Skip to content

data: add independent DeepSeek V4 Pro anchoring replication - #5

Open
NineThoughts0521 wants to merge 2 commits into
xiaobright:mainfrom
NineThoughts0521:codex/deepseek-v4-independent-replication
Open

data: add independent DeepSeek V4 Pro anchoring replication#5
NineThoughts0521 wants to merge 2 commits into
xiaobright:mainfrom
NineThoughts0521:codex/deepseek-v4-independent-replication

Conversation

@NineThoughts0521

Copy link
Copy Markdown

Summary

This PR contributes an independent third-party replication by @NineThoughts0521 of the DeepSeek V4 Pro first-request tool-schema anchoring result on frozen Project2 V4.1b.

These runs remain separate from the maintainer's formal n, rankings, worst scores, means, and sample index. The frozen task, verifier, evaluator, and scoring rules are unchanged.

Results

All DSH runs used DeepSeek V4 Pro, reasoning max, DSH 0.1.0-rc.6, the same candidate prompt, a reset frozen workspace, and the original V4.1b evaluator.

Harness / preset Ability Ship Class Hidden ESP static Account cost
DSH Anchored Standard 96 96 A 44/45 9/9 ¥2.58
DSH Standard 89 89 B+ 43/45 7/9 ¥1.64
DSH Minimal-Full 85.5 85.5 B+ 42/45 6/9 ¥1.13
OpenCode 1.18.17 replacement 93 93 B+ 43/45 8/9 ¥1.53

Observed descriptive contrasts:

  • Anchored - Standard: +7
  • Anchored - Minimal-Full: +10.5
  • Standard - Minimal-Full: +3.5

Minimal-Full ablation

Minimal-Full retains the Anchored Minimal complete system condition, complete: true, includeRuntimeContext: false, and the Standard capability roster, but removes the first-request bootstrap.

The zero-cost schema gate confirms that Minimal-Full request 1 tools equal both Standard request 1 and Anchored request 2; Minimal-Full and Anchored also have equal system and request 1 non-tool hashes. Anchored request 1 exposes only pwsh/read, then restores the full 25-tool catalog.

The result supports an additional contribution from first-request tool-schema anchoring on this task and is consistent with the maintainer's original Anchored-over-Standard direction. It does not establish universal causality: each condition has one run, and the Anchored treatment also includes the catalog transition and its timing.

OpenCode disclosure

The original OpenCode run was interrupted by the outer process lifecycle after model activity had started. It is permanently retained as an unscored infrastructure failure, with its ¥0.24 cost and partial trajectory aggregate.

One explicitly approved replacement changed only the process lifecycle to a detached/background wrapper. It used OpenCode 1.18.17, direct DeepSeek, deepseek-v4-pro, --variant max, and https://api.deepseek.com. The replacement scored 93 and was not rerun. OpenCode remains a post-preregistered exploratory harness comparison and is excluded from the DSH mechanism ablation.

Cost, evidence, and privacy

Total account cost, including the partial run, was ¥7.12. Usage recomputation using the official live CNY prices applicable to each run was ¥7.075304.

Public artifacts include scores, token/time/cost aggregates, first-request tool-catalog snapshots, catalog transitions, trajectory fingerprints, balance window hashes, infrastructure events, and an evidence manifest. Trajectory wording is treated only as a behavior fingerprint.

Raw DSH/OpenCode sessions, full reasoning/CoT, credentials, and private absolute paths remain local and Git-ignored. The public manifest records raw evidence SHA-256 values without publishing raw content.

Validation

  • Zero-cost mock schema gate: all 7 assertions passed.
  • Public JSON parsing and Python script compilation passed.
  • Candidate prompt and public artifact hashes were verified.
  • Markdown links, privacy scans, and git diff --check passed.

The optional real ESP-IDF build was not run. DeepSWE and Terminal-Bench were not executed, so cross-task generalization remains an open question.

- 增加 Minimal-Full 消融 preset 与零成本 schema gate
- 发布 3+1 结果、成本、轨迹聚合和证据哈希
- 保留 OpenCode partial/replacement 与第三方统计边界
- 区分维护者正式统计与 NineThoughts0521 独立结果
- 接入 Minimal-Full、trajectory、成本与证据限制
- 如实披露 OpenCode partial 和 replacement
Copilot AI lite review requested due to automatic review settings August 16, 2026 10:29

Copilot AI left a comment

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Pull request overview

This PR adds a third-party, independent replication package for DeepSeek V4 Pro “first-request tool-schema anchoring” on the frozen Project2 V4.1b baseline, including a new minimal-full ablation preset plus scripts and published (privacy-preserving) derived artifacts.

Changes:

  • Add minimal-full DSH preset (full Standard tool catalog from request 1, Minimal complete system condition) and document its role in the ablation.
  • Add an experiments/deepseek-v4-pro-anchoring/ bundle (runners, schema gate, analyzers, comparison builder, and derived/public artifacts).
  • Update top-level docs/scoreboard indices to reference the independent replication, and extend .gitignore to keep private evidence out of Git.

Reviewed changes

Copilot reviewed 55 out of 60 changed files in this pull request and generated 4 comments.

Show a summary per file
File Description
tools/deepseek-harness-presets/README.md Documents the new minimal-full preset and its validation requirements.
tools/deepseek-harness-presets/minimal-full/preset.yml Registers the minimal-full preset (name/description/order).
tools/deepseek-harness-presets/minimal-full/agent.cordis.yml Defines the minimal-full agent-plane composition (Minimal system condition + Standard tool roster).
README.md Adds a section summarizing the independent replication and links to the report and experiment directory.
experiments/deepseek-v4-pro-anchoring/scripts/validate_presets.py Validates schema-gate request snapshots via hashes and tool catalogs.
experiments/deepseek-v4-pro-anchoring/scripts/run_project2.ps1 Automates DSH runs + evaluator execution + session analysis for Project2.
experiments/deepseek-v4-pro-anchoring/scripts/run_opencode_replacement_detached.cmd Runs the OpenCode replacement runner in a detached wrapper style with redirected logs.
experiments/deepseek-v4-pro-anchoring/scripts/run_opencode_agent.py Runs OpenCode against DeepSeek API with balance polling and exports a private session artifact.
experiments/deepseek-v4-pro-anchoring/scripts/run_mock_gate.ps1 Executes a zero-cost mock schema gate for Standard/Minimal-Full/Anchored tool-catalog assertions.
experiments/deepseek-v4-pro-anchoring/scripts/mock_deepseek_server.mjs Local mock DeepSeek-compatible SSE server used for the schema gate.
experiments/deepseek-v4-pro-anchoring/scripts/launch_detached.py Helper to launch a Python runner with stdout/stderr redirected.
experiments/deepseek-v4-pro-anchoring/scripts/extract_candidate_prompt.py Extracts the candidate prompt block from CANDIDATE_PROMPT.md and emits its SHA-256.
experiments/deepseek-v4-pro-anchoring/scripts/dsh_session_driver.mjs Drives DSH Web via RPC to run a prompt and export session history/meta.
experiments/deepseek-v4-pro-anchoring/scripts/collect_price.py Captures a DeepSeek pricing-page snapshot and writes a structured price record.
experiments/deepseek-v4-pro-anchoring/scripts/collect_balance.py Captures a DeepSeek balance snapshot and writes a structured balance record.
experiments/deepseek-v4-pro-anchoring/scripts/capture_baseline.py Captures environment + pinned versions + preset tree hashes for reproducibility.
experiments/deepseek-v4-pro-anchoring/scripts/calculate_balance_delta.py Computes per-window balance deltas between two balance snapshots.
experiments/deepseek-v4-pro-anchoring/scripts/build_evidence_manifest.py Builds a manifest of public artifact hashes and private-raw hash pointers.
experiments/deepseek-v4-pro-anchoring/scripts/build_comparison.py Aggregates per-run artifacts into a machine-readable comparison JSON.
experiments/deepseek-v4-pro-anchoring/scripts/analyze_session.py Derives redacted/public stats from a DSH session JSONL and evaluator summary.
experiments/deepseek-v4-pro-anchoring/scripts/analyze_opencode_session.py Derives redacted/public stats from an OpenCode export + gate/price/balance inputs.
experiments/deepseek-v4-pro-anchoring/run-matrix.json Declares the run plan/status matrix for the 3+1 bundle.
experiments/deepseek-v4-pro-anchoring/RESULTS.md Publishes the summarized 3+1 results table and bounded conclusions.
experiments/deepseek-v4-pro-anchoring/README.md Provides experiment directory overview and privacy/evidence boundaries.
experiments/deepseek-v4-pro-anchoring/preregistration.json Captures preregistration metadata, pins, budget, and amendments.
experiments/deepseek-v4-pro-anchoring/mock-prompt.txt Prompt used for the mock schema gate.
experiments/deepseek-v4-pro-anchoring/artifacts/schema-gate.json Published output of the mock schema gate assertions.
experiments/deepseek-v4-pro-anchoring/artifacts/runs/P2-20260815-04b-opencode-replacement.json Published derived artifact for the OpenCode replacement run.
experiments/deepseek-v4-pro-anchoring/artifacts/runs/P2-20260815-04b-opencode-replacement-balance.json Published balance delta for the OpenCode replacement run.
experiments/deepseek-v4-pro-anchoring/artifacts/runs/P2-20260815-04-opencode-partial.json Published derived artifact for the preserved partial OpenCode run.
experiments/deepseek-v4-pro-anchoring/artifacts/runs/P2-20260815-04-opencode-partial-balance.json Published balance delta for the preserved partial OpenCode run.
experiments/deepseek-v4-pro-anchoring/artifacts/runs/P2-20260815-03-minimal-full-balance.json Published balance delta for the Minimal-Full run.
experiments/deepseek-v4-pro-anchoring/artifacts/runs/P2-20260815-02-standard-balance.json Published balance delta for the Standard run.
experiments/deepseek-v4-pro-anchoring/artifacts/runs/P2-20260815-01-anchored-balance.json Published balance delta for the Anchored run.
experiments/deepseek-v4-pro-anchoring/artifacts/price-snapshot.json Published pricing snapshot used for cost recomputation.
experiments/deepseek-v4-pro-anchoring/artifacts/price-P2-20260815-04b-opencode-replacement.json Per-run pricing snapshot for OpenCode replacement.
experiments/deepseek-v4-pro-anchoring/artifacts/price-P2-20260815-04-opencode.json Per-run pricing snapshot for initial OpenCode attempt.
experiments/deepseek-v4-pro-anchoring/artifacts/price-P2-20260815-03-minimal-full.json Per-run pricing snapshot for Minimal-Full.
experiments/deepseek-v4-pro-anchoring/artifacts/price-P2-20260815-02-standard.json Per-run pricing snapshot for Standard.
experiments/deepseek-v4-pro-anchoring/artifacts/price-P2-20260815-01-anchored.json Per-run pricing snapshot for Anchored.
experiments/deepseek-v4-pro-anchoring/artifacts/opencode-gate.json Published gate confirming OpenCode configuration/tool catalog assumptions.
experiments/deepseek-v4-pro-anchoring/artifacts/metadata-corrections.json Records post-run corrections to derived metadata (with reasons).
experiments/deepseek-v4-pro-anchoring/artifacts/infrastructure-events.json Records infra events and retry boundaries for the run set.
experiments/deepseek-v4-pro-anchoring/artifacts/evidence-manifest.json Published manifest of public artifact hashes and private evidence hashes.
experiments/deepseek-v4-pro-anchoring/artifacts/environment-baseline.json Published environment/pin snapshot including preset tree hashes.
experiments/deepseek-v4-pro-anchoring/artifacts/deepswe-gate.json Published environment gate output for deferred DeepSWE pairing.
experiments/deepseek-v4-pro-anchoring/artifacts/balance-events.json Published balance window events and per-run window hashes/deltas.
evaluator/trajectory_evidence/README.md Clarifies separation between maintainer baseline evidence and third-party replication aggregates.
evaluator/reports/v4.1b_scoreboard.md Adds a “third-party replication” section excluded from formal n.
evaluator/reports/README.md Adds the independent replication report link to the report index.
docs/v4.1/README.md Adds the independent replication report link to the v4.1 doc index.
docs/v4.1/DEEPSEEK_V4_TRIGGER_MECHANISM_EXPERIMENTS_20260814.md Adds an update section referencing the third-party ablation/replication.
docs/v4.1/DEEPSEEK_V4_TRAJECTORY_ANALYSIS_20260814.md Adds fingerprint summary rows for the third-party runs.
docs/v4.1/DEEPSEEK_V4_PRO_INDEPENDENT_REPLICATION_20260815.md New full report describing methodology, boundaries, results, and limitations.
docs/v4.1/DEEPSEEK_V4_PRO_HARNESS_ANALYSIS_20260814.md Adds an appendix referencing the third-party independent replication findings.
.gitignore Ignores private experiment evidence directories and environment files.
Suppressed comments (1)

experiments/deepseek-v4-pro-anchoring/scripts/validate_presets.py:28

  • load() iterates value before confirming the JSON root is a list, so a non-list input will raise TypeError instead of the intended validation error. Validate the type first, and (optionally) filter only dict entries before accessing .get().

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

Comment on lines +13 to +20
def request_fingerprint(request: dict) -> dict:
tools = request.get("tools", [])
return {
"system_sha256": hashlib.sha256(canonical(request.get("messages", [])[0])).hexdigest(),
"tool_names": [item.get("function", {}).get("name") for item in tools],
"tools_sha256": hashlib.sha256(canonical(tools)).hexdigest(),
"non_tools_sha256": hashlib.sha256(canonical({key: value for key, value in request.items() if key != "tools"})).hexdigest(),
}
Comment on lines +15 to +23
def credential() -> str:
value = os.environ.get("DEEPSEEK_API_KEY")
if value:
return value
path = Path(os.environ.get("DSH_HOME", Path.home() / ".dsh")) / ".credentials.yaml"
for line in path.read_text(encoding="utf-8").splitlines():
if line.startswith("DEEPSEEK_API_KEY:"):
return line.split(":", 1)[1].strip().strip("'\"")
raise RuntimeError("未找到 DSH DeepSeek credential")
Comment on lines +20 to +25
$sourceHashes = Get-ChildItem -Recurse -File $source | ForEach-Object {
[pscustomobject]@{ Relative = [IO.Path]::GetRelativePath($source, $_.FullName); Hash = (Get-FileHash $_.FullName -Algorithm SHA256).Hash }
}
$targetHashes = Get-ChildItem -Recurse -File $target | ForEach-Object {
[pscustomobject]@{ Relative = [IO.Path]::GetRelativePath($target, $_.FullName); Hash = (Get-FileHash $_.FullName -Algorithm SHA256).Hash }
}
Comment on lines +24 to +32
def credential() -> str:
value = os.environ.get("DEEPSEEK_API_KEY")
if value:
return value
path = Path(os.environ.get("DSH_HOME", Path.home() / ".dsh")) / ".credentials.yaml"
for line in path.read_text(encoding="utf-8").splitlines():
if line.startswith("DEEPSEEK_API_KEY:"):
return line.split(":", 1)[1].strip().strip("'\"")
raise RuntimeError("未找到 DSH DeepSeek credential")
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants