --bench recall_benchmark: its recall report (10K random 128-D vectors, HnswParams::max_recall index, 100 queries, k=10) at the current presets, 2026-09-10 on 6.0.0 (Fast ef=96: 97.4%, Balanced ef=160: 99.8%, Accurate ef=512: 100.0%) - the run behind readme_recall_fast_mode, readme_recall_balanced_mode, readme_recall_accurate_mode and docs/BENCHMARKS.md section 5 'HNSW Recall Profiles (10K/128D)'.",
      "measured_on": "2026-09-10",
      "measured_machine": "Apple M5 Pro",
      "measured_version": "6.0.0",
      "last_updated": "2026-09-13"
    },
    {
      "id": "benchmarks_recall_profile_accurate",
      "file": "docs/BENCHMARKS.md",
      "must_contain": "| Accurate | 512 | 100.0% |",
      "validation_command": "cargo bench -p velesdb-core --bench recall_benchmark",
      "executable": false,
      "source": "cargo bench -p velesdb-core --bench recall_benchmark: its recall report (10K random 128-D vectors, HnswParams::max_recall index, 100 queries, k=10) at the current presets, 2026-09-10 on 6.0.0 (Fast ef=96: 97.4%, Balanced ef=160: 99.8%, Accurate ef=512: 100.0%) - the run behind readme_recall_fast_mode, readme_recall_balanced_mode, readme_recall_accurate_mode and docs/BENCHMARKS.md section 5 'HNSW Recall Profiles (10K/128D)'.",
      "measured_on": "2026-09-10",
      "measured_machine": "Apple M5 Pro",
      "measured_version": "6.0.0",
      "last_updated": "2026-09-13"
    },
    {
      "id": "benchmarks_recall_profile_perfect",
      "file": "docs/BENCHMARKS.md",
      "must_contain": "| Perfect | exhaustive | 100.0% |",
      "validation_command": "cargo bench -p velesdb-core --bench recall_benchmark",
      "executable": false,
      "source": "cargo bench -p velesdb-core --bench recall_benchmark: its recall report (10K random 128-D vectors, HnswParams::max_recall index, 100 queries, k=10) at the current presets, 2026-09-10 on 6.0.0 (Fast ef=96: 97.4%, Balanced ef=160: 99.8%, Accurate ef=512: 100.0%) - the run behind readme_recall_fast_mode, readme_recall_balanced_mode, readme_recall_accurate_mode and docs/BENCHMARKS.md section 5 'HNSW Recall Profiles (10K/128D)'. The same report runs Perfect (the exhaustive scan: SearchQuality::Perfect in benches/recall_benchmark.rs), recorded at 100% in this row's table.",
      "measured_on": "2026-09-10",
      "measured_machine": "Apple M5 Pro",
      "measured_version": "6.0.0",
      "last_updated": "2026-09-13"
    },
    {
      "id": "search_modes_benchmarks_fast",
      "file": "docs/guides/SEARCH_MODES.md",
      "must_contain": "| Fast | 96 | 97.4% |",
      "validation_command": "cargo bench -p velesdb-core --bench recall_benchmark",
      "executable": false,
      "source": "cargo bench -p velesdb-core --bench recall_benchmark: its recall report (10K random 128-D vectors, HnswParams::max_recall index, 100 queries, k=10) at the current presets, 2026-09-10 on 6.0.0 (Fast ef=96: 97.4%, Balanced ef=160: 99.8%, Accurate ef=512: 100.0%) - the run behind readme_recall_fast_mode, readme_recall_balanced_mode, readme_recall_accurate_mode and docs/BENCHMARKS.md section 5 'HNSW Recall Profiles (10K/128D)'.",
      "measured_on": "2026-09-10",
      "measured_machine": "Apple M5 Pro",
      "measured_version": "6.0.0",
      "last_updated": "2026-09-13"
    },
    {
      "id": "search_modes_benchmarks_balanced",
      "file": "docs/guides/SEARCH_MODES.md",
      "must_contain": "| Balanced | 160 | 99.8% |",
      "validation_command": "cargo bench -p velesdb-core --bench recall_benchmark",
      "executable": false,
      "source": "cargo bench -p velesdb-core --bench recall_benchmark: its recall report (10K random 128-D vectors, HnswParams::max_recall index, 100 queries, k=10) at the current presets, 2026-09-10 on 6.0.0 (Fast ef=96: 97.4%, Balanced ef=160: 99.8%, Accurate ef=512: 100.0%) - the run behind readme_recall_fast_mode, readme_recall_balanced_mode, readme_recall_accurate_mode and docs/BENCHMARKS.md section 5 'HNSW Recall Profiles (10K/128D)'.",
      "measured_on": "2026-09-10",
      "measured_machine": "Apple M5 Pro",
      "measured_version": "6.0.0",
      "last_updated": "2026-09-13"
    },
    {
      "id": "search_modes_benchmarks_accurate",
      "file": "docs/guides/SEARCH_MODES.md",
      "must_contain": "| Accurate | 512 | 100.0% |",
      "validation_command": "cargo bench -p velesdb-core --bench recall_benchmark",
      "executable": false,
      "source": "cargo bench -p velesdb-core --bench recall_benchmark: its recall report (10K random 128-D vectors, HnswParams::max_recall index, 100 queries, k=10) at the current presets, 2026-09-10 on 6.0.0 (Fast ef=96: 97.4%, Balanced ef=160: 99.8%, Accurate ef=512: 100.0%) - the run behind readme_recall_fast_mode, readme_recall_balanced_mode, readme_recall_accurate_mode and docs/BENCHMARKS.md section 5 'HNSW Recall Profiles (10K/128D)'.",
      "measured_on": "2026-09-10",
      "measured_machine": "Apple M5 Pro",
      "measured_version": "6.0.0",
      "last_updated": "2026-09-13"
    },
    {
      "id": "search_modes_benchmarks_perfect",
      "file": "docs/guides/SEARCH_MODES.md",
      "must_contain": "| Perfect | exhaustive | 100.0% |",
      "validation_command": "cargo bench -p velesdb-core --bench recall_benchmark",
      "executable": false,
      "source": "cargo bench -p velesdb-core --bench recall_benchmark: its recall report (10K random 128-D vectors, HnswParams::max_recall index, 100 queries, k=10) at the current presets, 2026-09-10 on 6.0.0 (Fast ef=96: 97.4%, Balanced ef=160: 99.8%, Accurate ef=512: 100.0%) - the run behind readme_recall_fast_mode, readme_recall_balanced_mode, readme_recall_accurate_mode and docs/BENCHMARKS.md section 5 'HNSW Recall Profiles (10K/128D)'. The same report runs Perfect (the exhaustive scan: SearchQuality::Perfect in benches/recall_benchmark.rs), recorded at 100% in this row's table.",
      "measured_on": "2026-09-10",
      "measured_machine": "Apple M5 Pro",
      "measured_version": "6.0.0",
      "last_updated": "2026-09-13"
    },
    {
      "id": "core_perf_recall_balanced_row",
      "file": "docs/guides/CORE_PERFORMANCE.md",
      "must_contain": "| 10K/128D | Balanced | 160 | **99.8%** |",
      "validation_command": "cargo bench -p velesdb-core --bench recall_benchmark",
      "executable": false,
      "source": "cargo bench -p velesdb-core --bench recall_benchmark: its recall report (10K random 128-D vectors, HnswParams::max_recall index, 100 queries, k=10) at the current presets, 2026-09-10 on 6.0.0 (Fast ef=96: 97.4%, Balanced ef=160: 99.8%, Accurate ef=512: 100.0%) - the run behind readme_recall_fast_mode, readme_recall_balanced_mode, readme_recall_accurate_mode and docs/BENCHMARKS.md section 5 'HNSW Recall Profiles (10K/128D)'.",
      "measured_on": "2026-09-10",
      "measured_machine": "Apple M5 Pro",
      "measured_version": "6.0.0",
      "last_updated": "2026-09-13"
    },
    {
      "id": "core_perf_recall_accurate_row",
      "file": "docs/guides/CORE_PERFORMANCE.md",
      "must_contain": "| 10K/128D | Accurate | 512 | **100.0%** |",
      "validation_command": "cargo bench -p velesdb-core --bench recall_benchmark",
      "executable": false,
      "source": "cargo bench -p velesdb-core --bench recall_benchmark: its recall report (10K random 128-D vectors, HnswParams::max_recall index, 100 queries, k=10) at the current presets, 2026-09-10 on 6.0.0 (Fast ef=96: 97.4%, Balanced ef=160: 99.8%, Accurate ef=512: 100.0%) - the run behind readme_recall_fast_mode, readme_recall_balanced_mode, readme_recall_accurate_mode and docs/BENCHMARKS.md section 5 'HNSW Recall Profiles (10K/128D)'.",
      "measured_on": "2026-09-10",
      "measured_machine": "Apple M5 Pro",
      "measured_version": "6.0.0",
      "last_updated": "2026-09-13"
    },
    {
      "id": "core_perf_recall_perfect_row",
      "file": "docs/guides/CORE_PERFORMANCE.md",
      "must_contain": "| 10K/128D | Perfect | exhaustive | **100%** |",
      "validation_command": "cargo bench -p velesdb-core --bench recall_benchmark",
      "executable": false,
      "source": "cargo bench -p velesdb-core --bench recall_benchmark: its recall report (10K random 128-D vectors, HnswParams::max_recall index, 100 queries, k=10) at the current presets, 2026-09-10 on 6.0.0 (Fast ef=96: 97.4%, Balanced ef=160: 99.8%, Accurate ef=512: 100.0%) - the run behind readme_recall_fast_mode, readme_recall_balanced_mode, readme_recall_accurate_mode and docs/BENCHMARKS.md section 5 'HNSW Recall Profiles (10K/128D)'. The same report runs Perfect (the exhaustive scan: SearchQuality::Perfect in benches/recall_benchmark.rs), recorded at 100% in this row's table.",
      "measured_on": "2026-09-10",
      "measured_machine": "Apple M5 Pro",
      "measured_version": "6.0.0",
      "last_updated": "2026-09-13"
    },
    {
      "id": "tuning_recall_fast_row",
      "file": "docs/guides/TUNING_GUIDE.md",
      "must_contain": "| `Fast` | 96 | max(96, k*3) | 97.4%* |",
      "validation_command": "cargo bench -p velesdb-core --bench recall_benchmark",
      "executable": false,
      "source": "cargo bench -p velesdb-core --bench recall_benchmark: its recall report (10K random 128-D vectors, HnswParams::max_recall index, 100 queries, k=10) at the current presets, 2026-09-10 on 6.0.0 (Fast ef=96: 97.4%, Balanced ef=160: 99.8%, Accurate ef=512: 100.0%) - the run behind readme_recall_fast_mode, readme_recall_balanced_mode, readme_recall_accurate_mode and docs/BENCHMARKS.md section 5 'HNSW Recall Profiles (10K/128D)'.",
      "measured_on": "2026-09-10",
      "measured_machine": "Apple M5 Pro",
      "measured_version": "6.0.0",
      "last_updated": "2026-09-13"
    },
    {
      "id": "tuning_recall_balanced_row",
      "file": "docs/guides/TUNING_GUIDE.md",
      "must_contain": "| `Balanced` (default) | 160 | max(160, k*5) | 99.8%* |",
      "validation_command": "cargo bench -p velesdb-core --bench recall_benchmark",
      "executable": false,
      "source": "cargo bench -p velesdb-core --bench recall_benchmark: its recall report (10K random 128-D vectors, HnswParams::max_recall index, 100 queries, k=10) at the current presets, 2026-09-10 on 6.0.0 (Fast ef=96: 97.4%, Balanced ef=160: 99.8%, Accurate ef=512: 100.0%) - the run behind readme_recall_fast_mode, readme_recall_balanced_mode, readme_recall_accurate_mode and docs/BENCHMARKS.md section 5 'HNSW Recall Profiles (10K/128D)'.",
      "measured_on": "2026-09-10",
      "measured_machine": "Apple M5 Pro",
      "measured_version": "6.0.0",
      "last_updated": "2026-09-13"
    },
    {
      "id": "tuning_recall_accurate_row",
      "file": "docs/guides/TUNING_GUIDE.md",
      "must_contain": "| max(512, k*16) | 100%*;",
      "validation_command": "cargo bench -p velesdb-core --bench recall_benchmark",
      "executable": false,
      "source": "cargo bench -p velesdb-core --bench recall_benchmark: its recall report (10K random 128-D vectors, HnswParams::max_recall index, 100 queries, k=10) at the current presets, 2026-09-10 on 6.0.0 (Fast ef=96: 97.4%, Balanced ef=160: 99.8%, Accurate ef=512: 100.0%) - the run behind readme_recall_fast_mode, readme_recall_balanced_mode, readme_recall_accurate_mode and docs/BENCHMARKS.md section 5 'HNSW Recall Profiles (10K/128D)'.",
      "measured_on": "2026-09-10",
      "measured_machine": "Apple M5 Pro",
      "measured_version": "6.0.0",
      "last_updated": "2026-09-13"
    },
    {
      "id": "python_perf_recall_fast_row",
      "file": "docs/guides/PYTHON_PERFORMANCE.md",
      "must_contain": "| `\"fast\"` | 96 | 97.4%* |",
      "validation_command": "cargo bench -p velesdb-core --bench recall_benchmark",
      "executable": false,
      "source": "cargo bench -p velesdb-core --bench recall_benchmark: its recall report (10K random 128-D vectors, HnswParams::max_recall index, 100 queries, k=10) at the current presets, 2026-09-10 on 6.0.0 (Fast ef=96: 97.4%, Balanced ef=160: 99.8%, Accurate ef=512: 100.0%) - the run behind readme_recall_fast_mode, readme_recall_balanced_mode, readme_recall_accurate_mode and docs/BENCHMARKS.md section 5 'HNSW Recall Profiles (10K/128D)'.",
      "measured_on": "2026-09-10",
      "measured_machine": "Apple M5 Pro",
      "measured_version": "6.0.0",
      "last_updated": "2026-09-13"
    },
    {
      "id": "python_perf_recall_balanced_row",
      "file": "docs/guides/PYTHON_PERFORMANCE.md",
      "must_contain": "| `\"balanced\"` | 160 | 99.8%* |",
      "validation_command": "cargo bench -p velesdb-core --bench recall_benchmark",
      "executable": false,
      "source": "cargo bench -p velesdb-core --bench recall_benchmark: its recall report (10K random 128-D vectors, HnswParams::max_recall index, 100 queries, k=10) at the current presets, 2026-09-10 on 6.0.0 (Fast ef=96: 97.4%, Balanced ef=160: 99.8%, Accurate ef=512: 100.0%) - the run behind readme_recall_fast_mode, readme_recall_balanced_mode, readme_recall_accurate_mode and docs/BENCHMARKS.md section 5 'HNSW Recall Profiles (10K/128D)'.",
      "measured_on": "2026-09-10",
      "measured_machine": "Apple M5 Pro",
      "measured_version": "6.0.0",
      "last_updated": "2026-09-13"
    },
    {
      "id": "tuning_accurate_sift1m",
      "file": "docs/guides/TUNING_GUIDE.md",
      "must_contain": "0.98 on SIFT1M's 1M |",
      "validation_command": "cargo bench -p velesdb-core --bench sift1m_recall --features bench-sift1m",
      "executable": false,
      "source": "docs/BENCHMARKS.md section 11.3, production path: Accurate recall@10 0.9803 on SIFT1M (1M base x 10K queries, M=16, ef_construction=200, L2), 'First reproducible run, VelesDB v3.3.0' on Apple Silicon, otherwise idle. Recorded by #1225, squashed as 4b62338f (2026-06-24), whose message reads 'Adds report_recall_quality() printing RECALL_REPORT_QUALITY lines for Accurate (ef~1024 -> 0.9803)'; the PR commit that added that report, b33d7ebd, is dated 2026-06-24 too. Section 11.3 names the version and host but not the day, so measured_on is the day of the commit that records the run.",
      "measured_on": "2026-06-24 (date of commit 4b62338f, which records the run)",
      "measured_machine": "Apple Silicon (M-series, ARM64/NEON), otherwise idle - docs/BENCHMARKS.md section 11.3",
      "measured_version": "3.3.0",
      "stale_accepted": {
        "reason": "Published with its version (3.3.0) and host in docs/BENCHMARKS.md section 11.3. No full 1M SIFT1M re-run has been recorded since: the nightly SIFT1M job gates recall@10 >= 0.90 at ef=128 but commits no figures, so the figure stays pinned to the run that produced it.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "rustdoc_config_accurate_sift1m",
      "file": "crates/velesdb-core/src/config.rs",
      "must_contain": "0.98 on SIFT1M's 1M (`docs/BENCHMARKS.md`)",
      "validation_command": "cargo bench -p velesdb-core --bench sift1m_recall --features bench-sift1m",
      "executable": false,
      "source": "docs/BENCHMARKS.md section 11.3, production path: Accurate recall@10 0.9803 on SIFT1M (1M base x 10K queries, M=16, ef_construction=200, L2), 'First reproducible run, VelesDB v3.3.0' on Apple Silicon, otherwise idle. Recorded by #1225, squashed as 4b62338f (2026-06-24), whose message reads 'Adds report_recall_quality() printing RECALL_REPORT_QUALITY lines for Accurate (ef~1024 -> 0.9803)'; the PR commit that added that report, b33d7ebd, is dated 2026-06-24 too. Section 11.3 names the version and host but not the day, so measured_on is the day of the commit that records the run.",
      "measured_on": "2026-06-24 (date of commit 4b62338f, which records the run)",
      "measured_machine": "Apple Silicon (M-series, ARM64/NEON), otherwise idle - docs/BENCHMARKS.md section 11.3",
      "measured_version": "3.3.0",
      "stale_accepted": {
        "reason": "Published with its version (3.3.0) and host in docs/BENCHMARKS.md section 11.3. No full 1M SIFT1M re-run has been recorded since: the nightly SIFT1M job gates recall@10 >= 0.90 at ef=128 but commits no figures, so the figure stays pinned to the run that produced it.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "rustdoc_params_accurate_sift1m",
      "file": "crates/velesdb-core/src/index/hnsw/params.rs",
      "must_contain": "0.98 on SIFT1M's 1M (`docs/BENCHMARKS.md`)",
      "validation_command": "cargo bench -p velesdb-core --bench sift1m_recall --features bench-sift1m",
      "executable": false,
      "source": "docs/BENCHMARKS.md section 11.3, production path: Accurate recall@10 0.9803 on SIFT1M (1M base x 10K queries, M=16, ef_construction=200, L2), 'First reproducible run, VelesDB v3.3.0' on Apple Silicon, otherwise idle. Recorded by #1225, squashed as 4b62338f (2026-06-24), whose message reads 'Adds report_recall_quality() printing RECALL_REPORT_QUALITY lines for Accurate (ef~1024 -> 0.9803)'; the PR commit that added that report, b33d7ebd, is dated 2026-06-24 too. Section 11.3 names the version and host but not the day, so measured_on is the day of the commit that records the run.",
      "measured_on": "2026-06-24 (date of commit 4b62338f, which records the run)",
      "measured_machine": "Apple Silicon (M-series, ARM64/NEON), otherwise idle - docs/BENCHMARKS.md section 11.3",
      "measured_version": "3.3.0",
      "stale_accepted": {
        "reason": "Published with its version (3.3.0) and host in docs/BENCHMARKS.md section 11.3. No full 1M SIFT1M re-run has been recorded since: the nightly SIFT1M job gates recall@10 >= 0.90 at ef=128 but commits no figures, so the figure stays pinned to the run that produced it.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "rustdoc_mobile_accurate_sift1m",
      "file": "crates/velesdb-mobile/src/types.rs",
      "must_contain": "`recall_benchmark`, 0.98 on SIFT1M's 1M.",
      "validation_command": "cargo bench -p velesdb-core --bench sift1m_recall --features bench-sift1m",
      "executable": false,
      "source": "docs/BENCHMARKS.md section 11.3, production path: Accurate recall@10 0.9803 on SIFT1M (1M base x 10K queries, M=16, ef_construction=200, L2), 'First reproducible run, VelesDB v3.3.0' on Apple Silicon, otherwise idle. Recorded by #1225, squashed as 4b62338f (2026-06-24), whose message reads 'Adds report_recall_quality() printing RECALL_REPORT_QUALITY lines for Accurate (ef~1024 -> 0.9803)'; the PR commit that added that report, b33d7ebd, is dated 2026-06-24 too. Section 11.3 names the version and host but not the day, so measured_on is the day of the commit that records the run.",
      "measured_on": "2026-06-24 (date of commit 4b62338f, which records the run)",
      "measured_machine": "Apple Silicon (M-series, ARM64/NEON), otherwise idle - docs/BENCHMARKS.md section 11.3",
      "measured_version": "3.3.0",
      "stale_accepted": {
        "reason": "Published with its version (3.3.0) and host in docs/BENCHMARKS.md section 11.3. No full 1M SIFT1M re-run has been recorded since: the nightly SIFT1M job gates recall@10 >= 0.90 at ef=128 but commits no figures, so the figure stays pinned to the run that produced it.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "benchmarks_pq_full_precision_row",
      "file": "docs/BENCHMARKS.md",
      "must_contain": "| **Full Precision** | 98.4% |",
      "validation_command": "cargo bench -p velesdb-core --bench pq_recall_benchmark -- full_precision_baseline",
      "executable": false,
      "source": "commit e35d9612 (2026-03-08, 'restore 0.92 recall thresholds with uniform random dataset'): moved pq_recall_benchmark from clustered to uniform-random data, restored the full-precision assert to recall >= 0.95, and recorded 'All 6 benchmark variants pass: PQ=0.994, Full=0.984, ...'. The 87.6% docs/BENCHMARKS.md showed before came from the clustered data of 7f05d0f0 (2026-03-07). measured_version is the workspace version of commit e35d9612.",
      "measured_on": "2026-03-08 (date of commit e35d9612, which records the run)",
      "measured_machine": "unknown (commit e35d9612 does not name it)",
      "measured_version": "1.5.0",
      "stale_accepted": {
        "reason": "A recall figure published with the commit and date of its run (e35d9612, 2026-03-08); re-running pq_recall_benchmark takes a cargo bench run on a quiet machine, not made in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "search_modes_recall_scale_exact",
      "file": "docs/guides/SEARCH_MODES.md",
      "must_contain": "| **100%** | All true neighbors found (exact search) |",
      "validation_command": "none: a scale, not a measurement",
      "executable": false,
      "source": "Written by commit 5ca54a75 (2026-06-12): a scale for reading recall values under '### What is Recall?', not a figure VelesDB measured.",
      "notes": "Not a measurement: nothing here records a run.",
      "measured_on": "n/a",
      "measured_machine": "n/a",
      "measured_version": "n/a",
      "last_updated": "2026-09-13"
    },
    {
      "id": "search_modes_recall_scale_95_99",
      "file": "docs/guides/SEARCH_MODES.md",
      "must_contain": "| **95-99%** | Excellent",
      "validation_command": "none: a scale, not a measurement",
      "executable": false,
      "source": "Written by commit 32bfd2df (2026-03-07): a scale for reading recall values under '### What is Recall?', not a figure VelesDB measured.",
      "notes": "Not a measurement: nothing here records a run.",
      "measured_on": "n/a",
      "measured_machine": "n/a",
      "measured_version": "n/a",
      "last_updated": "2026-09-13"
    },
    {
      "id": "search_modes_recall_scale_90_95",
      "file": "docs/guides/SEARCH_MODES.md",
      "must_contain": "| **90-95%** | Acceptable for exploration/prototyping |",
      "validation_command": "none: a scale, not a measurement",
      "executable": false,
      "source": "Written by commit 5ca54a75 (2026-06-12): a scale for reading recall values under '### What is Recall?', not a figure VelesDB measured.",
      "notes": "Not a measurement: nothing here records a run.",
      "measured_on": "n/a",
      "measured_machine": "n/a",
      "measured_version": "n/a",
      "last_updated": "2026-09-13"
    },
    {
      "id": "search_modes_recall_scale_below_90",
      "file": "docs/guides/SEARCH_MODES.md",
      "must_contain": "| **< 90%** | Risk of missing important results |",
      "validation_command": "none: a scale, not a measurement",
      "executable": false,
      "source": "Written by commit 5ca54a75 (2026-06-12): a scale for reading recall values under '### What is Recall?', not a figure VelesDB measured.",
      "notes": "Not a measurement: nothing here records a run.",
      "measured_on": "n/a",
      "measured_machine": "n/a",
      "measured_version": "n/a",
      "last_updated": "2026-09-13"
    },
    {
      "id": "rustdoc_params_scale_target_10k",
      "file": "crates/velesdb-core/src/index/hnsw/params.rs",
      "must_contain": "| ≤10K | 32 | 200 | ≥98% |",
      "validation_command": "none: no benchmark or test in the repository measures these targets",
      "executable": false,
      "source": "Written by commit f7c9dcdc (2026-01-08): the 'Target Recall' column of HnswParams's '# Parameters by Scale' table: a target per dataset size, which no run in the tree records.",
      "notes": "Not a measurement: nothing here records a run.",
      "measured_on": "n/a",
      "measured_machine": "n/a",
      "measured_version": "n/a",
      "last_updated": "2026-09-13"
    },
    {
      "id": "rustdoc_params_scale_target_100k",
      "file": "crates/velesdb-core/src/index/hnsw/params.rs",
      "must_contain": "| ≤100K | 64 | 800 | ≥95% |",
      "validation_command": "none: no benchmark or test in the repository measures these targets",
      "executable": false,
      "source": "Written by commit f7c9dcdc (2026-01-08): the 'Target Recall' column of HnswParams's '# Parameters by Scale' table: a target per dataset size, which no run in the tree records.",
      "notes": "Not a measurement: nothing here records a run.",
      "measured_on": "n/a",
      "measured_machine": "n/a",
      "measured_version": "n/a",
      "last_updated": "2026-09-13"
    },
    {
      "id": "rustdoc_params_scale_target_500k",
      "file": "crates/velesdb-core/src/index/hnsw/params.rs",
      "must_contain": "| ≤500K | 96 | 1200 | ≥95% |",
      "validation_command": "none: no benchmark or test in the repository measures these targets",
      "executable": false,
      "source": "Written by commit f7c9dcdc (2026-01-08): the 'Target Recall' column of HnswParams's '# Parameters by Scale' table: a target per dataset size, which no run in the tree records.",
      "notes": "Not a measurement: nothing here records a run.",
      "measured_on": "n/a",
      "measured_machine": "n/a",
      "measured_version": "n/a",
      "last_updated": "2026-09-13"
    },
    {
      "id": "rustdoc_params_scale_target_1m",
      "file": "crates/velesdb-core/src/index/hnsw/params.rs",
      "must_contain": "| ≤1M | 128 | 1600 | ≥95% |",
      "validation_command": "none: no benchmark or test in the repository measures these targets",
      "executable": false,
      "source": "Written by commit f7c9dcdc (2026-01-08): the 'Target Recall' column of HnswParams's '# Parameters by Scale' table: a target per dataset size, which no run in the tree records.",
      "notes": "Not a measurement: nothing here records a run.",
      "measured_on": "n/a",
      "measured_machine": "n/a",
      "measured_version": "n/a",
      "last_updated": "2026-09-13"
    },
    {
      "id": "architecture_diagrams_recall_gate",
      "file": "docs/reference/ARCHITECTURE_DIAGRAMS.md",
      "must_contain": "| Recall ≥ 0.95 |",
      "validation_command": "cargo test -p velesdb-core --lib precision_contract_tests",
      "executable": false,
      "source": "Written by commit 1564f57a (2026-04-04): a quality-gate row: the floor the precision contract tests assert, not a measured value.",
      "notes": "Not a measurement: nothing here records a run.",
      "measured_on": "n/a",
      "measured_machine": "n/a",
      "measured_version": "n/a",
      "last_updated": "2026-09-13"
    },
    {
      "id": "rustdoc_sq8_recall_contract_floor",
      "file": "crates/velesdb-core/src/index/hnsw/native/sq8_precision.rs",
      "must_contain": "0.95 contract is pinned on the default configuration",
      "validation_command": "cargo test -p velesdb-core --lib precision_contract_tests",
      "executable": false,
      "source": "Written by commit d5ce370e (2026-08-22): the recall@10 floor the precision contract tests assert on the default configuration, not a measured value.",
      "notes": "Not a measurement: nothing here records a run.",
      "measured_on": "n/a",
      "measured_machine": "n/a",
      "measured_version": "n/a",
      "last_updated": "2026-09-13"
    },
    {
      "id": "rustdoc_precision_contract_recall_floor_line148",
      "file": "crates/velesdb-core/src/index/hnsw/native/precision_contract_tests.rs",
      "must_contain": "recall@k vs brute-force is >= 0.95.",
      "validation_command": "cargo test -p velesdb-core --lib precision_contract_tests",
      "executable": false,
      "source": "Written by commit c1d12b25 (2026-08-22): the floor this test module asserts, not a measured value.",
      "notes": "Not a measurement: nothing here records a run.",
      "measured_on": "n/a",
      "measured_machine": "n/a",
      "measured_version": "n/a",
      "last_updated": "2026-09-13"
    },
    {
      "id": "rustdoc_rabitq_precision_recall_floor",
      "file": "crates/velesdb-core/src/index/hnsw/native/rabitq_precision_tests.rs",
      "must_contain": "recall@k >= 0.95.",
      "validation_command": "cargo test -p velesdb-core --lib rabitq_precision_tests",
      "executable": false,
      "source": "Written by commit 1ede859e (2026-06-10): the floor this test module asserts, not a measured value.",
      "notes": "Not a measurement: nothing here records a run.",
      "measured_on": "n/a",
      "measured_machine": "n/a",
      "measured_version": "n/a",
      "last_updated": "2026-09-13"
    },
    {
      "id": "rustdoc_graduated_ef_recall_floor",
      "file": "crates/velesdb-core/src/index/hnsw/native/backend_adapter_tests.rs",
      "must_contain": "maintains recall >= 0.90",
      "validation_command": "cargo test -p velesdb-core --lib test_graduated_ef_construction_recall",
      "executable": false,
      "source": "Written by commit d6c6a575 (2026-04-02): the floor test_graduated_ef_construction_recall asserts, not a measured value.",
      "notes": "Not a measurement: nothing here records a run.",
      "measured_on": "n/a",
      "measured_machine": "n/a",
      "measured_version": "n/a",
      "last_updated": "2026-09-13"
    },
    {
      "id": "rustdoc_gpu_search_auto_recall_floor",
      "file": "crates/velesdb-core/src/index/hnsw/gpu_search_auto_tests.rs",
      "must_contain": "recall against brute force is above 0.80",
      "validation_command": "cargo test -p velesdb-core --lib search_auto_wiring_preserves_recall_small_dataset",
      "executable": false,
      "source": "Written by commit dec8264f (2026-04-23): the floor search_auto_wiring_preserves_recall_small_dataset asserts (recall >= 0.80), not a measured value.",
      "notes": "Not a measurement: nothing here records a run.",
      "measured_on": "n/a",
      "measured_machine": "n/a",
      "measured_version": "n/a",
      "last_updated": "2026-09-13"
    },
    {
      "id": "rustdoc_neighbors_recall_floors",
      "file": "crates/velesdb-core/src/index/hnsw/native/graph/neighbors.rs",
      "must_contain": "Recall quality is enforced by tests: >= 0.80 in unit tests",
      "validation_command": "none: the comment names no test",
      "executable": false,
      "source": "Written by commit d123c589 (2026-04-02): floors the HNSW tests assert (>= 0.80 at 1K vectors, >= 0.90 at 100K), not measured values; the comment names no test.",
      "notes": "Not a measurement: nothing here records a run.",
      "measured_on": "n/a",
      "measured_machine": "n/a",
      "measured_version": "n/a",
      "last_updated": "2026-09-13"
    },
    {
      "id": "tuning_new_recall_target",
      "file": "docs/guides/TUNING_GUIDE.md",
      "must_contain": "| ≥95% target¹ |",
      "validation_command": "python benchmarks/velesdb_benchmark.py --recall",
      "executable": false,
      "source": "Written by commit d498c4c1 (2026-09-13): the design target of the auto() default. Footnote ¹ gives what a run measured on that path: 97.6% recall@10 (benchmarks/report_1.7.2_2026-03-27.json).",
      "notes": "Not a measurement: nothing here records a run.",
      "measured_on": "n/a",
      "measured_machine": "n/a",
      "measured_version": "n/a",
      "last_updated": "2026-09-13"
    },
    {
      "id": "native_hnsw_new_recall_target",
      "file": "docs/reference/NATIVE_HNSW.md",
      "must_contain": "| ≥95% target¹ |",
      "validation_command": "python benchmarks/velesdb_benchmark.py --recall",
      "executable": false,
      "source": "Written by commit d498c4c1 (2026-09-13): the design target of the auto() default. Footnote ¹ gives what a run measured on that path: 97.6% recall@10 (benchmarks/report_1.7.2_2026-03-27.json).",
      "notes": "Not a measurement: nothing here records a run.",
      "measured_on": "n/a",
      "measured_machine": "n/a",
      "measured_version": "n/a",
      "last_updated": "2026-09-13"
    },
    {
      "id": "benchmarks_pq07_floor_cell",
      "file": "docs/BENCHMARKS.md",
      "must_contain": "PQ-07 contract asserts recall>=0.92)",
      "validation_command": "cargo bench -p velesdb-core --bench pq_recall_benchmark",
      "executable": false,
      "source": "Written by commit c4d8acbe (2026-06-21): the floor pq_recall_benchmark asserts for PQ m=8 k=256 with rescore ('PQ-07 contract: recall@10 >= 0.92'); the cell says no run re-measures the recall.",
      "notes": "Not a measurement: nothing here records a run.",
      "measured_on": "n/a",
      "measured_machine": "n/a",
      "measured_version": "n/a",
      "last_updated": "2026-09-13"
    },
    {
      "id": "benchmarks_pq07_floor_note",
      "file": "docs/BENCHMARKS.md",
      "must_contain": "the PQ-07 contract asserts recall>=0.92, so the prior 30.6% figure",
      "validation_command": "cargo bench -p velesdb-core --bench pq_recall_benchmark",
      "executable": false,
      "source": "Written by commit c4d8acbe (2026-06-21): the floor pq_recall_benchmark asserts (recall@10 >= 0.92), and 30.6%, the figure this note retires as an obsolete artifact rather than a result.",
      "notes": "Not a measurement: nothing here records a run.",
      "measured_on": "n/a",
      "measured_machine": "n/a",
      "measured_version": "n/a",
      "last_updated": "2026-09-13"
    },
    {
      "id": "server_examples_quickstart_goal",
      "file": "crates/velesdb-server/examples/README.md",
      "must_contain": "The README \"first success in 60 seconds\"",
      "validation_command": "none: a goal, not a measurement",
      "executable": false,
      "source": "Written by commit 5afeee5f (2026-07-25): the name of the README's quickstart: a goal no run times.",
      "notes": "Not a measurement: nothing here records a run.",
      "measured_on": "n/a",
      "measured_machine": "n/a",
      "measured_version": "n/a",
      "last_updated": "2026-09-13"
    },
    {
      "id": "wasm_examples_quickstart_goal",
      "file": "crates/velesdb-wasm/examples/README.md",
      "must_contain": "The README's \"first success in 60 seconds\"",
      "validation_command": "none: a goal, not a measurement",
      "executable": false,
      "source": "Written by commit 5afeee5f (2026-07-25): the name of the README's quickstart: a goal no run times.",
      "notes": "Not a measurement: nothing here records a run.",
      "measured_on": "n/a",
      "measured_machine": "n/a",
      "measured_version": "n/a",
      "last_updated": "2026-09-13"
    },
    {
      "id": "memory_readme_multihop_hash",
      "file": "crates/velesdb-memory/README.md",
      "must_contain": "| `hash` (deterministic) | 100% | 0% | **100%** |",
      "validation_command": "cargo run --release -p velesdb-memory --example bench_multihop",
      "executable": false,
      "source": "commit 422848f4 (2026-06-26) wrote this table from the bench_multihop example: 'Graph contribution to multi-hop answer recall: +100pp (deterministic) and +67pp (vector 33%→graph 100%, real all-minilm), direct recall 100%.' measured_version is the workspace version of that commit.",
      "measured_on": "2026-06-26 (date of commit 422848f4, which records the run)",
      "measured_machine": "platform-independent (the deterministic hash embedder)",
      "measured_version": "3.3.0",
      "stale_accepted": {
        "reason": "Recorded with the change that introduced the multi-hop benchmark (422848f4, 3.3.0); re-running bench_multihop is not done in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "memory_readme_multihop_minilm",
      "file": "crates/velesdb-memory/README.md",
      "must_contain": "| real model (Ollama `all-minilm`) | 100% | 33% | **100%** |",
      "validation_command": "VELESDB_MEMORY_EMBEDDER=ollama cargo run --release -p velesdb-memory --features embedder-http --example bench_multihop",
      "executable": false,
      "source": "commit 422848f4 (2026-06-26) wrote this row from the bench_multihop example run on a local Ollama all-minilm: '+67pp (vector 33%→graph 100%, real all-minilm), direct recall 100%.' measured_version is the workspace version of that commit.",
      "measured_on": "2026-06-26 (date of commit 422848f4, which records the run)",
      "measured_machine": "unknown (commit 422848f4 names none)",
      "measured_version": "3.3.0",
      "stale_accepted": {
        "reason": "Recorded with the change that introduced the multi-hop benchmark (422848f4, 3.3.0); re-running it needs Ollama serving all-minilm, not done in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "migrate_readme_expected_output",
      "file": "crates/velesdb-migrate/README.md",
      "must_contain": "0 failed in 0.12s (24 pts/sec)",
      "validation_command": "velesdb-migrate run --config migration.yaml",
      "executable": false,
      "source": "commit 871d21af (2026-07-25) pasted this log as the README's expected output (timestamps 2026-07-25T16:30:04); its message is silent on it, and the README says the timings vary with the machine. measured_version is the workspace version of that commit.",
      "measured_on": "2026-07-25 (date of commit 871d21af, which records the run)",
      "measured_machine": "unknown (neither commit 871d21af nor the README names it)",
      "measured_version": "4.0.0",
      "stale_accepted": {
        "reason": "An example run's log, recorded with the README rewrite (871d21af, 4.0.0); by the README's own words its timings vary with the machine.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "rustdoc_hnsw_concurrent_latency_66pct",
      "file": "crates/velesdb-core/src/index/hnsw/native/graph/mod.rs",
      "must_contain": "with per-batch latency 66% higher than single-threaded.",
      "validation_command": "cargo bench -p velesdb-core --bench scalability_benchmark -- concurrent_queries",
      "executable": false,
      "source": "commit 185660d3 (2026-06-01) wrote this line with the ~5× beside it, from scalability_benchmark::concurrent_queries on an M5 Pro; its message states the ~5× ('~5× throughput at 8 threads instead of the ideal 8×') and not the 66%. measured_version is the workspace version of that commit.",
      "measured_on": "2026-06-01 (date of commit 185660d3, which records the run; the rustdoc dates the closed investigation 2026-06-03)",
      "measured_machine": "Apple M5 Pro (as commit 185660d3 and the rustdoc name it)",
      "measured_version": "1.16.0",
      "stale_accepted": {
        "reason": "Recorded by commit 185660d3 (1.16.0) as the closing measurement of #967; re-running it needs the same M5 Pro at rest, not done in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "rustdoc_memory_write_attribution_row219",
      "file": "crates/velesdb-memory/src/migration/tests/performance.rs",
      "must_contain": "| full (content+project)      |  3457.8 |",
      "validation_command": "cargo test -p velesdb-memory --release -- --ignored the_per_point_write_cost_is_attributed_to_payload_or_vector",
      "executable": false,
      "source": "commit 83784a92 (2026-08-03) recorded this table in the doc comment of the_per_point_write_cost_is_attributed_to_payload_or_vector ('Measured 2026-08-03 (macOS, --release, n = 2 000, batch = 1 024)'); 6a64778e moved it here unchanged. measured_version is the workspace version of that commit.",
      "measured_on": "2026-08-03 (date of commit 83784a92, which records the run)",
      "measured_machine": "macOS, --release (the doc comment names no model)",
      "measured_version": "4.2.0",
      "stale_accepted": {
        "reason": "The test's own recorded measurement (83784a92, 4.2.0); re-running it writes 2 000 facts three times on a machine at rest, not done in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "rustdoc_memory_write_attribution_row220",
      "file": "crates/velesdb-memory/src/migration/tests/performance.rs",
      "must_contain": "| no content key (still text) |  3418.7 |",
      "validation_command": "cargo test -p velesdb-memory --release -- --ignored the_per_point_write_cost_is_attributed_to_payload_or_vector",
      "executable": false,
      "source": "commit 83784a92 (2026-08-03) recorded this table in the doc comment of the_per_point_write_cost_is_attributed_to_payload_or_vector ('Measured 2026-08-03 (macOS, --release, n = 2 000, batch = 1 024)'); 6a64778e moved it here unchanged. measured_version is the workspace version of that commit.",
      "measured_on": "2026-08-03 (date of commit 83784a92, which records the run)",
      "measured_machine": "macOS, --release (the doc comment names no model)",
      "measured_version": "4.2.0",
      "stale_accepted": {
        "reason": "The test's own recorded measurement (83784a92, 4.2.0); re-running it writes 2 000 facts three times on a machine at rest, not done in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "rustdoc_memory_write_attribution_row221",
      "file": "crates/velesdb-memory/src/migration/tests/performance.rs",
      "must_contain": "| numeric only (no text)      |    15.1 |",
      "validation_command": "cargo test -p velesdb-memory --release -- --ignored the_per_point_write_cost_is_attributed_to_payload_or_vector",
      "executable": false,
      "source": "commit 83784a92 (2026-08-03) recorded this table in the doc comment of the_per_point_write_cost_is_attributed_to_payload_or_vector ('Measured 2026-08-03 (macOS, --release, n = 2 000, batch = 1 024)'); 6a64778e moved it here unchanged. measured_version is the workspace version of that commit.",
      "measured_on": "2026-08-03 (date of commit 83784a92, which records the run)",
      "measured_machine": "macOS, --release (the doc comment names no model)",
      "measured_version": "4.2.0",
      "stale_accepted": {
        "reason": "The test's own recorded measurement (83784a92, 4.2.0); re-running it writes 2 000 facts three times on a machine at rest, not done in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "rustdoc_memory_write_attribution_row222",
      "file": "crates/velesdb-memory/src/migration/tests/performance.rs",
      "must_contain": "| no payload at all           |    10.2 |",
      "validation_command": "cargo test -p velesdb-memory --release -- --ignored the_per_point_write_cost_is_attributed_to_payload_or_vector",
      "executable": false,
      "source": "commit 83784a92 (2026-08-03) recorded this table in the doc comment of the_per_point_write_cost_is_attributed_to_payload_or_vector ('Measured 2026-08-03 (macOS, --release, n = 2 000, batch = 1 024)'); 6a64778e moved it here unchanged. measured_version is the workspace version of that commit.",
      "measured_on": "2026-08-03 (date of commit 83784a92, which records the run)",
      "measured_machine": "macOS, --release (the doc comment names no model)",
      "measured_version": "4.2.0",
      "stale_accepted": {
        "reason": "The test's own recorded measurement (83784a92, 4.2.0); re-running it writes 2 000 facts three times on a machine at rest, not done in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "rustdoc_memory_batched_write_row296",
      "file": "crates/velesdb-memory/src/migration/tests/performance.rs",
      "must_contain": "| 2 000  | 3 408.8      | 20.5        | 48 794        |",
      "validation_command": "cargo test -p velesdb-memory --release -- --ignored batched_write_cost_stays_flat_as_the_volume_doubles",
      "executable": false,
      "source": "commit a4950038 (2026-08-03) recorded this table in the doc comment of batched_write_cost_stays_flat_as_the_volume_doubles, 're-measured with the same code on the same machine, --release, batch 1 024'; 6a64778e moved it here unchanged. measured_version is the workspace version of that commit.",
      "measured_on": "2026-08-03 (date of commit a4950038, which records the run)",
      "measured_machine": "unknown (the doc comment says 'an idle machine'; commit a4950038 names none)",
      "measured_version": "4.2.0",
      "stale_accepted": {
        "reason": "The test's own recorded measurement (a4950038, 4.2.0); re-running it writes 31 000 facts on a machine at rest, not done in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "rustdoc_memory_batched_write_row297",
      "file": "crates/velesdb-memory/src/migration/tests/performance.rs",
      "must_contain": "| 4 000  | 3 341.6      | 17.7        | 56 518        |",
      "validation_command": "cargo test -p velesdb-memory --release -- --ignored batched_write_cost_stays_flat_as_the_volume_doubles",
      "executable": false,
      "source": "commit a4950038 (2026-08-03) recorded this table in the doc comment of batched_write_cost_stays_flat_as_the_volume_doubles, 're-measured with the same code on the same machine, --release, batch 1 024'; 6a64778e moved it here unchanged. measured_version is the workspace version of that commit.",
      "measured_on": "2026-08-03 (date of commit a4950038, which records the run)",
      "measured_machine": "unknown (the doc comment says 'an idle machine'; commit a4950038 names none)",
      "measured_version": "4.2.0",
      "stale_accepted": {
        "reason": "The test's own recorded measurement (a4950038, 4.2.0); re-running it writes 31 000 facts on a machine at rest, not done in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "rustdoc_memory_batched_write_row298",
      "file": "crates/velesdb-memory/src/migration/tests/performance.rs",
      "must_contain": "| 8 000  | 3 374.0      | 16.7        | 59 720        |",
      "validation_command": "cargo test -p velesdb-memory --release -- --ignored batched_write_cost_stays_flat_as_the_volume_doubles",
      "executable": false,
      "source": "commit a4950038 (2026-08-03) recorded this table in the doc comment of batched_write_cost_stays_flat_as_the_volume_doubles, 're-measured with the same code on the same machine, --release, batch 1 024'; 6a64778e moved it here unchanged. measured_version is the workspace version of that commit.",
      "measured_on": "2026-08-03 (date of commit a4950038, which records the run)",
      "measured_machine": "unknown (the doc comment says 'an idle machine'; commit a4950038 names none)",
      "measured_version": "4.2.0",
      "stale_accepted": {
        "reason": "The test's own recorded measurement (a4950038, 4.2.0); re-running it writes 31 000 facts on a machine at rest, not done in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "rustdoc_memory_batched_write_row299",
      "file": "crates/velesdb-memory/src/migration/tests/performance.rs",
      "must_contain": "| 16 000 | 3 345.8      | **16.3**    | **61 524**    |",
      "validation_command": "cargo test -p velesdb-memory --release -- --ignored batched_write_cost_stays_flat_as_the_volume_doubles",
      "executable": false,
      "source": "commit a4950038 (2026-08-03) recorded this table in the doc comment of batched_write_cost_stays_flat_as_the_volume_doubles, 're-measured with the same code on the same machine, --release, batch 1 024'; 6a64778e moved it here unchanged. measured_version is the workspace version of that commit.",
      "measured_on": "2026-08-03 (date of commit a4950038, which records the run)",
      "measured_machine": "unknown (the doc comment says 'an idle machine'; commit a4950038 names none)",
      "measured_version": "4.2.0",
      "stale_accepted": {
        "reason": "The test's own recorded measurement (a4950038, 4.2.0); re-running it writes 31 000 facts on a machine at rest, not done in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "rustdoc_memory_paging_cost_row412",
      "file": "crates/velesdb-memory/src/migration/tests/performance.rs",
      "must_contain": "| 250   | 10.3 ms  | 41.1    | —                 |",
      "validation_command": "cargo test -p velesdb-memory -- --ignored paging_cost_grows_faster_than_the_store",
      "executable": false,
      "source": "commit 83784a92 (2026-08-03) recorded this table in the doc comment of paging_cost_grows_faster_than_the_store ('Measured 2026-08-03 (macOS, page size 100)'); 6a64778e moved it here unchanged. measured_version is the workspace version of that commit.",
      "measured_on": "2026-08-03 (date of commit 83784a92, which records the run)",
      "measured_machine": "macOS (the doc comment names no model)",
      "measured_version": "4.2.0",
      "stale_accepted": {
        "reason": "The test's own recorded measurement (83784a92, 4.2.0); re-running it writes 100 001 facts and, per its #[ignore], wants a machine at rest - not done in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "rustdoc_memory_paging_cost_row413",
      "file": "crates/velesdb-memory/src/migration/tests/performance.rs",
      "must_contain": "| 500   | 24.4 ms  | 48.8    | ×2.37             |",
      "validation_command": "cargo test -p velesdb-memory -- --ignored paging_cost_grows_faster_than_the_store",
      "executable": false,
      "source": "commit 83784a92 (2026-08-03) recorded this table in the doc comment of paging_cost_grows_faster_than_the_store ('Measured 2026-08-03 (macOS, page size 100)'); 6a64778e moved it here unchanged. measured_version is the workspace version of that commit.",
      "measured_on": "2026-08-03 (date of commit 83784a92, which records the run)",
      "measured_machine": "macOS (the doc comment names no model)",
      "measured_version": "4.2.0",
      "stale_accepted": {
        "reason": "The test's own recorded measurement (83784a92, 4.2.0); re-running it writes 100 001 facts and, per its #[ignore], wants a machine at rest - not done in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "rustdoc_memory_paging_cost_row414",
      "file": "crates/velesdb-memory/src/migration/tests/performance.rs",
      "must_contain": "| 1 000 | 69.7 ms  | 69.7    | ×2.86             |",
      "validation_command": "cargo test -p velesdb-memory -- --ignored paging_cost_grows_faster_than_the_store",
      "executable": false,
      "source": "commit 83784a92 (2026-08-03) recorded this table in the doc comment of paging_cost_grows_faster_than_the_store ('Measured 2026-08-03 (macOS, page size 100)'); 6a64778e moved it here unchanged. measured_version is the workspace version of that commit.",
      "measured_on": "2026-08-03 (date of commit 83784a92, which records the run)",
      "measured_machine": "macOS (the doc comment names no model)",
      "measured_version": "4.2.0",
      "stale_accepted": {
        "reason": "The test's own recorded measurement (83784a92, 4.2.0); re-running it writes 100 001 facts and, per its #[ignore], wants a machine at rest - not done in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "rustdoc_memory_paging_cost_row415",
      "file": "crates/velesdb-memory/src/migration/tests/performance.rs",
      "must_contain": "| 2 000 | 252.9 ms | 126.4   | ×3.63             |",
      "validation_command": "cargo test -p velesdb-memory -- --ignored paging_cost_grows_faster_than_the_store",
      "executable": false,
      "source": "commit 83784a92 (2026-08-03) recorded this table in the doc comment of paging_cost_grows_faster_than_the_store ('Measured 2026-08-03 (macOS, page size 100)'); 6a64778e moved it here unchanged. measured_version is the workspace version of that commit.",
      "measured_on": "2026-08-03 (date of commit 83784a92, which records the run)",
      "measured_machine": "macOS (the doc comment names no model)",
      "measured_version": "4.2.0",
      "stale_accepted": {
        "reason": "The test's own recorded measurement (83784a92, 4.2.0); re-running it writes 100 001 facts and, per its #[ignore], wants a machine at rest - not done in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "rustdoc_memory_cursor_vs_offset_row473",
      "file": "crates/velesdb-memory/src/migration/tests/performance.rs",
      "must_contain": "| 250   | 2.55 ms  | 10.2    | 9.87 ms     | 39.5    |",
      "validation_command": "cargo test -p velesdb-memory -- --ignored the_cursor_cost_per_fact_does_not_grow_like_the_offset_walk",
      "executable": false,
      "source": "commit 8ef88b6a (2026-08-05, #1816) recorded this table: 'Measured on aarch64-apple-darwin, debug, machine at rest, sequentially', with 'offset=242.21ms (121.1 us/fact)' among its rows. measured_version is the workspace version of that commit.",
      "measured_on": "2026-08-05 (date of commit 8ef88b6a, which records the run)",
      "measured_machine": "aarch64-apple-darwin, debug build, machine at rest (commit 8ef88b6a names no model)",
      "measured_version": "4.2.0",
      "stale_accepted": {
        "reason": "One configuration recorded with the change it informed (8ef88b6a, 4.2.0); re-running it needs bge-m3 served locally and a machine at rest, not done in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "rustdoc_memory_cursor_vs_offset_row474",
      "file": "crates/velesdb-memory/src/migration/tests/performance.rs",
      "must_contain": "| 500   | 4.20 ms  | 8.4     | 24.12 ms    | 48.2    |",
      "validation_command": "cargo test -p velesdb-memory -- --ignored the_cursor_cost_per_fact_does_not_grow_like_the_offset_walk",
      "executable": false,
      "source": "commit 8ef88b6a (2026-08-05, #1816) recorded this table: 'Measured on aarch64-apple-darwin, debug, machine at rest, sequentially', with 'offset=242.21ms (121.1 us/fact)' among its rows. measured_version is the workspace version of that commit.",
      "measured_on": "2026-08-05 (date of commit 8ef88b6a, which records the run)",
      "measured_machine": "aarch64-apple-darwin, debug build, machine at rest (commit 8ef88b6a names no model)",
      "measured_version": "4.2.0",
      "stale_accepted": {
        "reason": "One configuration recorded with the change it informed (8ef88b6a, 4.2.0); re-running it needs bge-m3 served locally and a machine at rest, not done in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "rustdoc_memory_cursor_vs_offset_row475",
      "file": "crates/velesdb-memory/src/migration/tests/performance.rs",
      "must_contain": "| 1 000 | 10.04 ms | 10.0    | 66.52 ms    | 66.5    |",
      "validation_command": "cargo test -p velesdb-memory -- --ignored the_cursor_cost_per_fact_does_not_grow_like_the_offset_walk",
      "executable": false,
      "source": "commit 8ef88b6a (2026-08-05, #1816) recorded this table: 'Measured on aarch64-apple-darwin, debug, machine at rest, sequentially', with 'offset=242.21ms (121.1 us/fact)' among its rows. measured_version is the workspace version of that commit.",
      "measured_on": "2026-08-05 (date of commit 8ef88b6a, which records the run)",
      "measured_machine": "aarch64-apple-darwin, debug build, machine at rest (commit 8ef88b6a names no model)",
      "measured_version": "4.2.0",
      "stale_accepted": {
        "reason": "One configuration recorded with the change it informed (8ef88b6a, 4.2.0); re-running it needs bge-m3 served locally and a machine at rest, not done in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "rustdoc_memory_cursor_vs_offset_row476",
      "file": "crates/velesdb-memory/src/migration/tests/performance.rs",
      "must_contain": "| 2 000 | 30.89 ms | 15.4    | 242.21 ms   | 121.1   |",
      "validation_command": "cargo test -p velesdb-memory -- --ignored the_cursor_cost_per_fact_does_not_grow_like_the_offset_walk",
      "executable": false,
      "source": "commit 8ef88b6a (2026-08-05, #1816) recorded this table: 'Measured on aarch64-apple-darwin, debug, machine at rest, sequentially', with 'offset=242.21ms (121.1 us/fact)' among its rows. measured_version is the workspace version of that commit.",
      "measured_on": "2026-08-05 (date of commit 8ef88b6a, which records the run)",
      "measured_machine": "aarch64-apple-darwin, debug build, machine at rest (commit 8ef88b6a names no model)",
      "measured_version": "4.2.0",
      "stale_accepted": {
        "reason": "One configuration recorded with the change it informed (8ef88b6a, 4.2.0); re-running it needs bge-m3 served locally and a machine at rest, not done in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "rustdoc_memory_reinsert_decomposition",
      "file": "crates/velesdb-memory/src/migration/tests/performance.rs",
      "must_contain": "µs/fact, vector width from 4 to 1024 ≈ 445, the bare write ≈ 15.",
      "validation_command": "cargo test -p velesdb-memory -- --ignored the_embedder_dominates_only_when_the_model_changes",
      "executable": false,
      "source": "commit 8ef88b6a (2026-08-05): 'Decomposed: text indexing ~3440 us/fact, vector width from 4 to 1024 ~445, the bare write ~15', from three measurements (88 462 / 3 900 µs, the 460 µs probe with no text, and 83784a92's attribution table). measured_version is the workspace version of that commit.",
      "measured_on": "2026-08-05 (date of commit 8ef88b6a, which records the run)",
      "measured_machine": "aarch64-apple-darwin, debug build, machine at rest (commit 8ef88b6a names no model)",
      "measured_version": "4.2.0",
      "stale_accepted": {
        "reason": "One configuration recorded with the change it informed (8ef88b6a, 4.2.0); re-running it needs bge-m3 served locally and a machine at rest, not done in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "rustdoc_memory_seeding_1242s",
      "file": "crates/velesdb-memory/src/migration/tests/performance.rs",
      "must_contain": "seeding took 1242 s for 100 001",
      "validation_command": "cargo test -p velesdb-memory --release -- --ignored past_the_ceiling_the_offset_walk_truncates_and_the_cursor_does_not",
      "executable": false,
      "source": "commit 83784a92 (2026-08-03) recorded it in the doc comment of past_the_ceiling_the_offset_walk_truncates_and_the_cursor_does_not ('Measured 2026-08-03 (macOS, 100 001 facts seeded in 1242 s)'); 6a64778e moved it here unchanged. measured_version is the workspace version of that commit.",
      "measured_on": "2026-08-03 (date of commit 83784a92, which records the run)",
      "measured_machine": "macOS (the doc comment names no model)",
      "measured_version": "4.2.0",
      "stale_accepted": {
        "reason": "The test's own recorded measurement (83784a92, 4.2.0); re-running it writes 100 001 facts and, per its #[ignore], wants a machine at rest - not done in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "timing_results_node_median",
      "file": "docs/quickstart/timing-results.md",
      "must_contain": "The Node scenario median (0.48 s)",
      "validation_command": "bash scripts/dx-timing/run_all.sh",
      "executable": false,
      "source": "commit f06468bd (2026-04-29) recorded the first reference run: 'benchmarks/dx-timing/results-2026-04-29T07-06-34Z.json — first reference run (Python 4.95 s / Rust 25.40 s / TS 0.48 s / Server 45.84 s, all under SLO)'; the doc names the host (MINGW64_NT-10.0-26200, Windows / Docker Desktop, x86_64).",
      "measured_on": "2026-04-29",
      "measured_machine": "MINGW64_NT-10.0-26200 (Windows / Docker Desktop), x86_64 (the JSON's host field)",
      "measured_version": "1.13.7",
      "stale_accepted": {
        "reason": "A frozen onboarding measurement kept with the release it installed (1.13.7), as the page says; re-running the harness installs a current release and would be a new measurement, not done in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "benchmarks_sift1m_accurate_row",
      "file": "docs/BENCHMARKS.md",
      "must_contain": "| Accurate | ~1024 | 0.9803 |",
      "validation_command": "cargo bench -p velesdb-core --bench sift1m_recall --features bench-sift1m",
      "executable": false,
      "source": "commit 4b62338f (2026-06-24, #1225) added report_recall_quality() printing RECALL_REPORT_QUALITY lines: 'Accurate (ef~1024 -> 0.9803) and Perfect (ef~8192 -> 0.9994) at 1M'; the doc names Apple Silicon (M-series, ARM64/NEON), rustc release build. measured_version is the workspace version of that commit.",
      "measured_on": "2026-06-24 (date of commit 4b62338f, which records the run)",
      "measured_machine": "Apple Silicon (M-series, ARM64/NEON), otherwise idle - docs/BENCHMARKS.md section 11.3",
      "measured_version": "3.3.0",
      "stale_accepted": {
        "reason": "Published with its version (3.3.0) and host in docs/BENCHMARKS.md section 11.3. No full 1M SIFT1M re-run has been recorded since: the nightly SIFT1M job gates recall@10 >= 0.90 at ef=128 but commits no figures, so the figure stays pinned to the run that produced it.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "benchmarks_sift1m_perfect_row",
      "file": "docs/BENCHMARKS.md",
      "must_contain": "| Perfect (exhaustive scan) ¹ | — | 0.9994 |",
      "validation_command": "cargo bench -p velesdb-core --bench sift1m_recall --features bench-sift1m",
      "executable": false,
      "source": "commit 4b62338f (2026-06-24, #1225) added report_recall_quality() printing RECALL_REPORT_QUALITY lines: 'Accurate (ef~1024 -> 0.9803) and Perfect (ef~8192 -> 0.9994) at 1M'; the doc names Apple Silicon (M-series, ARM64/NEON), rustc release build. measured_version is the workspace version of that commit.",
      "measured_on": "2026-06-24 (date of commit 4b62338f, which records the run)",
      "measured_machine": "Apple Silicon (M-series, ARM64/NEON), otherwise idle - docs/BENCHMARKS.md section 11.3",
      "measured_version": "3.3.0",
      "stale_accepted": {
        "reason": "Published with its version (3.3.0) and host in docs/BENCHMARKS.md section 11.3. No full 1M SIFT1M re-run has been recorded since: the nightly SIFT1M job gates recall@10 >= 0.90 at ef=128 but commits no figures, so the figure stays pinned to the run that produced it.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "rustdoc_perf_optimizations_import_target",
      "file": "crates/velesdb-core/src/perf_optimizations.rs",
      "must_contain": "Bulk import: 50K+ vectors/sec at 768D",
      "validation_command": "none: no benchmark or test in the repository measures this target",
      "executable": false,
      "source": "Written by the root commit 9a8ff23a (2026-01-04) under '# Performance Targets': a target, which no run in the tree records.",
      "notes": "Not a measurement: nothing here records a run.",
      "measured_on": "n/a",
      "measured_machine": "n/a",
      "measured_version": "n/a",
      "last_updated": "2026-09-13"
    },
    {
      "id": "examples_real_session_caching_savings",
      "file": "examples/real-session-benchmark/README.md",
      "must_contain": "does NOT mean 20.3%/17.2% lower $ cost",
      "validation_command": "cd examples/real-session-benchmark && node offline.mjs",
      "executable": false,
      "source": "This README's two token savings, 17.2% (lossless) and 20.3% (window-enforcement), recorded by commit e34dd2cb (2026-07-18) from node offline.mjs, which also wrote this caching note. The note said '21%/17%' until #2266: no figure in the README reads 21% (an earlier run's '21.4% saved'). measured_version is the workspace version of that commit.",
      "measured_on": "2026-07-18 (date of commit e34dd2cb, which records the run)",
      "measured_machine": "platform-independent (deterministic token counts)",
      "measured_version": "3.12.0",
      "stale_accepted": {
        "reason": "Deterministic token counts published with their campaign in examples/real-session-benchmark; re-running offline.mjs needs the Node binding built from the current tree, not done in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "rustdoc_memory_seeding_1242s_header",
      "file": "crates/velesdb-memory/src/migration/tests/performance.rs",
      "must_contain": "Measured 2026-08-03 (macOS, 100 001 facts seeded in 1242 s):",
      "validation_command": "cargo test -p velesdb-memory --release -- --ignored past_the_ceiling_the_offset_walk_truncates_and_the_cursor_does_not",
      "executable": false,
      "source": "commit 83784a92 (2026-08-03) recorded it in the doc comment of past_the_ceiling_the_offset_walk_truncates_and_the_cursor_does_not ('Measured 2026-08-03 (macOS, 100 001 facts seeded in 1242 s)'); 6a64778e moved it here unchanged. measured_version is the workspace version of that commit.",
      "measured_on": "2026-08-03 (date of commit 83784a92, which records the run)",
      "measured_machine": "macOS (the doc comment names no model)",
      "measured_version": "4.2.0",
      "stale_accepted": {
        "reason": "The test's own recorded measurement (83784a92, 4.2.0); re-running it writes 100 001 facts and, per its #[ignore], wants a machine at rest - not done in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "migrate_readme_expected_output_duration",
      "file": "crates/velesdb-migrate/README.md",
      "must_contain": "Duration:  0.12s",
      "validation_command": "velesdb-migrate run --config migration.yaml",
      "executable": false,
      "source": "commit 871d21af (2026-07-25) pasted this log as the README's expected output (timestamps 2026-07-25T16:30:04); its message is silent on it, and the README says the timings vary with the machine. measured_version is the workspace version of that commit.",
      "measured_on": "2026-07-25 (date of commit 871d21af, which records the run)",
      "measured_machine": "unknown (neither commit 871d21af nor the README names it)",
      "measured_version": "4.0.0",
      "stale_accepted": {
        "reason": "An example run's log, recorded with the README rewrite (871d21af, 4.0.0); by the README's own words its timings vary with the machine.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "rustdoc_contiguous_reorder_range",
      "file": "crates/velesdb-core/src/contiguous_ops.rs",
      "must_contain": "0.044–0.060 s in place against 0.158–0.209 s",
      "validation_command": "cargo run --release -p velesdb-core --example resident_set --features persistence -- reorder",
      "executable": false,
      "source": "commit 34cda293 (2026-08-24): 'the reorder cost is published as a range over eight runs (0.044-0.060 s in place, 0.158-0.209 s for the copying shape)'; the doc comment names a 4-vCPU Linux container. measured_version is the workspace version of that commit.",
      "measured_on": "2026-08-24 (date of commit 34cda293, which records the runs)",
      "measured_machine": "4-vCPU Linux container (the doc comment's words; commit 34cda293 says 'a shared host')",
      "measured_version": "5.2.0",
      "stale_accepted": {
        "reason": "Recorded with the change it informed (34cda293, 5.2.0); re-running the resident_set example needs the same container at rest, not done in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "rustdoc_embedder_cold_load",
      "file": "crates/velesdb-memory/src/embedder.rs",
      "must_contain": "measured on this repo's extraction model, 14.19 s cold",
      "validation_command": "none: commit d89af64d names no command",
      "executable": false,
      "source": "commit d89af64d (2026-07-26, #1598): 'Measured against this repo's documented example extraction model (qwen3.6:35b-mlx, 21.9 GB) on an idle daemon: cold 14.19 s warm 0.22 s'. measured_version is the workspace version of that commit.",
      "measured_on": "2026-07-26 (date of commit d89af64d, which records the run)",
      "measured_machine": "unknown (commit d89af64d names the model, not the machine)",
      "measured_version": "4.0.0",
      "stale_accepted": {
        "reason": "Recorded with the change it informed (d89af64d, 4.0.0); re-running it needs the 21.9 GB model served locally, not done in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "rustdoc_embedder_warm_call",
      "file": "crates/velesdb-memory/src/embedder.rs",
      "must_contain": "against 0.22 s warm",
      "validation_command": "none: commit d89af64d names no command",
      "executable": false,
      "source": "commit d89af64d (2026-07-26, #1598): 'Measured against this repo's documented example extraction model (qwen3.6:35b-mlx, 21.9 GB) on an idle daemon: cold 14.19 s warm 0.22 s'. measured_version is the workspace version of that commit.",
      "measured_on": "2026-07-26 (date of commit d89af64d, which records the run)",
      "measured_machine": "unknown (commit d89af64d names the model, not the machine)",
      "measured_version": "4.0.0",
      "stale_accepted": {
        "reason": "Recorded with the change it informed (d89af64d, 4.0.0); re-running it needs the 21.9 GB model served locally, not done in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "rustdoc_extract_runaway_generation",
      "file": "crates/velesdb-memory/src/extract.rs",
      "must_contain": "1 min 59 s spent generating JSON",
      "validation_command": "none: commit 380fd655 names no command",
      "executable": false,
      "source": "commit 380fd655 (2026-08-09, #1846): 'measured 3 933 completion tokens for a twelve-word sentence — 1 min 59 s … Capped at 600 tokens, the same call measured 14.9 s.' measured_version is the workspace version of that commit.",
      "measured_on": "2026-08-09 (date of commit 380fd655, which records the run)",
      "measured_machine": "unknown (commit 380fd655 names none)",
      "measured_version": "4.3.0",
      "stale_accepted": {
        "reason": "Recorded with the fix it informed (380fd655, 4.3.0); re-running it needs the production extractor served locally, not done in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "rustdoc_extract_capped_generation",
      "file": "crates/velesdb-memory/src/extract.rs",
      "must_contain": "14.9 s. 512 sits just under that",
      "validation_command": "none: commit 380fd655 names no command",
      "executable": false,
      "source": "commit 380fd655 (2026-08-09, #1846): 'measured 3 933 completion tokens for a twelve-word sentence — 1 min 59 s … Capped at 600 tokens, the same call measured 14.9 s.' measured_version is the workspace version of that commit.",
      "measured_on": "2026-08-09 (date of commit 380fd655, which records the run)",
      "measured_machine": "unknown (commit 380fd655 names none)",
      "measured_version": "4.3.0",
      "stale_accepted": {
        "reason": "Recorded with the fix it informed (380fd655, 4.3.0); re-running it needs the production extractor served locally, not done in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "rustdoc_service_autograph_46s",
      "file": "crates/velesdb-memory/src/service.rs",
      "must_contain": "46 s measured for a 12-word fact on the production",
      "validation_command": "none: commit ef8e33ef names no command; tests/autograph_async_bdd.rs tests the decoupling, not the timing",
      "executable": false,
      "source": "commit ef8e33ef (2026-08-09, #1846) wrote these figures with the bounded autograph queue; its message is silent on them, and the text names the production daemon (the #1839 investigation). 64 × 46 s is arithmetic on the measured 46 s. measured_version is the workspace version of that commit.",
      "measured_on": "2026-08-09 (date of commit ef8e33ef, which records the figures)",
      "measured_machine": "unknown (the text says only 'the production daemon')",
      "measured_version": "4.3.0",
      "stale_accepted": {
        "reason": "Recorded with the fix it informed (ef8e33ef, 4.3.0); re-running it needs the production extractor, not done in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "rustdoc_service_embedding_012s",
      "file": "crates/velesdb-memory/src/service.rs",
      "must_contain": "versus a 0.12 s embedding",
      "validation_command": "none: commit ef8e33ef names no command; tests/autograph_async_bdd.rs tests the decoupling, not the timing",
      "executable": false,
      "source": "commit ef8e33ef (2026-08-09, #1846) wrote these figures with the bounded autograph queue; its message is silent on them, and the text names the production daemon (the #1839 investigation). 64 × 46 s is arithmetic on the measured 46 s. measured_version is the workspace version of that commit.",
      "measured_on": "2026-08-09 (date of commit ef8e33ef, which records the figures)",
      "measured_machine": "unknown (the text says only 'the production daemon')",
      "measured_version": "4.3.0",
      "stale_accepted": {
        "reason": "Recorded with the fix it informed (ef8e33ef, 4.3.0); re-running it needs the production extractor, not done in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "rustdoc_service_queue_64x46s",
      "file": "crates/velesdb-memory/src/service.rs",
      "must_contain": "64 of them at 46 s each would hold",
      "validation_command": "none: commit ef8e33ef names no command; tests/autograph_async_bdd.rs tests the decoupling, not the timing",
      "executable": false,
      "source": "commit ef8e33ef (2026-08-09, #1846) wrote these figures with the bounded autograph queue; its message is silent on them, and the text names the production daemon (the #1839 investigation). 64 × 46 s is arithmetic on the measured 46 s. measured_version is the workspace version of that commit.",
      "measured_on": "2026-08-09 (date of commit ef8e33ef, which records the figures)",
      "measured_machine": "unknown (the text says only 'the production daemon')",
      "measured_version": "4.3.0",
      "stale_accepted": {
        "reason": "Recorded with the fix it informed (ef8e33ef, 4.3.0); re-running it needs the production extractor, not done in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "rustdoc_graph_wiring_autograph_46_52s",
      "file": "crates/velesdb-memory/src/service_graph_wiring.rs",
      "must_contain": "held every `remember` for 46-52 s",
      "validation_command": "none: commit ef8e33ef names no command; tests/autograph_async_bdd.rs tests the decoupling, not the timing",
      "executable": false,
      "source": "commit ef8e33ef (2026-08-09, #1846) wrote these figures with the bounded autograph queue; its message is silent on them, and the text names the production daemon (the #1839 investigation). 64 × 46 s is arithmetic on the measured 46 s. measured_version is the workspace version of that commit.",
      "measured_on": "2026-08-09 (date of commit ef8e33ef, which records the figures)",
      "measured_machine": "unknown (the text says only 'the production daemon')",
      "measured_version": "4.3.0",
      "stale_accepted": {
        "reason": "Recorded with the fix it informed (ef8e33ef, 4.3.0); re-running it needs the production extractor, not done in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "rustdoc_graph_wiring_embedding_012s",
      "file": "crates/velesdb-memory/src/service_graph_wiring.rs",
      "must_contain": "0.12 s — and the MCP client timed out mid-generation",
      "validation_command": "none: commit ef8e33ef names no command; tests/autograph_async_bdd.rs tests the decoupling, not the timing",
      "executable": false,
      "source": "commit ef8e33ef (2026-08-09, #1846) wrote these figures with the bounded autograph queue; its message is silent on them, and the text names the production daemon (the #1839 investigation). 64 × 46 s is arithmetic on the measured 46 s. measured_version is the workspace version of that commit.",
      "measured_on": "2026-08-09 (date of commit ef8e33ef, which records the figures)",
      "measured_machine": "unknown (the text says only 'the production daemon')",
      "measured_version": "4.3.0",
      "stale_accepted": {
        "reason": "Recorded with the fix it informed (ef8e33ef, 4.3.0); re-running it needs the production extractor, not done in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "perf_slo_insert_ci_variance",
      "file": "docs/reference/PERFORMANCE_SLO.md",
      "must_contain": "(7.86–10.42s across 5 runs)",
      "validation_command": "cargo bench -p velesdb-core --bench smoke_test -- --noplot",
      "executable": false,
      "source": "commit 39a06f8e (2026-03-19, #333): 'IO-bound, 7.86-10.42s variance on shared CI runners'; the same diff's benchmarks/baseline.json lists '7.86/9.87/9.47/8.80/10.42 across 5 CI runs' on ubuntu-latest GitHub Actions runners (2-core AMD). measured_version is the workspace version of that commit.",
      "measured_on": "2026-03-19 (date of commit 39a06f8e, which records the runs)",
      "measured_machine": "GitHub Actions ubuntu-latest (2-core AMD), per the baseline.json note of the same diff",
      "measured_version": "1.5.1",
      "stale_accepted": {
        "reason": "CI-runner figures recorded with the threshold change (39a06f8e, 1.5.1); the smoke bench runs in CI's perf-smoke job, which keeps its own baseline.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "perf_slo_insert_local_836",
      "file": "docs/reference/PERFORMANCE_SLO.md",
      "must_contain": "`smoke_insert/10k_128d`: 8.36s",
      "validation_command": "cargo bench -p velesdb-core --bench smoke_test -- --noplot",
      "executable": false,
      "source": "commit 399fc2fc (2026-03-25, release v1.7.2) wrote it; its message is silent, and the line then read 'Local validation (i9-14900KF)'. measured_version is the workspace version of that commit.",
      "measured_on": "2026-03-25 (date of commit 399fc2fc, which records the run)",
      "measured_machine": "Intel Core i9-14900KF (the line's original words)",
      "measured_version": "1.7.2",
      "stale_accepted": {
        "reason": "A local run recorded with the 1.7.2 release (399fc2fc); re-measuring it belongs on the idle i9-14900KF reference machine.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-13"
    },
    {
      "id": "quantization_arena_optout_build_time",
      "file": "docs/guides/QUANTIZATION.md",
      "must_contain": "most 1.35 s of build time",
      "validation_command": "cargo run --release --example resident_set --features persistence -- write",
      "executable": false,
      "source": "commit adc85f5b (#2128, 2026-08-24, 'decide the arena is not configurable, and say why') added the `write` mode of examples/resident_set.rs and this table: filling 100 000 x 768-d into each backing, no graph and no quantizer - heap 58-62 ms, stable across runs; file-mapped 0.30-1.35 s, rising with each consecutive run (five runs). The commit names no machine. measured_version is the workspace version of that commit. This line restates the worst of those runs.",
      "measured_on": "2026-08-24 (date of commit adc85f5b, which records the run)",
      "measured_machine": "unknown (commit adc85f5b names none)",
      "measured_version": "5.2.0",
      "stale_accepted": {
        "reason": "Recorded with the decision it informed (adc85f5b, 5.2.0); re-running the `write` mode five times was not done in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-14"
    },
    {
      "id": "quantization_arena_fill_counterweight",
      "file": "docs/guides/QUANTIZATION.md",
      "must_contain": "the arena fill time (1.35 s at worst)",
      "validation_command": "cargo run --release --example resident_set --features persistence -- write",
      "executable": false,
      "source": "commit adc85f5b (#2128, 2026-08-24, 'decide the arena is not configurable, and say why') added the `write` mode of examples/resident_set.rs and this table: filling 100 000 x 768-d into each backing, no graph and no quantizer - heap 58-62 ms, stable across runs; file-mapped 0.30-1.35 s, rising with each consecutive run (five runs). The commit names no machine. measured_version is the workspace version of that commit. This line restates the worst of those runs.",
      "measured_on": "2026-08-24 (date of commit adc85f5b, which records the run)",
      "measured_machine": "unknown (commit adc85f5b names none)",
      "measured_version": "5.2.0",
      "stale_accepted": {
        "reason": "Recorded with the decision it informed (adc85f5b, 5.2.0); re-running the `write` mode five times was not done in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-14"
    },
    {
      "id": "rustdoc_memory_million_fact_rebuild",
      "file": "crates/velesdb-memory/src/migration/tests/performance.rs",
      "must_contain": "rebuild moves from roughly 56 minutes to roughly 16 seconds",
      "validation_command": "none: arithmetic, no run",
      "executable": false,
      "source": "Written by commit a4950038 (2026-08-03): arithmetic on the per-fact costs it measured (3345.8 → 16.3 µs/fact) times 10⁶ facts; no million-fact run exists (the test writes 31 000 facts in total).",
      "notes": "Not a measurement: an extrapolation the line now says is one.",
      "measured_on": "n/a",
      "measured_machine": "n/a",
      "measured_version": "n/a",
      "last_updated": "2026-09-13"
    },
    {
      "id": "rustdoc_memory_offset_ceiling_21min",
      "file": "crates/velesdb-memory/src/migration/tests/performance.rs",
      "must_contain": "and it takes ~21 minutes",
      "validation_command": "cargo test -p velesdb-memory --release -- --ignored past_the_ceiling_the_offset_walk_truncates_and_the_cursor_does_not",
      "executable": false,
      "source": "The seeding of commit 83784a92's run ('Measured 2026-08-03 (macOS, 100 001 facts seeded in 1242 s)'), stated in minutes; 6a64778e moved it here unchanged. measured_version is the workspace version of that commit.",
      "measured_on": "2026-08-03 (date of commit 83784a92, which records the run)",
      "measured_machine": "macOS (the doc comment names no model)",
      "measured_version": "4.2.0",
      "stale_accepted": {
        "reason": "The test's own recorded measurement (83784a92, 4.2.0); re-running it writes 100 001 facts and, per its #[ignore], wants a machine at rest - not done in this change.",
        "granted": "2026-09-13",
        "for_workspace_major": 6
      },
      "last_updated": "2026-09-14"
    },
    {
      "id": "hnsw_batch_pool_handoff_flat_cost",
      "file": "crates/velesdb-core/src/index/hnsw/native/backend_adapter.rs",
      "must_contain": "a flat ~32 us of thread hand-off, measured",
      "validation_command": "n/a — rayon thread hand-off cost, not a repository benchmark",
      "executable": false,
      "source": "Two measurements, kept apart because conflating them is what made the first version of this claim wrong. (1) The hand-off: a standalone rayon micro-benchmark, release, 2 000 iterations after warm-up, shows `ThreadPool::install` costs a flat ~32 us to a caller outside the pool, independent of the work it wraps. (2) What a batch actually costs: `HnswIndex::insert_batch_parallel` itself, 16 dimensions, release, medians of 20 runs on a fresh index — n=1 250 ns, n=10 7 958 ns, n=50 161 333 ns, n=99 1 283 125 ns, n=1 000 6 457 916 ns. The hand-off is therefore 128x the work at n=1, 4x at n=10, 20% at n=50 and 2.5% at n=99, which is why `insert_batch_parallel` installs the pool only at or above `PARALLEL_BATCH_MIN`. An earlier version of this claim compared the hand-off against a SYNTHETIC loop rather than against this call and reported 61x at n=10; that figure was wrong and is not used anywhere. The ratio also falls with dimension: docs/BENCHMARKS.md measures ~300 us per vector at 768D against ~6.5 us here. A platform- and dimension-dependent scheduling cost, not a product promise.",
      "last_updated": "2026-09-18",
      "measured_on": "2026-09-18",
      "measured_machine": "Apple M5 Pro (18 cores, macOS 26.6.2, release, idle)",
      "measured_version": "6.0.0"
    }
  ],
  "claim_families": [
    {
      "id": "binary_size",
      "canonical_claim_id": "readme_binary_size_9mb",
      "canonical_value": "~14 MB",
      "members": [
        {
          "file": "README.md",
          "value": "~14 MB",
          "must_contain": "One ~14 MB binary"
        },
        {
          "file": "README.md",
          "value": "~14 MB",
          "must_contain": "| **Binary size** | ~14 MB |"
        },
        {
          "file": "docs/README.md",
          "value": "~14 MB",
          "must_contain": "One ~14 MB binary"
        },
        {
          "file": "docs/WHY_VELESDB.md",
          "value": "~14 MB",
          "must_contain": "single embeddable ~14 MB binary"
        },
        {
          "file": "ARCHITECTURE.md",
          "value": "~14 MB",
          "must_contain": "single ~14 MB binary"
        },
        {
          "file": "docs/BUSINESS_MODEL.md",
          "value": "~14 MB",
          "must_contain": "single binary (~14 MB)"
        },
        {
          "file": "docs/guides/BUSINESS_SCENARIOS.md",
          "value": "~14 MB",
          "must_contain": "**~14 MB binary**"
        },
        {
          "file": "AGENTS.md",
          "value": "~14 MB",
          "must_contain": "single ~14 MB binary"
        },
        {
          "file": "CLAUDE.md",
          "value": "~14 MB",
          "must_contain": "single ~14 MB binary"
        }
      ]
    },
    {
      "id": "wasm_bundle_size",
      "canonical_claim_id": "readme_wasm_bundle_gzipped",
      "canonical_value": "~710 KB",
      "members": [
        {
          "file": "README.md",
          "value": "~710 KB",
          "must_contain": "WASM build is ~710 KB gzipped"
        },
        {
          "file": "README.md",
          "value": "~710 KB",
          "must_contain": "WASM, ~710 KB gzipped"
        },
        {
          "file": "docs/wasm/bundle-optimization.md",
          "value": "~710 KB",
          "must_contain": "| Release (gzipped) | < 200 KB | ~710 KB |"
        },
        {
          "file": "docs/wasm/bundle-optimization.md",
          "value": "~710 KB",
          "must_contain": "measured figure (~710 KB gzipped,"
        },
        {
          "file": "docs/wasm/bundle-optimization.md",
          "value": "~710 KB",
          "must_contain": "artifact (~710 KB gzipped, ~2.0 MB raw)"
        }
      ]
    },
    {
      "id": "rest_endpoint_count",
      "canonical_claim_id": "readme_rest_endpoint_count",
      "canonical_value": "54 REST endpoints",
      "members": [
        {
          "file": "README.md",
          "value": "54 REST endpoints",
          "must_contain": "benchmark figures, \"54 REST endpoints\""
        },
        {
          "file": "README.md",
          "value": "54 REST endpoints",
          "must_contain": "**REST:** 54 REST endpoints"
        },
        {
          "file": "README.md",
          "value": "54 REST endpoints",
          "must_contain": "| 54 REST endpoints —"
        },
        {
          "file": "README.md",
          "value": "54 REST endpoints",
          "must_contain": "Serving it over HTTP — the 54 REST endpoints"
        },
        {
          "file": "docs/reference/ARCHITECTURE.md",
          "value": "54 REST endpoints",
          "must_contain": "- 54 REST endpoints (61 method+path operations)"
        },
        {
          "file": "docs/NEW_USER_TROUBLESHOOTING.md",
          "value": "54 REST endpoints",
          "must_contain": "All 54 REST endpoints with request/response examples"
        }
      ]
    }
  ]
}