diff --git a/DEVLOG.md b/DEVLOG.md
index 003f723..e904f37 100644
--- a/DEVLOG.md
+++ b/DEVLOG.md
@@ -4,6 +4,36 @@
 
 ---
 
+## 2026-09-30：AC1b 兩次嘗試失敗；修正各 seed 選不同聚合時的畫圖缺陷，新增嘗試帳本
+
+### 本次工作 / 執行摘要
+- 狀態不變：RQ1 到 RQ5 完成；**Tier 1 驗收未完成，AC1b 尚未通過**。
+- **AC1b 第 1 次**（`3992840ddb3f`，2026-09-29 07:40 到 14:24 UTC）：訓練、AC2、pilot 全過；Haiku 前的 `count_tokens` 在 Anthropic 認證事故期間連 5 次 HTTP 503，花費 US$0。comparison 機械判定 FAIL；Drew 判定為基礎設施事故、不視為 AC1b 失敗，帳本記為 `INFRASTRUCTURE INTERRUPTED`，同一 id 與身分續跑。
+- **AC1b 第 2 次**（續跑，2026-09-30 01:13 到 01:31 UTC）：Haiku 8,600/8,600、US$3.178751、0 parse failed、20 筆預測與原 run 不同；analysis、latency、cost 完成；`make figures` 失敗：ModernBERT k=100 三個 seed 在 validation 選了 argmax、summed、summed，`figures.selected_aggregation` 直接丟 ValueError。原始 run 三個 seed 都選 argmax，所以這條路徑從沒跑過。判 FAIL（程式缺陷）。
+- **修正**：找過所有讀聚合方式的地方。`analysis` 本來就讓每個 seed 用自己 validation 選的聚合（`final`、`diagnostics` 逐 seed 算好再合併）；`report` 的聚合說明已會列出逐 seed 選擇；`cost`、`comparison`、學習曲線、router 圖、門檻圖都讀 `final` 或 `diagnostics`，不受影響。只有 risk-coverage 圖假設三個 seed 同一種。`curves.json` 對每種聚合只存三 seed 的 mean 與 std，沒有逐 seed 曲線，而補存會改動原始 `curves.json` 的位元組，所以改成：seed 一致的 k 維持一個面板；不一致的 k 每種被選到的聚合各一個面板（固定 argmax、summed 順序），標題寫出哪些 seed 的 final router 用它；README 圖說在這種情況自動加一句說明。沒有任何選擇用到 test，也沒有固定成某一種。
+- **帳本**：`docs/ac1b/attempts.json` 記錄每次嘗試（commit、id、UTC 起訖、結果與原因、實際 Haiku 支出），`docs/ac1b/attempt-<n>/` 放 comparison 與事故說明。comparison 與 README 的預算改成三部分：原始實驗 US$3.19（固定）、每次嘗試支出與原因、reproduction-validation 累計。PLAN §5.1 預算規則補一行：每次嘗試各自獨立 US$5 上限（Drew，2026-09-30）。
+- **聚合選擇只列不判**（第 3 次嘗試實跑前寫定，PLAN §5.1 實跑前補充追加一行）：comparison 新增一節，對每個有 final router 的組（25 組）列出原始與重跑每個 seed 在 validation 選的 8 類聚合與是否相同，狀態固定 `LISTED, NOT JUDGED`，不計入 REVIEW、不影響判定。用第 2 次重跑輸出實測：7 組不同，其中 ModernBERT k=100 原始 argmax/argmax/argmax、重跑 argmax/summed/summed（seeds 43、44 改選），可解釋 small-only 91.87 → 91.42 的 REVIEW；REVIEW 計數仍是 192。
+- **帳本結果分類（PR #20 審查後）**：依 Drew 2026-09-29 的決定（由協調者轉達），第 1 次記為 `INFRASTRUCTURE INTERRUPTED`（`failure_kind: infrastructure`），附外部事故證據（status.claude.com 事故、2026-09-29 14:21 UTC、503 credential validation failed）與決定原文；comparison 的機械判定 FAIL 另存為 `comparison_verdict`，證據檔不改寫。收緊條件：`failure_kind` 只能是 infrastructure、program defect、ac2；INTERRUPTED 只允許 infrastructure 且需證據與決定；同一 infrastructure 錯誤類型連續第二次必須是 `INFRASTRUCTURE BLOCKED`（AC1b 暫停）；程式缺陷只能是 FAIL。AC1b 最終判定仍只有 PASS 或 FAIL：INTERRUPTED 不算失敗也不算通過，只有某次嘗試 PASS 才完成驗收。risk-coverage 混合面板標題改為「mean of 3 seeds; final router of seed …」。
+- 下一步依 Drew 決定：合併後在新 commit 完整重跑 AC1b（新的一次嘗試，新 id、新 journal、獨立 US$5）；不拿第 2 次的輸出配新 commit 的圖。
+
+### 核心發現 / 數據
+- 用第 2 次重跑輸出的副本跑 `figures` 與 `report`（`--results-root`、`--readme-out`）：4 張圖與 report 都完成；risk-coverage 為 (a) k=10 summed、(b) k=100 argmax（seed 42）、(c) k=100 summed（seeds 43、44）。
+- 第 2 次重跑中其他聚合不一致的組：BERT k=10、25、50、100，ModernBERT k=25，OOS 0 消融 k=100，TF-IDF k=5、10、25。
+- 原始結果上 `make figures`、`make report` 兩次，`results/` 與 README 生成區塊逐位元組不變。
+
+### Blockers / 遇到的問題
+- (無)
+
+### Next
+- [ ] 合併後在新 commit 完整重跑 AC1b（第 3 次嘗試），結果寫進 `docs/ac1b/attempts.json`
+
+### Files / Budget
+- 新增：`src/tinyrouter/ac1b_ledger.py`；`tests/test_ac1b_ledger.py`；`docs/ac1b/`（`attempts.json`、`README.md`、`attempt-1/`、`attempt-2/`）
+- 修改：`src/tinyrouter/figures.py`、`report.py`、`comparison.py`；`tests/test_figures.py`、`test_report.py`、`test_comparison.py`；`README.md`、`docs/PLAN.md`、`DEVLOG.md`
+- API 花費：本次 US$0。AC1b 累計（reproduction-validation）：US$3.178751；原始實驗成本維持 US$3.19
+
+---
+
 ## 2026-09-29（深夜，七）：PR #19 審查修正（R1 到 R6），AC1b 實跑前固定比較範圍
 
 ### 本次工作 / 執行摘要
diff --git a/README.md b/README.md
index 4a9b26f..8365087 100644
--- a/README.md
+++ b/README.md
@@ -237,7 +237,25 @@ Training on Apple MPS is not bit-for-bit deterministic, so a rerun is not expect
 
 A clean clone is not a clean cache: the base models and the CLINC150 files come from the shared Hugging Face cache (`~/.cache/huggingface` by default). That does not affect correctness, since every model is pinned to a revision and every data file is checked against its SHA-256, but to download everything afresh point the cache at an empty directory, for example `HF_HOME=/tmp/tinyrouter-hf make reproduce`.
 
-**Status: AC1b has not been run yet.** Until it has passed, the status is RQ1 to RQ5 complete, Tier 1 acceptance not complete.
+**Status: AC1b has not passed yet.** It has been attempted twice at `3992840ddb3f`: attempt 1 was interrupted by an external authentication incident (recorded as infrastructure-interrupted, neither a failure nor a pass), and attempt 2 failed on a defect in the figure code when seeds chose different 8-way aggregations; every attempt is recorded in [`docs/ac1b/`](docs/ac1b/README.md). Until it has passed, the status is RQ1 to RQ5 complete, Tier 1 acceptance not complete.
+
+AC1b budget. Every attempt has its own US$5 Haiku cap, enforced per reproduction id:
+
+<!-- BEGIN AC1b budget, generated from docs/ac1b/attempts.json; do not edit by hand -->
+- Original experiment (AC6, fixed): US$3.19
+- AC1b attempts (reproduction-validation, each capped at US$5):
+
+| attempt | reproduction id | result | reason | Haiku spend |
+|---|---|---|---|---|
+| 1 | `3992840ddb3f` | INFRASTRUCTURE INTERRUPTED | infrastructure: Haiku count_tokens returned HTTP 503 credential validation failed 5 times during an Anthropic authentication incident; no classification request sent | US$0.000000 |
+| 2 | `3992840ddb3f` | FAIL | program defect: make figures stopped because seeds chose different 8-way aggregations for modernbert/k100 (argmax, summed, summed); Haiku 8600/8600, 0 parse failures, 20 predictions differ | US$3.178751 |
+
+- Reproduction-validation total (all AC1b attempts): US$3.178751
+
+Reproduction-validation spend is never reported as original experiment cost and does not change the AC6 conclusion (Haiku spend at most US$5).
+
+INFRASTRUCTURE INTERRUPTED is neither an AC1b failure nor a pass; AC1b acceptance is completed only by an attempt whose result is PASS.
+<!-- END AC1b budget -->
 
 ## Reproducibility notes
 
diff --git a/docs/PLAN.md b/docs/PLAN.md
index b7f9709..712837e 100644
--- a/docs/PLAN.md
+++ b/docs/PLAN.md
@@ -127,9 +127,11 @@ TinyRouter 問三件事：**小模型要多少標註資料才夠？它知不知
 
 - 判定範圍（偏離原始 mean ± std 標 `REVIEW REQUIRED`）固定為：README 首屏數字；README router 表（LLM-only，以及 ModernBERT k=10 與 k=100 的 small-only、hybrid 目標 2% 與 5%、oracle，各報 8 類準確率、OOS recall、高信心 OOS 誤派、LLM 呼叫率）；學習曲線（兩個 encoder、每個 k 的 8 類準確率與 OOS recall）；`ablation_comparison`（OOS 250 對 0 的 AUROC、AUPRC 與 router 行為）；所有門檻診斷統計。
 - **只列不判**：延遲（`results/efficiency/*`）、k=100 的訓練時間與峰值記憶體、成本模型（`results/cost/cost.json`）隨機器變動。comparison 列出原始值、重跑值與差異，狀態固定為 `LISTED, NOT JUDGED`，不計入 `REVIEW REQUIRED`，也不影響判定。
+- **聚合選擇只列不判**（2026-09-30，第 2 次嘗試觀察到差異後追加，第 3 次嘗試實跑前寫定）：每個有 final router 的組，comparison 列出原始與重跑每個 seed 在 validation 選的 8 類聚合方式與是否相同，狀態固定為 `LISTED, NOT JUDGED`，不計入 `REVIEW REQUIRED`，也不影響判定。
 - 續跑只限同一身分：每一步記錄 HEAD、`uv.lock` 與 `configs/` 的 SHA-256、Python、torch 與 transformers 版本；同一個 reproduction 目錄若由不同身分開始，拒絕續跑，必須換新 id，所以一份 comparison 不會混到兩個 commit 的輸出（第 1 條的落實方式）。
 
 **預算規則**：原始 AC6 實驗的 Haiku 花費固定是 US$3.19（完整 run US$3.18 加 smoke），這個數字不再變動。AC1b 的 Haiku 是另一次、明確標記為 reproduction-validation 的執行，有獨立的 US$5 上限，花費寫在 reproduction 目錄自己的紀錄裡；兩者不得混稱為原始實驗成本，也不相加。
+- 每次 AC1b 嘗試各自獨立 US$5 上限（Drew，2026-09-30）。嘗試帳本：`docs/ac1b/attempts.json`（人讀版 `docs/ac1b/README.md`，各次證據在 `docs/ac1b/attempt-<n>/`）。
 
 ## 6. 風險
 
diff --git a/docs/ac1b/README.md b/docs/ac1b/README.md
new file mode 100644
index 0000000..1b7d46e
--- /dev/null
+++ b/docs/ac1b/README.md
@@ -0,0 +1,35 @@
+# AC1b attempts
+
+AC1b (docs/PLAN.md section 5.1) is the whole study rerun from a clean clone with `make setup && make reproduce`. This directory records every attempt, whether it passed or not, so that a failed attempt and its cost are never hidden by a later one.
+
+- `attempts.json`: the ledger, one entry per attempt: commit, reproduction id, start and end (UTC), result and reason, actual Haiku spend. `python -m tinyrouter.ac1b_ledger` checks it and rewrites the budget block below and in the top-level README; `tests/test_ac1b_ledger.py` fails when either is stale.
+- `attempt-<n>/`: the attempt's `comparison.md` as written by `make reproduce`, and `INCIDENT.md`, a note on why it stopped. Evidence files are kept as written: attempt 1's `INCIDENT.md` says it started "about 07:45" UTC; the ledger uses 07:40:02 UTC, the start of the first step in `steps.json`, which is authoritative.
+
+Results (validated by `tinyrouter.ac1b_ledger`):
+
+- `PASS` or `FAIL`: the comparison's verdict under the frozen criteria. `comparison_verdict` always records that mechanical verdict, and the evidence is never rewritten.
+- `INFRASTRUCTURE INTERRUPTED`: allowed only with `failure_kind: infrastructure`, external incident evidence (incident, time, source, error type) and the decision that classified it. INFRASTRUCTURE INTERRUPTED is neither an AC1b failure nor a pass; AC1b acceptance is completed only by an attempt whose result is PASS. Attempt 1 is recorded this way under Drew's decision of 2026-09-29, while its comparison verdict stays FAIL.
+- `INFRASTRUCTURE BLOCKED`: required when the same infrastructure error type recurs on the next attempt. AC1b is then paused, not resumed again.
+- `failure_kind` is one of `infrastructure`, `program defect`, `ac2`; a program defect can only be `FAIL`.
+
+Budget rules (Drew, 2026-09-29 and 2026-09-30):
+
+- The original AC6 experiment cost is fixed at US$3.19. No AC1b attempt changes it, and reproduction-validation spend is never reported as original experiment cost or used to revisit the AC6 conclusion (Haiku spend at most US$5).
+- Every AC1b attempt has its own US$5 cap. `make reproduce` writes the Haiku journal under `reproduction/<id>/results` and runs it with `MAX_USD=5`; the cap counts only that journal, so an attempt at a new commit starts at US$0. Attempt 2 resumed attempt 1 under the same id and identity and so shared its journal and cap (attempt 1 spent US$0).
+- A comparison never mixes the outputs of two commits. After a code fix, AC1b is run again in full at the new commit.
+
+<!-- BEGIN AC1b budget, generated from docs/ac1b/attempts.json; do not edit by hand -->
+- Original experiment (AC6, fixed): US$3.19
+- AC1b attempts (reproduction-validation, each capped at US$5):
+
+| attempt | reproduction id | result | reason | Haiku spend |
+|---|---|---|---|---|
+| 1 | `3992840ddb3f` | INFRASTRUCTURE INTERRUPTED | infrastructure: Haiku count_tokens returned HTTP 503 credential validation failed 5 times during an Anthropic authentication incident; no classification request sent | US$0.000000 |
+| 2 | `3992840ddb3f` | FAIL | program defect: make figures stopped because seeds chose different 8-way aggregations for modernbert/k100 (argmax, summed, summed); Haiku 8600/8600, 0 parse failures, 20 predictions differ | US$3.178751 |
+
+- Reproduction-validation total (all AC1b attempts): US$3.178751
+
+Reproduction-validation spend is never reported as original experiment cost and does not change the AC6 conclusion (Haiku spend at most US$5).
+
+INFRASTRUCTURE INTERRUPTED is neither an AC1b failure nor a pass; AC1b acceptance is completed only by an attempt whose result is PASS.
+<!-- END AC1b budget -->
diff --git a/docs/ac1b/attempt-1/INCIDENT.md b/docs/ac1b/attempt-1/INCIDENT.md
new file mode 100644
index 0000000..461bc4e
--- /dev/null
+++ b/docs/ac1b/attempt-1/INCIDENT.md
@@ -0,0 +1,13 @@
+# AC1b attempt 1: interrupted by an external authentication incident
+
+- Reproduction id: `3992840ddb3f` (clean clone of `main` at `3992840ddb3f`, `uv.lock` SHA-256 prefix `d6b5940a86b5c28c`).
+- Started: 2026-09-29 about 07:45 UTC. Failed: 2026-09-29 about 14:24 UTC.
+- Completed and passed before the failure (all at `3992840ddb3f`): original-haiku, ac2, pilot-lr, pilot-steps, baselines, curve-bert, curve-modernbert, oos-ablation, verify-logits.
+  - AC2 rerun: 96.60 / 96.36 / 96.49 percent (seeds 42 / 43 / 44), all at or above 95.7.
+  - Pilots selected the same values as the original run (lr 5e-5 for both encoders, S_min 400).
+- Failed step: `llm`. The pre-run token count (`messages.count_tokens`) returned HTTP 503 `InternalServerError` with the message "credential validation failed" on all 5 attempts. No classification request was sent.
+- Cost of attempt 1: US$0.00 (token counting is not billed; the reproduction journal holds no records).
+- Cause: an Anthropic incident, "Elevated errors on claude.ai, Claude Code and Claude Cowork", Investigating since 2026-09-29 14:21 UTC, described as failed requests and being asked to sign in again (status.claude.com). The API component was listed as operational, but requests from this machine received 503 during the incident window.
+- Verdict of attempt 1 under the frozen criteria: FAIL (flow did not complete). Drew's decision (2026-09-29): this is an infrastructure availability failure, not a failure of training or of the protocol; resume after the incident with the same reproduction id and identity. If the same authentication failure recurs, pause and mark AC1b infrastructure-blocked.
+- Before resuming (2026-09-30 01:13 UTC): a free token count succeeded; HEAD, working tree, and `uv.lock` unchanged; the reproduction LLM journal was empty, so no prediction is paid for twice. No hyperparameter, threshold, or result file was changed.
+- Evidence kept: `comparison.md`, `comparison.json`, `steps.json`, and `run.log` from attempt 1 (no API key in any of them).
diff --git a/docs/ac1b/attempt-1/comparison.md b/docs/ac1b/attempt-1/comparison.md
new file mode 100644
index 0000000..c77a299
--- /dev/null
+++ b/docs/ac1b/attempt-1/comparison.md
@@ -0,0 +1,53 @@
+# AC1b comparison
+
+**Verdict: FAIL** (0 item(s) REVIEW REQUIRED). Criteria: docs/PLAN.md section 5.1 (frozen 2026-09-29). FAIL only when the flow or AC2 fails.
+
+## Flow
+
+| step | status | commit | completion line(s) |
+|---|---|---|---|
+| original-haiku | PASS | 3992840ddb3f | `verified 1/1 release files` |
+| ac2 | PASS | 3992840ddb3f | `wrote reproduction/3992840ddb3f/results/ac2.json: PASS` |
+| pilot-lr | PASS | 3992840ddb3f | `wrote reproduction/3992840ddb3f/results/pilots/lr.json` |
+| pilot-steps | PASS | 3992840ddb3f | `wrote reproduction/3992840ddb3f/results/pilots/steps.json` |
+| baselines | PASS | 3992840ddb3f | `completed 36/36 baseline points` |
+| curve-bert | PASS | 3992840ddb3f | `completed 18/18 encoder points (bert)` |
+| curve-modernbert | PASS | 3992840ddb3f | `completed 18/18 encoder points (modernbert)` |
+| oos-ablation | PASS | 3992840ddb3f | `completed 3/3 ablation points` |
+| verify-logits | PASS | 3992840ddb3f | `OK 75 archive(s) match reproduction/3992840ddb3f/results/logits-manifest.json` |
+| llm | FAIL | 3992840ddb3f | none |
+
+## AC2 (threshold 95.7%, every seed): PASS
+
+| seed | rerun (%) | original (%) | passed |
+|---|---|---|---|
+| 42 | 96.60 | 96.51 | True |
+| 43 | 96.36 | 96.40 | True |
+| 44 | 96.49 | 96.22 | True |
+
+## Haiku (reproduction-validation run)
+
+```json
+{
+  "reached": false,
+  "checks_passed": false,
+  "problems": [
+    "no reproduction run"
+  ]
+}
+```
+
+## Budget (separate, never summed)
+
+- original AC6 experiment: US$3.19 (fixed)
+- reproduction-validation: US$None (cap US$5)
+
+## Pilots
+
+| pilot | original | rerun | status |
+|---|---|---|---|
+| lr | {'bert': 5e-05, 'modernbert': 5e-05} | {'bert': 5e-05, 'modernbert': 5e-05} | OK |
+| steps | 400 | 400 | OK |
+
+Numbers: not reached (the analysis step did not finish).
+
diff --git a/docs/ac1b/attempt-2/INCIDENT.md b/docs/ac1b/attempt-2/INCIDENT.md
new file mode 100644
index 0000000..7a6b7b2
--- /dev/null
+++ b/docs/ac1b/attempt-2/INCIDENT.md
@@ -0,0 +1,11 @@
+# AC1b attempt 2: stopped by a figure-code defect
+
+Written 2026-09-30 from the attempt's `run.log`, `steps.json` and `comparison.md` (the evidence directory had no incident note for this attempt).
+
+- Reproduction id: `3992840ddb3f`, the same id and identity as attempt 1 (commit `3992840ddb3f`, same `uv.lock`, `configs/`, Python, torch and transformers). It resumed attempt 1: the nine steps that had passed were skipped because their outputs still had the recorded SHA-256.
+- Started (resumed `llm` step): 2026-09-30 01:13:40 UTC. Failed: 2026-09-30 01:31:53 UTC.
+- Passed in this attempt: llm (Haiku 8,600/8,600, 0 parse failures, 20 predictions different from the original run), verify-llm, analysis, bench-cpu, llm-latency, cost.
+- Haiku spend: US$3.178751, under the US$5 cap of this reproduction id (attempt 1 spent US$0 on the same journal).
+- Failed step: `figures`. `figures.selected_aggregation` raised `ValueError: modernbert/k100: seeds chose different aggregations ['argmax', 'summed', 'summed']; pick one first`. In the original run all three ModernBERT k=100 seeds chose argmax on validation, so this path had never run.
+- Verdict under the frozen criteria (docs/PLAN.md 5.1): FAIL (the flow did not complete). This is a defect in the program, not an infrastructure failure.
+- Drew's decision, relayed by the coordinator (2026-09-30): fix the defect, then run AC1b again in full at the new commit as a new attempt with its own US$5 cap. The outputs of this attempt are not combined with figures or a report from another commit.
diff --git a/docs/ac1b/attempt-2/comparison.md b/docs/ac1b/attempt-2/comparison.md
new file mode 100644
index 0000000..d1fe52c
--- /dev/null
+++ b/docs/ac1b/attempt-2/comparison.md
@@ -0,0 +1,344 @@
+# AC1b comparison
+
+**Verdict: FAIL** (192 item(s) REVIEW REQUIRED). Criteria: docs/PLAN.md section 5.1 (frozen 2026-09-29). FAIL only when the flow or AC2 fails.
+
+## Flow
+
+| step | status | commit | completion line(s) |
+|---|---|---|---|
+| original-haiku | PASS (resumed) | 3992840ddb3f | `verified 1/1 release files` |
+| ac2 | PASS (resumed) | 3992840ddb3f | `wrote reproduction/3992840ddb3f/results/ac2.json: PASS` |
+| pilot-lr | PASS (resumed) | 3992840ddb3f | `wrote reproduction/3992840ddb3f/results/pilots/lr.json` |
+| pilot-steps | PASS (resumed) | 3992840ddb3f | `wrote reproduction/3992840ddb3f/results/pilots/steps.json` |
+| baselines | PASS (resumed) | 3992840ddb3f | `completed 36/36 baseline points` |
+| curve-bert | PASS (resumed) | 3992840ddb3f | `completed 18/18 encoder points (bert)` |
+| curve-modernbert | PASS (resumed) | 3992840ddb3f | `completed 18/18 encoder points (modernbert)` |
+| oos-ablation | PASS (resumed) | 3992840ddb3f | `completed 3/3 ablation points` |
+| verify-logits | PASS (resumed) | 3992840ddb3f | `OK 75 archive(s) match reproduction/3992840ddb3f/results/logits-manifest.json` |
+| llm | PASS | 3992840ddb3f | `completed 8600/8600 llm predictions` |
+| verify-llm | PASS | 3992840ddb3f | `completed 8600/8600 llm predictions` |
+| analysis | PASS | 3992840ddb3f | `completed analysis (75/75 archives, 8600/8600 llm rows, 25 groups)` |
+| bench-cpu | PASS | 3992840ddb3f | `wrote reproduction/3992840ddb3f/results/efficiency/cpu_latency.json` |
+| llm-latency | PASS | 3992840ddb3f | `wrote reproduction/3992840ddb3f/results/efficiency/haiku_latency.json` |
+| cost | PASS | 3992840ddb3f | `wrote reproduction/3992840ddb3f/results/cost/cost.json` |
+| figures | FAIL | 3992840ddb3f | none |
+
+## AC2 (threshold 95.7%, every seed): PASS
+
+| seed | rerun (%) | original (%) | passed |
+|---|---|---|---|
+| 42 | 96.60 | 96.51 | True |
+| 43 | 96.36 | 96.40 | True |
+| 44 | 96.49 | 96.22 | True |
+
+## Haiku (reproduction-validation run)
+
+```json
+{
+  "reached": true,
+  "rows": "8600/8600",
+  "parse_failed": 0,
+  "original_parse_failed": 0,
+  "reproduction_cost_usd": 3.178751,
+  "checks_passed": true,
+  "problems": [],
+  "predictions": {
+    "compared_rows": 8600,
+    "different_predictions": 20,
+    "only_in_original": 0,
+    "only_in_reproduction": 0
+  }
+}
+```
+
+## Budget (separate, never summed)
+
+- original AC6 experiment: US$3.19 (fixed)
+- reproduction-validation: US$3.178751 (cap US$5)
+
+## Pilots
+
+| pilot | original | rerun | status |
+|---|---|---|---|
+| lr | {'bert': 5e-05, 'modernbert': 5e-05} | {'bert': 5e-05, 'modernbert': 5e-05} | OK |
+| steps | 400 | 400 | OK |
+
+## README first screen (test)
+
+| metric | original mean (%) | original std (%) | rerun (%) | difference (pp) | status |
+|---|---|---|---|---|---|
+| ModernBERT k=100 small-only 8-way accuracy | 91.87 | 0.13 | 91.42 | -0.45 | REVIEW REQUIRED |
+| Haiku 4.5 8-way accuracy | 82.07 | n/a | 81.96 | -0.11 | REVIEW REQUIRED |
+| ModernBERT k=10 small-only 8-way accuracy | 81.50 | 0.69 | 80.56 | -0.94 | REVIEW REQUIRED |
+| ModernBERT k=10 hybrid 8-way accuracy | 87.95 | 0.39 | 87.60 | -0.35 | OK |
+| ModernBERT k=10 hybrid Haiku call rate | 23.88 | 3.76 | 27.24 | 3.36 | OK |
+| k=100 threshold, test selective risk | 7.34 | 0.86 | 7.75 | 0.40 | OK |
+| k=100 threshold, validation selective risk | 1.50 | 0.13 | 1.57 | 0.07 | OK |
+| validation OOS share | 3.23 | 0.00 | 3.23 | 0.00 | OK |
+| test OOS share | 18.18 | 0.00 | 18.18 | 0.00 | OK |
+| share of the gap explained by OOS share | 86.66 | 1.39 | 87.73 | 1.07 | OK |
+| k=100 kept-OOS error rate, test | 36.48 | 3.04 | 39.25 | 2.77 | OK |
+| k=100 kept-OOS error rate, validation | 19.91 | 2.93 | 22.75 | 2.84 | OK |
+| ModernBERT k=100 hybrid Haiku call rate | 1.27 | 1.59 | 1.27 | 0.00 | OK |
+| ModernBERT k=100 hybrid 8-way accuracy | 92.13 | 0.14 | 91.63 | -0.50 | REVIEW REQUIRED |
+
+## README router table (test, 8-way)
+
+| metric | original mean (%) | original std (%) | rerun (%) | difference (pp) | status |
+|---|---|---|---|---|---|
+| LLM-only accuracy_8 | 82.07 | n/a | 81.96 | -0.11 | REVIEW REQUIRED |
+| LLM-only oos_recall | 56.80 | n/a | 56.60 | -0.20 | REVIEW REQUIRED |
+| LLM-only llm_call_rate | 100.00 | n/a | 100.00 | 0.00 | OK |
+| modernbert/k10 small-only accuracy_8 | 81.50 | 0.69 | 80.56 | -0.94 | REVIEW REQUIRED |
+| modernbert/k10 small-only oos_recall | 23.13 | 3.18 | 20.67 | -2.47 | OK |
+| modernbert/k10 small-only high_conf_oos_misroute_rate | 76.87 | 3.18 | 79.33 | 2.47 | OK |
+| modernbert/k10 small-only llm_call_rate | 0.00 | 0.00 | 0.00 | 0.00 | OK |
+| modernbert/k10 hybrid 0.02 accuracy_8 | 87.95 | 0.39 | 87.60 | -0.35 | OK |
+| modernbert/k10 hybrid 0.02 oos_recall | 53.47 | 1.46 | 53.13 | -0.33 | OK |
+| modernbert/k10 hybrid 0.02 high_conf_oos_misroute_rate | 15.37 | 4.72 | 13.23 | -2.13 | OK |
+| modernbert/k10 hybrid 0.02 llm_call_rate | 23.88 | 3.76 | 27.24 | 3.36 | OK |
+| modernbert/k10 hybrid 0.05 accuracy_8 | 86.05 | 0.59 | 85.75 | -0.30 | OK |
+| modernbert/k10 hybrid 0.05 oos_recall | 43.53 | 3.54 | 43.00 | -0.53 | OK |
+| modernbert/k10 hybrid 0.05 high_conf_oos_misroute_rate | 40.07 | 7.53 | 37.90 | -2.17 | OK |
+| modernbert/k10 hybrid 0.05 llm_call_rate | 11.75 | 3.28 | 13.32 | 1.56 | OK |
+| modernbert/k10 oracle accuracy_8 | 91.45 | 0.22 | 91.10 | -0.36 | REVIEW REQUIRED |
+| modernbert/k10 oracle oos_recall | 61.23 | 0.68 | 60.57 | -0.67 | OK |
+| modernbert/k10 oracle high_conf_oos_misroute_rate | 0.00 | 0.00 | 0.00 | 0.00 | OK |
+| modernbert/k10 oracle llm_call_rate | 18.50 | 0.69 | 19.44 | 0.94 | REVIEW REQUIRED |
+| modernbert/k100 small-only accuracy_8 | 91.87 | 0.13 | 91.42 | -0.45 | REVIEW REQUIRED |
+| modernbert/k100 small-only oos_recall | 61.07 | 0.38 | 58.47 | -2.60 | REVIEW REQUIRED |
+| modernbert/k100 small-only high_conf_oos_misroute_rate | 38.93 | 0.38 | 41.53 | 2.60 | REVIEW REQUIRED |
+| modernbert/k100 small-only llm_call_rate | 0.00 | 0.00 | 0.00 | 0.00 | OK |
+| modernbert/k100 hybrid 0.02 accuracy_8 | 92.13 | 0.14 | 91.63 | -0.50 | REVIEW REQUIRED |
+| modernbert/k100 hybrid 0.02 oos_recall | 62.43 | 1.33 | 59.47 | -2.97 | REVIEW REQUIRED |
+| modernbert/k100 hybrid 0.02 high_conf_oos_misroute_rate | 34.53 | 5.53 | 36.83 | 2.30 | OK |
+| modernbert/k100 hybrid 0.02 llm_call_rate | 1.27 | 1.59 | 1.27 | 0.00 | OK |
+| modernbert/k100 hybrid 0.05 accuracy_8 | 91.93 | 0.13 | 91.57 | -0.36 | REVIEW REQUIRED |
+| modernbert/k100 hybrid 0.05 oos_recall | 61.37 | 0.21 | 59.23 | -2.13 | REVIEW REQUIRED |
+| modernbert/k100 hybrid 0.05 high_conf_oos_misroute_rate | 38.30 | 0.17 | 39.90 | 1.60 | REVIEW REQUIRED |
+| modernbert/k100 hybrid 0.05 llm_call_rate | 0.16 | 0.06 | 0.35 | 0.19 | REVIEW REQUIRED |
+| modernbert/k100 oracle accuracy_8 | 95.15 | 0.05 | 94.83 | -0.32 | REVIEW REQUIRED |
+| modernbert/k100 oracle oos_recall | 75.60 | 0.60 | 74.00 | -1.60 | REVIEW REQUIRED |
+| modernbert/k100 oracle high_conf_oos_misroute_rate | 0.00 | 0.00 | 0.00 | 0.00 | OK |
+| modernbert/k100 oracle llm_call_rate | 8.13 | 0.13 | 8.58 | 0.45 | REVIEW REQUIRED |
+
+## OOS ablation (OOS 250 against OOS 0)
+
+| metric | original mean (%) | original std (%) | rerun (%) | difference (pp) | status |
+|---|---|---|---|---|---|
+| oos_0/hybrid_test/0.02/accuracy_8 | 88.32 | 0.18 | 87.95 | -0.37 | REVIEW REQUIRED |
+| oos_0/hybrid_test/0.02/high_conf_oos_misroute_rate | 37.67 | 2.90 | 38.57 | 0.90 | OK |
+| oos_0/hybrid_test/0.02/llm_call_rate | 12.07 | 0.73 | 11.82 | -0.25 | OK |
+| oos_0/hybrid_test/0.02/oos_recall | 40.60 | 1.04 | 38.90 | -1.70 | REVIEW REQUIRED |
+| oos_0/hybrid_test/0.05/accuracy_8 | 81.03 | 0.07 | 81.00 | -0.03 | OK |
+| oos_0/hybrid_test/0.05/high_conf_oos_misroute_rate | 99.33 | 0.23 | 99.13 | -0.20 | OK |
+| oos_0/hybrid_test/0.05/llm_call_rate | 0.12 | 0.04 | 0.16 | 0.04 | REVIEW REQUIRED |
+| oos_0/hybrid_test/0.05/oos_recall | 0.53 | 0.21 | 0.73 | 0.20 | OK |
+| oos_0/oos_detection_test/auprc | 91.52 | 0.26 | 91.53 | 0.01 | OK |
+| oos_0/oos_detection_test/auroc | 97.80 | 0.04 | 97.84 | 0.04 | REVIEW REQUIRED |
+| oos_0/small_only_oos_recall | 0.00 | 0.00 | 0.00 | 0.00 | OK |
+| oos_250/hybrid_test/0.02/accuracy_8 | 92.13 | 0.14 | 91.63 | -0.50 | REVIEW REQUIRED |
+| oos_250/hybrid_test/0.02/high_conf_oos_misroute_rate | 34.53 | 5.53 | 36.83 | 2.30 | OK |
+| oos_250/hybrid_test/0.02/llm_call_rate | 1.27 | 1.59 | 1.27 | 0.00 | OK |
+| oos_250/hybrid_test/0.02/oos_recall | 62.43 | 1.33 | 59.47 | -2.97 | REVIEW REQUIRED |
+| oos_250/hybrid_test/0.05/accuracy_8 | 91.93 | 0.13 | 91.57 | -0.36 | REVIEW REQUIRED |
+| oos_250/hybrid_test/0.05/high_conf_oos_misroute_rate | 38.30 | 0.17 | 39.90 | 1.60 | REVIEW REQUIRED |
+| oos_250/hybrid_test/0.05/llm_call_rate | 0.16 | 0.06 | 0.35 | 0.19 | REVIEW REQUIRED |
+| oos_250/hybrid_test/0.05/oos_recall | 61.37 | 0.21 | 59.23 | -2.13 | REVIEW REQUIRED |
+| oos_250/oos_detection_test/auprc | 94.34 | 0.25 | 94.13 | -0.21 | OK |
+| oos_250/oos_detection_test/auroc | 98.33 | 0.07 | 98.26 | -0.06 | OK |
+| oos_250/small_only_oos_recall | 61.07 | 0.38 | 58.47 | -2.60 | REVIEW REQUIRED |
+
+## Machine-dependent numbers: 583 LISTED, NOT JUDGED (all rows in the JSON)
+
+Latency, k=100 training time and peak memory, and the cost model vary with the machine; they are listed with their differences and never marked REVIEW REQUIRED.
+
+## Learning curves (test, small model alone)
+
+| metric | original mean (%) | original std (%) | rerun (%) | difference (pp) | status |
+|---|---|---|---|---|---|
+| modernbert k=1 small-only accuracy_8 | 41.71 | 2.47 | 43.30 | 1.59 | OK |
+| modernbert k=1 small-only oos_recall | 1.73 | 0.71 | 2.43 | 0.70 | OK |
+| modernbert k=5 small-only accuracy_8 | 70.99 | 3.40 | 71.71 | 0.72 | OK |
+| modernbert k=5 small-only oos_recall | 11.00 | 4.29 | 12.30 | 1.30 | OK |
+| modernbert k=10 small-only accuracy_8 | 81.50 | 0.69 | 80.56 | -0.94 | REVIEW REQUIRED |
+| modernbert k=10 small-only oos_recall | 23.13 | 3.18 | 20.67 | -2.47 | OK |
+| modernbert k=25 small-only accuracy_8 | 86.27 | 0.52 | 87.21 | 0.94 | REVIEW REQUIRED |
+| modernbert k=25 small-only oos_recall | 36.43 | 3.00 | 40.93 | 4.50 | REVIEW REQUIRED |
+| modernbert k=50 small-only accuracy_8 | 90.39 | 0.38 | 90.25 | -0.14 | OK |
+| modernbert k=50 small-only oos_recall | 53.87 | 2.32 | 53.83 | -0.03 | OK |
+| modernbert k=100 small-only accuracy_8 | 91.87 | 0.13 | 91.42 | -0.45 | REVIEW REQUIRED |
+| modernbert k=100 small-only oos_recall | 61.07 | 0.38 | 58.47 | -2.60 | REVIEW REQUIRED |
+| bert k=1 small-only accuracy_8 | 54.93 | 1.32 | 55.28 | 0.35 | OK |
+| bert k=1 small-only oos_recall | 9.90 | 6.17 | 9.47 | -0.43 | OK |
+| bert k=5 small-only accuracy_8 | 78.99 | 1.61 | 78.79 | -0.20 | OK |
+| bert k=5 small-only oos_recall | 19.10 | 6.95 | 18.00 | -1.10 | OK |
+| bert k=10 small-only accuracy_8 | 81.90 | 1.18 | 82.16 | 0.27 | OK |
+| bert k=10 small-only oos_recall | 24.07 | 4.20 | 25.47 | 1.40 | OK |
+| bert k=25 small-only accuracy_8 | 86.52 | 0.65 | 86.16 | -0.36 | OK |
+| bert k=25 small-only oos_recall | 38.00 | 3.18 | 35.53 | -2.47 | OK |
+| bert k=50 small-only accuracy_8 | 88.85 | 0.30 | 88.86 | 0.01 | OK |
+| bert k=50 small-only oos_recall | 44.83 | 1.55 | 45.17 | 0.33 | OK |
+| bert k=100 small-only accuracy_8 | 91.05 | 0.53 | 90.76 | -0.29 | OK |
+| bert k=100 small-only oos_recall | 56.07 | 2.61 | 54.50 | -1.57 | OK |
+
+## Threshold diagnostics: 155 of 758 flagged (all rows in the JSON)
+
+| metric | original mean (%) | original std (%) | rerun (%) | difference (pp) | status |
+|---|---|---|---|---|---|
+| bert/k1 0.02/sensitivity_reweighted_validation/test/kept_oos_error_rate | 60.00 | n/a | 66.67 | 6.67 | REVIEW REQUIRED |
+| bert/k1 0.02/sensitivity_reweighted_validation/test/kept_oos_share | 3.88 | n/a | 4.11 | 0.23 | REVIEW REQUIRED |
+| bert/k1 0.02/sensitivity_reweighted_validation/test/selective_risk | 2.33 | n/a | 2.74 | 0.41 | REVIEW REQUIRED |
+| bert/k1 0.02/share_of_gap_explained_by_oos_share | 84.49 | n/a | 84.46 | -0.03 | REVIEW REQUIRED |
+| bert/k1 0.02/test/kept_oos_error_rate | 60.00 | n/a | 66.67 | 6.67 | REVIEW REQUIRED |
+| bert/k1 0.02/test/kept_oos_share | 3.88 | n/a | 4.11 | 0.23 | REVIEW REQUIRED |
+| bert/k1 0.02/test/selective_risk | 2.33 | n/a | 2.74 | 0.41 | REVIEW REQUIRED |
+| bert/k1 0.02/test_reweighted_to_validation_oos_share/kept_oos_error_rate | 60.00 | n/a | 66.67 | 6.67 | REVIEW REQUIRED |
+| bert/k1 0.02/test_reweighted_to_validation_oos_share/kept_oos_share | 0.60 | n/a | 0.64 | 0.04 | REVIEW REQUIRED |
+| bert/k1 0.02/test_reweighted_to_validation_oos_share/selective_risk | 0.36 | n/a | 0.43 | 0.07 | REVIEW REQUIRED |
+| bert/k1 0.05/sensitivity_reweighted_validation/test/high_conf_oos_misroute_rate | 0.40 | 0.35 | 0.77 | 0.37 | REVIEW REQUIRED |
+| bert/k1 0.05/sensitivity_reweighted_validation/validation_weighted_coverage | 3.02 | 2.90 | 5.98 | 2.97 | REVIEW REQUIRED |
+| bert/k1 0.05/validation/kept_oos_error_rate | 100.00 | 0.00 | 66.67 | -33.33 | REVIEW REQUIRED |
+| bert/k1 0.05/validation/kept_oos_share | 0.43 | 0.16 | 0.78 | 0.35 | REVIEW REQUIRED |
+| bert/k1 0.05/validation/selective_risk | 2.83 | 0.18 | 2.64 | -0.19 | REVIEW REQUIRED |
+| bert/k10 0.02/sensitivity_reweighted_validation/tau | 96.88 | 2.54 | n/a | n/a | REVIEW REQUIRED |
+| bert/k10 0.02/sensitivity_reweighted_validation/validation_weighted_coverage | 27.00 | 5.36 | 32.65 | 5.65 | REVIEW REQUIRED |
+| bert/k10 0.02/tau | 81.51 | 7.57 | n/a | n/a | REVIEW REQUIRED |
+| bert/k10 0.05/sensitivity_reweighted_validation/test/kept_oos_share | 3.85 | 0.84 | 4.84 | 0.99 | REVIEW REQUIRED |
+| bert/k10 0.05/validation/kept_oos_share | 1.84 | 0.02 | 1.95 | 0.11 | REVIEW REQUIRED |
+| bert/k100 0.02/test/high_conf_oos_misroute_rate | 27.27 | 1.07 | 29.57 | 2.30 | REVIEW REQUIRED |
+| bert/k100 0.02/test/kept_oos_error_rate | 35.48 | 2.24 | 38.06 | 2.58 | REVIEW REQUIRED |
+| bert/k100 0.02/test/selective_risk | 5.98 | 0.10 | 6.43 | 0.45 | REVIEW REQUIRED |
+| bert/k100 0.02/test_reweighted_to_validation_oos_share/high_conf_oos_misroute_rate | 27.27 | 1.07 | 29.57 | 2.30 | REVIEW REQUIRED |
+| bert/k100 0.02/test_reweighted_to_validation_oos_share/kept_oos_error_rate | 35.48 | 2.24 | 38.06 | 2.58 | REVIEW REQUIRED |
+| bert/k100 0.02/test_reweighted_to_validation_oos_share/selective_risk | 1.77 | 0.08 | 1.88 | 0.10 | REVIEW REQUIRED |
+| bert/k100 0.05/sensitivity_reweighted_validation/test/coverage | 93.12 | 0.62 | 93.79 | 0.68 | REVIEW REQUIRED |
+| bert/k100 0.05/sensitivity_reweighted_validation/test/high_conf_oos_misroute_rate | 21.17 | 1.62 | 24.53 | 3.37 | REVIEW REQUIRED |
+| bert/k100 0.05/sensitivity_reweighted_validation/test/kept_oos_error_rate | 31.11 | 2.89 | 34.91 | 3.80 | REVIEW REQUIRED |
+| bert/k100 0.05/sensitivity_reweighted_validation/test/selective_risk | 4.80 | 0.40 | 5.46 | 0.65 | REVIEW REQUIRED |
+| bert/k25 0.02/sensitivity_reweighted_validation/tau | 21.61 | 8.44 | n/a | n/a | REVIEW REQUIRED |
+| bert/k25 0.02/sensitivity_reweighted_validation/test/kept_oos_share | 5.54 | 0.98 | 4.17 | -1.37 | REVIEW REQUIRED |
+| bert/k25 0.02/tau | 11.31 | 2.09 | n/a | n/a | REVIEW REQUIRED |
+| bert/k25 0.02/validation/high_conf_oos_misroute_rate | 12.67 | 2.52 | 17.00 | 4.33 | REVIEW REQUIRED |
+| bert/k25 0.02/validation/in_scope_risk | 1.13 | 0.06 | 0.97 | -0.16 | REVIEW REQUIRED |
+| bert/k25 0.02/validation/kept_oos_error_rate | 27.27 | 2.40 | 32.30 | 5.03 | REVIEW REQUIRED |
+| bert/k25 0.02/validation/kept_oos_share | 1.62 | 0.23 | 1.89 | 0.27 | REVIEW REQUIRED |
+| bert/k25 0.05/sensitivity_reweighted_validation/tau | 10.62 | 1.65 | n/a | n/a | REVIEW REQUIRED |
+| bert/k25 0.05/sensitivity_reweighted_validation/test/coverage | 84.67 | 3.15 | 81.42 | -3.25 | REVIEW REQUIRED |
+| bert/k25 0.05/sensitivity_reweighted_validation/test/in_scope_risk | 1.33 | 0.26 | 0.94 | -0.38 | REVIEW REQUIRED |
+| bert/k25 0.05/sensitivity_reweighted_validation/validation_weighted_coverage | 86.37 | 2.25 | 83.66 | -2.70 | REVIEW REQUIRED |
+| bert/k25 0.05/sensitivity_reweighted_validation/validation_weighted_risk | 4.06 | 0.05 | 4.11 | 0.05 | REVIEW REQUIRED |
+| bert/k25 0.05/tau | 2.32 | 0.17 | n/a | n/a | REVIEW REQUIRED |
+| bert/k25 0.05/test/coverage | 99.96 | 0.03 | 99.59 | -0.37 | REVIEW REQUIRED |
+| bert/k25 0.05/test/kept_oos_share | 18.15 | 0.02 | 17.87 | -0.27 | REVIEW REQUIRED |
+| bert/k25 0.05/test_reweighted_to_validation_oos_share/coverage | 99.99 | 0.00 | 99.90 | -0.10 | REVIEW REQUIRED |
+| bert/k25 0.05/test_reweighted_to_validation_oos_share/kept_oos_share | 3.22 | 0.00 | 3.16 | -0.06 | REVIEW REQUIRED |
+| bert/k25 0.05/validation/coverage | 100.00 | 0.00 | 99.97 | -0.03 | REVIEW REQUIRED |
+| bert/k25 0.05/validation/kept_oos_share | 3.23 | 0.00 | 3.19 | -0.03 | REVIEW REQUIRED |
+| bert/k5 0.02/sensitivity_reweighted_validation/tau | n/a | n/a | 98.40 | n/a | REVIEW REQUIRED |
+| bert/k5 0.02/tau | n/a | n/a | 87.51 | n/a | REVIEW REQUIRED |
+| bert/k5 0.02/test/coverage | 58.16 | 0.86 | 59.78 | 1.61 | REVIEW REQUIRED |
+| bert/k5 0.02/test_reweighted_to_validation_oos_share/coverage | 66.07 | 0.97 | 67.95 | 1.88 | REVIEW REQUIRED |
+| bert/k5 0.02/validation/coverage | 66.81 | 1.41 | 68.41 | 1.60 | REVIEW REQUIRED |
+| bert/k5 0.05/sensitivity_reweighted_validation/validation_weighted_risk | 4.10 | 0.03 | 4.07 | -0.03 | REVIEW REQUIRED |
+| bert/k5 0.05/tau | 60.05 | 1.53 | 57.94 | -2.11 | REVIEW REQUIRED |
+| bert/k5 0.05/test/coverage | 81.46 | 1.09 | 82.74 | 1.28 | REVIEW REQUIRED |
+| bert/k5 0.05/test_reweighted_to_validation_oos_share/coverage | 88.55 | 1.02 | 89.60 | 1.05 | REVIEW REQUIRED |
+| bert/k5 0.05/validation/coverage | 89.11 | 0.47 | 89.76 | 0.66 | REVIEW REQUIRED |
+| bert/k50 0.02/sensitivity_reweighted_validation/tau | n/a | n/a | -114.56 | n/a | REVIEW REQUIRED |
+| bert/k50 0.02/sensitivity_reweighted_validation/validation_weighted_risk | 1.33 | 0.09 | 1.42 | 0.09 | REVIEW REQUIRED |
+| bert/k50 0.02/tau | n/a | n/a | -246.17 | n/a | REVIEW REQUIRED |
+| bert/k50 0.05/test/in_scope_risk | 1.36 | 0.03 | 1.42 | 0.05 | REVIEW REQUIRED |
+| bert/k50 0.05/test_reweighted_to_validation_oos_share/in_scope_risk | 1.36 | 0.03 | 1.42 | 0.05 | REVIEW REQUIRED |
+| modernbert-oos0/k100 0.02/sensitivity_reweighted_validation/tau | -13.82 | 1.04 | n/a | n/a | REVIEW REQUIRED |
+| modernbert-oos0/k100 0.02/sensitivity_reweighted_validation/test/in_scope_risk | 0.18 | 0.06 | 0.32 | 0.14 | REVIEW REQUIRED |
+| modernbert-oos0/k100 0.02/sensitivity_reweighted_validation/test/selective_risk | 2.19 | 0.57 | 2.81 | 0.62 | REVIEW REQUIRED |
+| modernbert-oos0/k100 0.02/share_of_gap_explained_by_oos_share | 93.87 | 1.34 | 92.22 | -1.65 | REVIEW REQUIRED |
+| modernbert-oos0/k100 0.02/tau | -119.38 | 44.08 | n/a | n/a | REVIEW REQUIRED |
+| modernbert-oos0/k100 0.02/test/in_scope_risk | 0.75 | 0.03 | 0.86 | 0.11 | REVIEW REQUIRED |
+| modernbert-oos0/k100 0.02/test_reweighted_to_validation_oos_share/in_scope_risk | 0.75 | 0.03 | 0.86 | 0.11 | REVIEW REQUIRED |
+| modernbert-oos0/k100 0.02/test_reweighted_to_validation_oos_share/selective_risk | 1.99 | 0.12 | 2.13 | 0.14 | REVIEW REQUIRED |
+| modernbert-oos0/k100 0.05/sensitivity_reweighted_validation/tau | n/a | n/a | 50.79 | n/a | REVIEW REQUIRED |
+| modernbert-oos0/k100 0.05/sensitivity_reweighted_validation/test/high_conf_oos_misroute_rate | 24.13 | 1.33 | 22.33 | -1.80 | REVIEW REQUIRED |
+| modernbert-oos0/k100 0.05/sensitivity_reweighted_validation/test/in_scope_risk | 0.50 | 0.04 | 0.56 | 0.06 | REVIEW REQUIRED |
+| modernbert-oos0/k100 0.05/sensitivity_reweighted_validation/test/kept_oos_share | 5.18 | 0.26 | 4.82 | -0.36 | REVIEW REQUIRED |
+| modernbert-oos0/k100 0.05/sensitivity_reweighted_validation/test/selective_risk | 5.65 | 0.30 | 5.35 | -0.30 | REVIEW REQUIRED |
+| modernbert-oos0/k100 0.05/sensitivity_reweighted_validation/validation_weighted_coverage | 83.84 | 0.25 | 83.57 | -0.27 | REVIEW REQUIRED |
+| modernbert-oos0/k100 0.05/share_of_gap_explained_by_oos_share | 97.97 | 0.26 | 97.33 | -0.64 | REVIEW REQUIRED |
+| modernbert-oos0/k100 0.05/tau | n/a | n/a | 10.34 | n/a | REVIEW REQUIRED |
+| modernbert-oos0/k100 0.05/test/coverage | 99.88 | 0.04 | 99.84 | -0.04 | REVIEW REQUIRED |
+| modernbert-oos0/k100 0.05/test/in_scope_risk | 1.08 | 0.05 | 1.16 | 0.07 | REVIEW REQUIRED |
+| modernbert-oos0/k100 0.05/test_reweighted_to_validation_oos_share/coverage | 99.98 | 0.01 | 99.96 | -0.01 | REVIEW REQUIRED |
+| modernbert-oos0/k100 0.05/test_reweighted_to_validation_oos_share/in_scope_risk | 1.08 | 0.05 | 1.16 | 0.07 | REVIEW REQUIRED |
+| modernbert-oos0/k100 0.05/test_reweighted_to_validation_oos_share/selective_risk | 4.25 | 0.05 | 4.32 | 0.07 | REVIEW REQUIRED |
+| modernbert/k1 0.05/sensitivity_reweighted_validation/test/kept_oos_error_rate | 100.00 | 0.00 | 91.67 | -8.33 | REVIEW REQUIRED |
+| modernbert/k1 0.05/test/kept_oos_error_rate | 100.00 | 0.00 | 91.67 | -8.33 | REVIEW REQUIRED |
+| modernbert/k1 0.05/test_reweighted_to_validation_oos_share/kept_oos_error_rate | 100.00 | 0.00 | 91.67 | -8.33 | REVIEW REQUIRED |
+| modernbert/k10 0.02/sensitivity_reweighted_validation/test/coverage | 61.01 | 0.49 | 62.42 | 1.41 | REVIEW REQUIRED |
+| modernbert/k10 0.02/sensitivity_reweighted_validation/test/high_conf_oos_misroute_rate | 3.93 | 0.78 | 5.77 | 1.83 | REVIEW REQUIRED |
+| modernbert/k10 0.02/sensitivity_reweighted_validation/test/in_scope_risk | 0.48 | 0.04 | 0.61 | 0.13 | REVIEW REQUIRED |
+| modernbert/k10 0.02/sensitivity_reweighted_validation/test/kept_oos_error_rate | 54.19 | 0.72 | 61.73 | 7.54 | REVIEW REQUIRED |
+| modernbert/k10 0.02/sensitivity_reweighted_validation/test/kept_oos_share | 2.16 | 0.42 | 2.75 | 0.59 | REVIEW REQUIRED |
+| modernbert/k10 0.02/sensitivity_reweighted_validation/test/selective_risk | 1.64 | 0.21 | 2.26 | 0.62 | REVIEW REQUIRED |
+| modernbert/k10 0.02/sensitivity_reweighted_validation/validation_weighted_coverage | 61.15 | 1.32 | 63.37 | 2.22 | REVIEW REQUIRED |
+| modernbert/k10 0.02/sensitivity_reweighted_validation/validation_weighted_risk | 1.28 | 0.07 | 1.39 | 0.11 | REVIEW REQUIRED |
+| modernbert/k10 0.02/validation/high_conf_oos_misroute_rate | 11.00 | 0.00 | 9.67 | -1.33 | REVIEW REQUIRED |
+| modernbert/k10 0.02/validation/kept_oos_share | 1.03 | 0.12 | 0.87 | -0.16 | REVIEW REQUIRED |
+| modernbert/k10 0.05/validation/selective_risk | 4.31 | 0.01 | 4.33 | 0.01 | REVIEW REQUIRED |
+| modernbert/k100 0.02/sensitivity_reweighted_validation/test/in_scope_risk | 0.53 | 0.10 | 0.39 | -0.14 | REVIEW REQUIRED |
+| modernbert/k100 0.05/sensitivity_reweighted_validation/test/in_scope_risk | 1.23 | 0.08 | 1.11 | -0.11 | REVIEW REQUIRED |
+| modernbert/k100 0.05/sensitivity_reweighted_validation/validation_weighted_risk | 4.12 | 0.06 | 4.05 | -0.07 | REVIEW REQUIRED |
+| modernbert/k100 0.05/test/coverage | 99.84 | 0.06 | 99.65 | -0.19 | REVIEW REQUIRED |
+| modernbert/k100 0.05/test/high_conf_oos_misroute_rate | 38.30 | 0.17 | 39.90 | 1.60 | REVIEW REQUIRED |
+| modernbert/k100 0.05/test/kept_oos_error_rate | 38.57 | 0.12 | 40.63 | 2.06 | REVIEW REQUIRED |
+| modernbert/k100 0.05/test/kept_oos_share | 18.08 | 0.07 | 17.91 | -0.17 | REVIEW REQUIRED |
+| modernbert/k100 0.05/test/selective_risk | 8.01 | 0.08 | 8.30 | 0.29 | REVIEW REQUIRED |
+| modernbert/k100 0.05/test_reweighted_to_validation_oos_share/coverage | 99.94 | 0.01 | 99.93 | -0.02 | REVIEW REQUIRED |
+| modernbert/k100 0.05/test_reweighted_to_validation_oos_share/high_conf_oos_misroute_rate | 38.30 | 0.17 | 39.90 | 1.60 | REVIEW REQUIRED |
+| modernbert/k100 0.05/test_reweighted_to_validation_oos_share/kept_oos_error_rate | 38.57 | 0.12 | 40.63 | 2.06 | REVIEW REQUIRED |
+| modernbert/k100 0.05/test_reweighted_to_validation_oos_share/kept_oos_share | 3.21 | 0.02 | 3.17 | -0.04 | REVIEW REQUIRED |
+| modernbert/k25 0.02/sensitivity_reweighted_validation/test/high_conf_oos_misroute_rate | 6.20 | 0.26 | 6.57 | 0.37 | REVIEW REQUIRED |
+| modernbert/k25 0.02/sensitivity_reweighted_validation/test/kept_oos_error_rate | 42.56 | 0.95 | 35.31 | -7.25 | REVIEW REQUIRED |
+| modernbert/k25 0.02/sensitivity_reweighted_validation/test/kept_oos_share | 3.57 | 0.15 | 4.56 | 0.99 | REVIEW REQUIRED |
+| modernbert/k25 0.02/test/kept_oos_error_rate | 49.52 | 4.21 | 44.33 | -5.19 | REVIEW REQUIRED |
+| modernbert/k25 0.02/test/kept_oos_share | 8.62 | 0.30 | 9.54 | 0.92 | REVIEW REQUIRED |
+| modernbert/k25 0.02/test_reweighted_to_validation_oos_share/kept_oos_error_rate | 49.52 | 4.21 | 44.33 | -5.19 | REVIEW REQUIRED |
+| modernbert/k25 0.02/test_reweighted_to_validation_oos_share/kept_oos_share | 1.40 | 0.05 | 1.56 | 0.16 | REVIEW REQUIRED |
+| modernbert/k25 0.02/validation/high_conf_oos_misroute_rate | 14.33 | 1.53 | 11.00 | -3.33 | REVIEW REQUIRED |
+| modernbert/k25 0.02/validation/in_scope_risk | 1.09 | 0.05 | 1.20 | 0.11 | REVIEW REQUIRED |
+| modernbert/k25 0.02/validation/kept_oos_error_rate | 32.81 | 1.66 | 24.00 | -8.81 | REVIEW REQUIRED |
+| modernbert/k25 0.02/validation/selective_risk | 1.56 | 0.00 | 1.56 | -0.01 | REVIEW REQUIRED |
+| modernbert/k25 0.05/sensitivity_reweighted_validation/test/coverage | 85.41 | 0.99 | 87.78 | 2.38 | REVIEW REQUIRED |
+| modernbert/k25 0.05/sensitivity_reweighted_validation/test/high_conf_oos_misroute_rate | 20.10 | 0.62 | 23.17 | 3.07 | REVIEW REQUIRED |
+| modernbert/k25 0.05/sensitivity_reweighted_validation/test/kept_oos_share | 8.70 | 0.54 | 10.34 | 1.65 | REVIEW REQUIRED |
+| modernbert/k25 0.05/sensitivity_reweighted_validation/test/selective_risk | 5.20 | 0.57 | 5.85 | 0.65 | REVIEW REQUIRED |
+| modernbert/k25 0.05/sensitivity_reweighted_validation/validation_weighted_coverage | 85.95 | 1.42 | 88.49 | 2.55 | REVIEW REQUIRED |
+| modernbert/k25 0.05/sensitivity_reweighted_validation/validation_weighted_risk | 4.09 | 0.02 | 4.05 | -0.03 | REVIEW REQUIRED |
+| modernbert/k25 0.05/test/kept_oos_error_rate | 63.18 | 3.12 | 59.02 | -4.16 | REVIEW REQUIRED |
+| modernbert/k25 0.05/test/selective_risk | 13.51 | 0.46 | 12.74 | -0.77 | REVIEW REQUIRED |
+| modernbert/k25 0.05/test_reweighted_to_validation_oos_share/kept_oos_error_rate | 63.18 | 3.12 | 59.02 | -4.16 | REVIEW REQUIRED |
+| modernbert/k25 0.05/validation/high_conf_oos_misroute_rate | 51.33 | 3.79 | 45.67 | -5.67 | REVIEW REQUIRED |
+| modernbert/k25 0.05/validation/kept_oos_error_rate | 52.28 | 4.96 | 45.67 | -6.61 | REVIEW REQUIRED |
+| modernbert/k5 0.02/test/high_conf_oos_misroute_rate | 4.00 | 0.69 | 4.97 | 0.97 | REVIEW REQUIRED |
+| modernbert/k5 0.02/test/selective_risk | 3.31 | 0.22 | 3.55 | 0.24 | REVIEW REQUIRED |
+| modernbert/k5 0.02/test_reweighted_to_validation_oos_share/high_conf_oos_misroute_rate | 4.00 | 0.69 | 4.97 | 0.97 | REVIEW REQUIRED |
+| modernbert/k5 0.02/validation/high_conf_oos_misroute_rate | 0.67 | 0.58 | 2.33 | 1.67 | REVIEW REQUIRED |
+| modernbert/k5 0.02/validation/in_scope_risk | 1.33 | 0.02 | 1.26 | -0.07 | REVIEW REQUIRED |
+| modernbert/k5 0.02/validation/kept_oos_error_rate | 8.33 | 7.22 | 15.87 | 7.54 | REVIEW REQUIRED |
+| modernbert/k5 0.02/validation/kept_oos_share | 0.45 | 0.12 | 0.58 | 0.13 | REVIEW REQUIRED |
+| modernbert/k5 0.05/sensitivity_reweighted_validation/tau | n/a | n/a | 202.52 | n/a | REVIEW REQUIRED |
+| modernbert/k5 0.05/tau | n/a | n/a | 144.06 | n/a | REVIEW REQUIRED |
+| modernbert/k5 0.05/validation/in_scope_risk | 3.68 | 0.04 | 3.60 | -0.08 | REVIEW REQUIRED |
+| modernbert/k50 0.02/sensitivity_reweighted_validation/tau | 79.38 | 11.04 | n/a | n/a | REVIEW REQUIRED |
+| modernbert/k50 0.02/sensitivity_reweighted_validation/test/in_scope_risk | 0.55 | 0.16 | 0.76 | 0.21 | REVIEW REQUIRED |
+| modernbert/k50 0.02/sensitivity_reweighted_validation/validation_weighted_risk | 1.35 | 0.02 | 1.42 | 0.07 | REVIEW REQUIRED |
+| modernbert/k50 0.02/tau | 35.06 | 7.34 | n/a | n/a | REVIEW REQUIRED |
+| modernbert/k50 0.05/sensitivity_reweighted_validation/tau | n/a | n/a | -238.19 | n/a | REVIEW REQUIRED |
+| modernbert/k50 0.05/sensitivity_reweighted_validation/test/in_scope_risk | 1.17 | 0.10 | 1.35 | 0.18 | REVIEW REQUIRED |
+| modernbert/k50 0.05/sensitivity_reweighted_validation/validation_weighted_risk | 4.11 | 0.04 | 4.16 | 0.04 | REVIEW REQUIRED |
+| modernbert/k50 0.05/share_of_gap_explained_by_oos_share | 89.41 | 1.38 | 88.02 | -1.39 | REVIEW REQUIRED |
+| modernbert/k50 0.05/tau | n/a | n/a | -371.87 | n/a | REVIEW REQUIRED |
+| modernbert/k50 0.05/test/in_scope_risk | 1.48 | 0.10 | 1.65 | 0.17 | REVIEW REQUIRED |
+| modernbert/k50 0.05/test_reweighted_to_validation_oos_share/in_scope_risk | 1.48 | 0.10 | 1.65 | 0.17 | REVIEW REQUIRED |
+| modernbert/k50 0.05/test_reweighted_to_validation_oos_share/selective_risk | 2.91 | 0.07 | 3.08 | 0.17 | REVIEW REQUIRED |
+
diff --git a/docs/ac1b/attempts.json b/docs/ac1b/attempts.json
new file mode 100644
index 0000000..cb27c97
--- /dev/null
+++ b/docs/ac1b/attempts.json
@@ -0,0 +1,47 @@
+{
+  "format_version": 1,
+  "original_ac6_usd": 3.19,
+  "cap_usd_per_attempt": 5.0,
+  "cap_rule": "each AC1b attempt has its own US$5 cap (Drew, 2026-09-30); enforced per reproduction id, whose Haiku journal starts empty",
+  "attempts": [
+    {
+      "attempt": 1,
+      "commit": "3992840ddb3ff1df8b879ba00aaee6277e3f7ce4",
+      "reproduction_id": "3992840ddb3f",
+      "started_utc": "2026-09-29T07:40:02Z",
+      "finished_utc": "2026-09-29T14:24:22Z",
+      "time_source": "steps.json: start of the first step, end of the failed llm step",
+      "result": "INFRASTRUCTURE INTERRUPTED",
+      "failure_kind": "infrastructure",
+      "comparison_verdict": "FAIL",
+      "reason": "infrastructure: Haiku count_tokens returned HTTP 503 credential validation failed 5 times during an Anthropic authentication incident; no classification request sent",
+      "evidence": {
+        "incident": "Elevated errors on claude.ai, Claude Code and Claude Cowork",
+        "time_utc": "2026-09-29 14:21 UTC, status Investigating",
+        "source": "status.claude.com",
+        "error_type": "HTTP 503 credential validation failed (messages.count_tokens)"
+      },
+      "decision": {
+        "text": "503 是外部服務／認證可用性事故，不代表訓練或實驗協定失效……若同一認證問題反覆發生，才暫停並把 AC1b 標為 infrastructure-blocked；目前不應視為 AC1b 失敗。",
+        "source": "Drew's decision, relayed by the coordinator (2026-09-29)"
+      },
+      "haiku_usd": 0.0,
+      "evidence_dir": "docs/ac1b/attempt-1/"
+    },
+    {
+      "attempt": 2,
+      "commit": "3992840ddb3ff1df8b879ba00aaee6277e3f7ce4",
+      "reproduction_id": "3992840ddb3f",
+      "started_utc": "2026-09-30T01:13:40Z",
+      "finished_utc": "2026-09-30T01:31:53Z",
+      "time_source": "steps.json: start of the resumed llm step, end of the failed figures step",
+      "result": "FAIL",
+      "failure_kind": "program defect",
+      "comparison_verdict": "FAIL",
+      "reason": "program defect: make figures stopped because seeds chose different 8-way aggregations for modernbert/k100 (argmax, summed, summed); Haiku 8600/8600, 0 parse failures, 20 predictions differ",
+      "haiku_usd": 3.178751,
+      "evidence_dir": "docs/ac1b/attempt-2/"
+    }
+  ],
+  "result_rule": "INFRASTRUCTURE INTERRUPTED is neither an AC1b failure nor a pass; AC1b acceptance is completed only by an attempt whose result is PASS. comparison_verdict keeps the comparison's own mechanical verdict unchanged."
+}
diff --git a/src/tinyrouter/ac1b_ledger.py b/src/tinyrouter/ac1b_ledger.py
new file mode 100644
index 0000000..1829a6a
--- /dev/null
+++ b/src/tinyrouter/ac1b_ledger.py
@@ -0,0 +1,260 @@
+"""The AC1b attempt ledger (``docs/ac1b/attempts.json``) and the three-part budget.
+
+docs/PLAN.md 5.1: the original AC6 experiment cost is fixed at US$3.19;
+every AC1b attempt is a reproduction-validation run with its own US$5
+cap; their spend is reported per attempt and as a reproduction-validation
+total, and never added to, or reported as, the original experiment cost.
+
+The cap is enforced per reproduction id: ``make reproduce`` writes the
+Haiku journal under ``reproduction/<id>/results`` and runs it with
+``MAX_USD=5``, and ``llm_run`` counts only that journal. A new commit gets
+a new id, so a new attempt starts from US$0. An attempt that resumes the
+same id under the same identity (attempt 2 resumed attempt 1) shares that
+id's journal and therefore its cap.
+"""
+
+from __future__ import annotations
+
+import json
+from pathlib import Path
+
+LEDGER = Path("docs/ac1b/attempts.json")
+REPO = Path(__file__).resolve().parents[2]
+ORIGINAL_AC6_USD = 3.19
+CAP_USD_PER_ATTEMPT = 5.0
+PASS = "PASS"
+FAIL = "FAIL"
+INTERRUPTED = "INFRASTRUCTURE INTERRUPTED"
+BLOCKED = "INFRASTRUCTURE BLOCKED"
+RESULTS = (PASS, FAIL, INTERRUPTED, BLOCKED)
+INFRASTRUCTURE = "infrastructure"
+FAILURE_KINDS = (INFRASTRUCTURE, "program defect", "ac2")
+VERDICTS = (PASS, FAIL)
+EVIDENCE_FIELDS = ("incident", "time_utc", "source", "error_type")
+DECISION_FIELDS = ("text", "source")
+FIELDS = (
+    "attempt",
+    "commit",
+    "reproduction_id",
+    "started_utc",
+    "finished_utc",
+    "result",
+    "failure_kind",
+    "comparison_verdict",
+    "reason",
+    "haiku_usd",
+)
+FINAL_RULE = (
+    "INFRASTRUCTURE INTERRUPTED is neither an AC1b failure nor a pass; AC1b acceptance is "
+    "completed only by an attempt whose result is PASS."
+)
+
+
+class LedgerError(ValueError):
+    """The ledger is malformed or breaks a budget rule."""
+
+
+def validate(ledger: dict) -> dict:
+    """Numbered attempts, known results and kinds, spend within the per-attempt cap.
+
+    An infrastructure result (INTERRUPTED or BLOCKED) needs failure_kind
+    ``infrastructure``, external incident evidence and, for INTERRUPTED, the
+    decision that classified it. A second consecutive infrastructure
+    attempt with the same error type must be BLOCKED: AC1b pauses instead
+    of resuming again. The comparison's own verdict is kept as recorded.
+    """
+    if ledger.get("original_ac6_usd") != ORIGINAL_AC6_USD:
+        raise LedgerError(f"original_ac6_usd must stay {ORIGINAL_AC6_USD}")
+    if ledger.get("cap_usd_per_attempt") != CAP_USD_PER_ATTEMPT:
+        raise LedgerError(f"cap_usd_per_attempt must be {CAP_USD_PER_ATTEMPT}")
+    attempts = ledger.get("attempts")
+    if not isinstance(attempts, list) or not attempts:
+        raise LedgerError("no attempts recorded")
+    previous: dict | None = None
+    for number, entry in enumerate(attempts, start=1):
+        check_attempt(number, entry)
+        check_repeat(number, entry, previous)
+        previous = entry
+    return ledger
+
+
+def check_attempt(number: int, entry: dict) -> None:
+    missing = [f for f in FIELDS if f not in entry]
+    if missing:
+        raise LedgerError(f"attempt {number}: missing {missing}")
+    if entry["attempt"] != number:
+        raise LedgerError(f"attempt numbers must run 1..n; found {entry['attempt']}")
+    result, kind = entry["result"], entry["failure_kind"]
+    if result not in RESULTS:
+        raise LedgerError(f"attempt {number}: result {result!r} not in {RESULTS}")
+    if entry["comparison_verdict"] not in VERDICTS:
+        raise LedgerError(f"attempt {number}: comparison_verdict must be one of {VERDICTS}")
+    if result == PASS and (kind is not None or entry["comparison_verdict"] != PASS):
+        raise LedgerError(f"attempt {number}: PASS needs a PASS comparison and no failure_kind")
+    if result != PASS and kind not in FAILURE_KINDS:
+        raise LedgerError(f"attempt {number}: failure_kind {kind!r} not in {FAILURE_KINDS}")
+    if result in (INTERRUPTED, BLOCKED):
+        check_infrastructure(number, entry)
+    if not str(entry["reason"]).strip():
+        raise LedgerError(f"attempt {number}: a result needs a reason")
+    spent = entry["haiku_usd"]
+    if not isinstance(spent, int | float) or not 0 <= spent <= CAP_USD_PER_ATTEMPT:
+        raise LedgerError(f"attempt {number}: haiku_usd {spent!r} outside 0..cap")
+
+
+def check_infrastructure(number: int, entry: dict) -> None:
+    if entry["failure_kind"] != INFRASTRUCTURE:
+        raise LedgerError(f"attempt {number}: {entry['result']} is for infrastructure only")
+    if not filled(entry.get("evidence"), EVIDENCE_FIELDS):
+        raise LedgerError(f"attempt {number}: {entry['result']} needs evidence {EVIDENCE_FIELDS}")
+    if entry["result"] == INTERRUPTED and not filled(entry.get("decision"), DECISION_FIELDS):
+        raise LedgerError(f"attempt {number}: INTERRUPTED needs the decision {DECISION_FIELDS}")
+
+
+def check_repeat(number: int, entry: dict, previous: dict | None) -> None:
+    """The same infrastructure error twice in a row pauses AC1b (BLOCKED)."""
+    repeated = (
+        previous is not None
+        and previous["failure_kind"] == INFRASTRUCTURE == entry["failure_kind"]
+        and error_type(previous) is not None
+        and error_type(previous) == error_type(entry)
+    )
+    if repeated and entry["result"] != BLOCKED:
+        raise LedgerError(
+            f"attempt {number}: the infrastructure error of attempt {number - 1} recurred; "
+            f"the result must be {BLOCKED} (AC1b paused)"
+        )
+
+
+def error_type(entry: dict) -> str | None:
+    evidence = entry.get("evidence")
+    return evidence.get("error_type") if isinstance(evidence, dict) else None
+
+
+def filled(value: object, fields: tuple[str, ...]) -> bool:
+    return isinstance(value, dict) and all(str(value.get(f) or "").strip() for f in fields)
+
+
+def load(path: Path = REPO / LEDGER) -> dict:
+    return validate(json.loads(path.read_text(encoding="utf-8")))
+
+
+def total_usd(attempts: list[dict]) -> float:
+    return round(sum(float(a["haiku_usd"]) for a in attempts), 6)
+
+
+def budget(ledger: dict, current: dict | None = None) -> dict[str, object]:
+    """Original (fixed), each attempt, and the reproduction-validation total.
+
+    ``current`` is the run being compared (``reproduction_id``, ``usd``)
+    when its id has no ledger entry yet; it is listed as pending and
+    counted in the total. When the id is already in the ledger, the
+    ledger's figures are used and ``current`` adds nothing.
+    """
+    attempts = [dict(a) for a in ledger["attempts"]]
+    recorded_ids = {a["reproduction_id"] for a in attempts}
+    if current is not None and current["reproduction_id"] not in recorded_ids:
+        attempts.append(
+            {
+                "attempt": len(attempts) + 1,
+                "commit": current.get("commit"),
+                "reproduction_id": current["reproduction_id"],
+                "result": "PENDING",
+                "reason": "this comparison; add it to docs/ac1b/attempts.json",
+                "haiku_usd": current.get("usd") or 0.0,
+            }
+        )
+    return {
+        "original_ac6_experiment": {
+            "usd": ORIGINAL_AC6_USD,
+            "note": "fixed: full run US$3.18 plus smoke; no AC1b attempt changes it",
+        },
+        "ac1b_attempts": [
+            {k: a.get(k) for k in ("attempt", "reproduction_id", "result", "reason", "haiku_usd")}
+            for a in attempts
+        ],
+        "reproduction_validation_total": {
+            "usd": total_usd(attempts),
+            "note": "sum over AC1b attempts only; never added to the original experiment cost",
+        },
+        "cap_usd_per_attempt": CAP_USD_PER_ATTEMPT,
+        "final_rule": FINAL_RULE,
+        "rule": (
+            "reproduction-validation spend is never reported as original experiment cost and "
+            "does not change the AC6 US$5 conclusion"
+        ),
+    }
+
+
+def markdown(section: dict) -> list[str]:
+    """The three-part budget as markdown lines."""
+    lines = [
+        f"- Original experiment (AC6, fixed): US${section['original_ac6_experiment']['usd']:.2f}",
+        f"- AC1b attempts (reproduction-validation, each capped at "
+        f"US${section['cap_usd_per_attempt']:g}):",
+        "",
+        "| attempt | reproduction id | result | reason | Haiku spend |",
+        "|---|---|---|---|---|",
+    ]
+    for a in section["ac1b_attempts"]:
+        lines.append(
+            f"| {a['attempt']} | `{a['reproduction_id']}` | {a['result']} | {a['reason']} | "
+            f"US${a['haiku_usd']:.6f} |"
+        )
+    total = section["reproduction_validation_total"]["usd"]
+    return [
+        *lines,
+        "",
+        f"- Reproduction-validation total (all AC1b attempts): US${total:.6f}",
+        "",
+        "Reproduction-validation spend is never reported as original experiment cost and does "
+        "not change the AC6 conclusion (Haiku spend at most US$5).",
+        "",
+        FINAL_RULE,
+    ]
+
+
+BEGIN = "<!-- BEGIN AC1b budget, generated from docs/ac1b/attempts.json; do not edit by hand -->"
+END = "<!-- END AC1b budget -->"
+BUDGET_DOCS = (Path("README.md"), Path("docs/ac1b/README.md"))
+
+
+def block(ledger: dict) -> str:
+    return "\n".join([BEGIN, *markdown(budget(ledger)), END])
+
+
+def with_block(text: str, ledger: dict) -> str:
+    start, end = text.find(BEGIN), text.find(END)
+    if start < 0 or end < start:
+        raise LedgerError("no AC1b budget block (BEGIN ... END markers)")
+    return text[:start] + block(ledger) + text[end + len(END) :]
+
+
+def main(argv: list[str] | None = None) -> None:
+    """Rewrite the budget block in README.md and docs/ac1b/README.md; ``--check`` only compares."""
+    import argparse
+    import sys
+
+    parser = argparse.ArgumentParser(description=main.__doc__)
+    parser.add_argument("--check", action="store_true")
+    args = parser.parse_args(argv)
+    ledger = load(REPO / LEDGER)
+    stale = []
+    for doc in BUDGET_DOCS:
+        path = REPO / doc
+        text = path.read_text(encoding="utf-8")
+        new = with_block(text, ledger)
+        if new == text:
+            continue
+        if args.check:
+            stale.append(str(doc))
+        else:
+            path.write_text(new, encoding="utf-8")
+            print(f"wrote {doc}")
+    if stale:
+        print(f"stale, run `python -m tinyrouter.ac1b_ledger`: {', '.join(stale)}", file=sys.stderr)
+        sys.exit(1)
+
+
+if __name__ == "__main__":
+    main()
diff --git a/src/tinyrouter/comparison.py b/src/tinyrouter/comparison.py
index 3a46e47..323e739 100644
--- a/src/tinyrouter/comparison.py
+++ b/src/tinyrouter/comparison.py
@@ -38,6 +38,7 @@
 from collections.abc import Iterator
 from pathlib import Path
 
+from tinyrouter import ac1b_ledger
 from tinyrouter.ac2 import SEEDS as AC2_SEEDS
 from tinyrouter.ac2 import THRESHOLD as AC2_THRESHOLD
 
@@ -48,9 +49,9 @@
 FAIL = "FAIL"
 EXACT_TOLERANCE = 1e-9
 EXPECTED_LLM_ROWS = 8600
-# docs/PLAN.md 5.1: the original AC6 spend is fixed; the rerun has its own cap.
-ORIGINAL_AC6_USD = 3.19
-REPRODUCTION_CAP_USD = 5.0
+# docs/PLAN.md 5.1: the original AC6 spend is fixed; each AC1b attempt has its own cap.
+ORIGINAL_AC6_USD = ac1b_ledger.ORIGINAL_AC6_USD
+REPRODUCTION_CAP_USD = ac1b_ledger.CAP_USD_PER_ATTEMPT
 TARGET = "0.02"
 CURVE_MODELS = ("modernbert", "bert")
 CURVE_KS = (1, 5, 10, 25, 50, 100)
@@ -223,6 +224,55 @@ def k100_training(root: Path) -> dict | None:
     return out
 
 
+SEEDS = (42, 43, 44)
+
+
+def seed_aggregations(summary: dict, group: str) -> list[str] | None:
+    """Each seed's validation-chosen 8-way aggregation, in seed order; None when absent.
+
+    ``summary.json`` stores one string when the seeds agree, the per-seed list otherwise.
+    """
+    chosen = at(summary, ("groups", group, "result", "selected_aggregation"))
+    if isinstance(chosen, str):
+        return [chosen] * len(SEEDS)
+    return list(chosen) if isinstance(chosen, list) else None
+
+
+def aggregation_rows(before: dict, after: dict) -> list[dict]:
+    """Per group with a final router: original and rerun aggregation per seed. Never judged.
+
+    Fixed before AC1b attempt 3 (docs/PLAN.md 5.1). It explains router
+    differences (a seed that switched aggregation routes differently) and
+    is ``LISTED, NOT JUDGED``: it adds no REVIEW REQUIRED and never changes
+    the verdict.
+    """
+    groups = set(before.get("groups", {})) | set(after.get("groups", {}))
+    rows = []
+    for group in sorted(g for g in groups if has_final(before, g) or has_final(after, g)):
+        original, rerun = seed_aggregations(before, group), seed_aggregations(after, group)
+        differing = (
+            list(SEEDS)
+            if original is None or rerun is None
+            else [s for s, a, b in zip(SEEDS, original, rerun, strict=True) if a != b]
+        )
+        rows.append(
+            {
+                "group": group,
+                "seeds": list(SEEDS),
+                "original": original,
+                "reproduced": rerun,
+                "same": not differing,
+                "differing_seeds": differing,
+                "status": LISTED,
+            }
+        )
+    return rows
+
+
+def has_final(summary: dict, group: str) -> bool:
+    return at(summary, ("groups", group, "result", "final")) is not None
+
+
 def machine_dependent_rows(original_root: Path, reproduced_root: Path) -> list[dict]:
     """Listed, not judged: latency, k=100 training time and peak memory, the cost model."""
     rows = []
@@ -331,23 +381,21 @@ def haiku_section(
     return body
 
 
-def budget_section(reproduced_summary: dict | None) -> dict[str, object]:
-    """Two separate budgets; the rerun's spend is never added to the original experiment's."""
+def budget_section(
+    reproduced_summary: dict | None, ledger: dict, reproduction_id: str | None = None
+) -> dict[str, object]:
+    """Three parts: the fixed original AC6 cost, each AC1b attempt, and their total.
+
+    The rerun's spend is never added to the original experiment's
+    (``ac1b_ledger.budget``).
+    """
     spent = None
     if reproduced_summary is not None:
         spent = round(float(reproduced_summary.get("totals", {}).get("cost_usd", 0.0)), 6)
-    return {
-        "original_ac6_experiment": {
-            "usd": ORIGINAL_AC6_USD,
-            "note": "fixed: full run US$3.18 plus smoke; not changed by any rerun",
-        },
-        "reproduction_validation": {
-            "usd": spent,
-            "cap_usd": REPRODUCTION_CAP_USD,
-            "note": "AC1b rerun only, its own journal and cap; not an experiment cost",
-        },
-        "rule": "separate budgets; never summed or reported as one number",
-    }
+    current = None
+    if reproduction_id is not None:
+        current = {"reproduction_id": reproduction_id, "usd": spent}
+    return {"this_run_usd": spent, **ac1b_ledger.budget(ledger, current)}
 
 
 def ac2_section(reproduced_ac2: dict | None, original_ac2: dict) -> dict[str, object]:
@@ -426,8 +474,14 @@ def build(
     steps: list[dict],
     expected_steps: list[str],
     context: dict[str, object],
+    ledger: dict | None = None,
 ) -> dict[str, object]:
-    """The whole comparison for one rerun; reads JSON only."""
+    """The whole comparison for one rerun; reads JSON only.
+
+    ``ledger`` defaults to this checkout's ``docs/ac1b/attempts.json``.
+    """
+    if ledger is None:
+        ledger = ac1b_ledger.load()
     before = read_json(original_root / "analysis" / "summary.json") or {}
     after = read_json(reproduced_root / "analysis" / "summary.json")
     rep_jsonl = reproduced_root / "llm" / "haiku-8way.jsonl"
@@ -452,6 +506,7 @@ def build(
         "ablation": None if after is None else ablation_rows(before, after),
         "threshold_diagnostics": None if after is None else diagnostic_rows(before, after),
         "machine_dependent": machine_dependent_rows(original_root, reproduced_root),
+        "aggregations": None if after is None else aggregation_rows(before, after),
     }
     flow_ok = flow_passed(steps, expected_steps)
     return {
@@ -467,7 +522,9 @@ def build(
         },
         "ac2": ac2,
         "haiku": haiku,
-        "budget": budget_section(rep_summary),
+        "budget": budget_section(
+            rep_summary, ledger, str(context.get("reproduction_id") or "") or None
+        ),
         **sections,
         **context,
     }
@@ -532,11 +589,12 @@ def render_markdown(body: dict) -> str:
         json.dumps(haiku, indent=2),
         "```",
         "",
-        "## Budget (separate, never summed)",
+        "## Budget",
         "",
-        f"- original AC6 experiment: US${budget['original_ac6_experiment']['usd']:.2f} (fixed)",
-        f"- reproduction-validation: US${budget['reproduction_validation']['usd']} "
-        f"(cap US${budget['reproduction_validation']['cap_usd']:g})",
+        f"This run's Haiku spend: US${budget['this_run_usd']} "
+        f"(cap US${budget['cap_usd_per_attempt']:g} for this reproduction id).",
+        "",
+        *ac1b_ledger.markdown(budget),
         "",
     ]
     lines += pilot_and_number_lines(body)
@@ -572,6 +630,7 @@ def pilot_and_number_lines(body: dict) -> list[str]:
         "Latency, k=100 training time and peak memory, and the cost model vary with the "
         "machine; they are listed with their differences and never marked REVIEW REQUIRED.",
         "",
+        *aggregation_lines(body["aggregations"]),
         "## Learning curves (test, small model alone)",
         "",
         *stat_table(body["learning_curves"]),
@@ -581,3 +640,24 @@ def pilot_and_number_lines(body: dict) -> list[str]:
         *stat_table(diag, only_review=True),
         "",
     ]
+
+
+def aggregation_lines(rows: list[dict]) -> list[str]:
+    differ = sum(not row["same"] for row in rows)
+    lines = [
+        f"## 8-way aggregation chosen on validation: {len(rows)} groups {LISTED} ({differ} differ)",
+        "",
+        "Each seed's final router uses the aggregation it chose on validation; a seed that "
+        "switched routes differently, which can explain a REVIEW REQUIRED row above.",
+        "",
+        "| group | original (42/43/44) | rerun (42/43/44) | same | differing seeds | status |",
+        "|---|---|---|---|---|---|",
+    ]
+    for row in rows:
+        before = "/".join(row["original"]) if row["original"] else "n/a"
+        after = "/".join(row["reproduced"]) if row["reproduced"] else "n/a"
+        seeds = ", ".join(map(str, row["differing_seeds"])) or "none"
+        lines.append(
+            f"| {row['group']} | {before} | {after} | {row['same']} | {seeds} | {row['status']} |"
+        )
+    return [*lines, ""]
diff --git a/src/tinyrouter/figures.py b/src/tinyrouter/figures.py
index 9ccd76c..5afc0e6 100644
--- a/src/tinyrouter/figures.py
+++ b/src/tinyrouter/figures.py
@@ -7,6 +7,10 @@
 palette, which stays distinguishable under the common colour-vision
 deficiencies; line styles and markers differ too, so the figures also work
 in greyscale. Bands and error bars are mean ± sample std over seeds 42, 43, 44.
+
+Each seed's final router uses the 8-way aggregation it chose on validation.
+In risk_coverage a k whose seeds chose differently gets one panel per
+aggregation, titled with the seeds that use it; nothing is chosen on test.
 """
 
 from __future__ import annotations
@@ -131,20 +135,67 @@ def learning_curves(summary: dict, plt) -> object:  # noqa: ANN001
     return fig
 
 
-def selected_aggregation(summary: dict, group: str) -> str:
+AGGREGATIONS = ("argmax", "summed")
+SEEDS = (42, 43, 44)
+
+
+def seed_aggregations(summary: dict, group: str) -> tuple[str, ...]:
+    """The 8-way aggregation each seed's final router chose on validation, in seed order.
+
+    ``summary.json`` stores one string when the three seeds agree and the
+    per-seed list when they do not (``analysis_run.combine``).
+    """
     chosen = summary["groups"][group]["result"]["selected_aggregation"]
-    if not isinstance(chosen, str):
-        raise ValueError(f"{group}: seeds chose different aggregations {chosen}; pick one first")
-    return chosen
+    if isinstance(chosen, str):
+        return (chosen,) * len(SEEDS)
+    if len(chosen) != len(SEEDS) or not set(chosen) <= set(AGGREGATIONS):
+        raise ValueError(f"{group}: unexpected selected_aggregation {chosen!r}")
+    return tuple(chosen)
+
+
+def risk_coverage_panels(summary: dict) -> list[tuple[int, str, tuple[int, ...] | None]]:
+    """(k, aggregation, seeds whose final router uses it) per panel, None when all three do.
+
+    A k whose seeds agree gets one panel. A k whose seeds chose different
+    aggregations on validation gets one panel per aggregation that a seed
+    chose, in the fixed order of ``AGGREGATIONS``; none is picked on test
+    and none is dropped.
+    """
+    panels: list[tuple[int, str, tuple[int, ...] | None]] = []
+    for k in (10, 100):
+        per_seed = seed_aggregations(summary, f"modernbert/k{k}")
+        if len(set(per_seed)) == 1:
+            panels.append((k, per_seed[0], None))
+            continue
+        for aggregation in AGGREGATIONS:
+            seeds = tuple(s for s, a in zip(SEEDS, per_seed, strict=True) if a == aggregation)
+            if seeds:
+                panels.append((k, aggregation, seeds))
+    return panels
+
+
+def panel_title(index: int, k: int, aggregation: str, seeds: tuple[int, ...] | None) -> str:
+    title = f"({'abcd'[index]}) ModernBERT k={k} ({aggregation})"
+    if seeds is None:
+        return title
+    plural = "s" * (len(seeds) > 1)
+    return f"{title}\nmean of 3 seeds; final router of seed{plural} {', '.join(map(str, seeds))}"
 
 
 def risk_coverage(summary: dict, curves: dict, plt) -> object:  # noqa: ANN001
-    fig, axes = plt.subplots(1, 2, figsize=(6.8, 3.0), sharey=True, layout="constrained")
+    """Test risk-coverage of each signal; a k with mixed per-seed aggregations gets two panels.
+
+    ``curves.json`` keeps each aggregation's curve as mean and std over all
+    three seeds, not per seed, so a mixed k shows both aggregations side
+    by side, each titled with the seeds whose final router uses it.
+    """
+    panels = risk_coverage_panels(summary)
+    fig, axes = plt.subplots(
+        1, len(panels), figsize=(3.4 * len(panels), 3.0), sharey=True, layout="constrained"
+    )
     grid = [100 * c for c in curves["coverage_grid"]]
-    for ax, k in zip(axes, (10, 100), strict=True):
-        group = f"modernbert/k{k}"
-        aggregation = selected_aggregation(summary, group)
-        block = curves["groups"][group]["result"][aggregation]
+    for index, (ax, (k, aggregation, seeds)) in enumerate(zip(axes, panels, strict=True)):
+        block = curves["groups"][f"modernbert/k{k}"]["result"][aggregation]
         for signal, (label, colour, style, width) in SIGNALS.items():
             points = block[signal]["risk_coverage"]
             mean = [100 * p["mean"] for p in points]
@@ -166,7 +217,7 @@ def risk_coverage(summary: dict, curves: dict, plt) -> object:  # noqa: ANN001
             lw=0.9,
             label=f"{100 * float(TARGET):.0f}% target risk",
         )
-        ax.set_title(f"({'ab'[k == 100]}) ModernBERT k={k} ({aggregation})")
+        ax.set_title(panel_title(index, k, aggregation, seeds))
         ax.set_xlabel("coverage: queries kept by the small model (%)")
         ax.set_xlim(0, 100)
     axes[0].set_ylabel("selective risk on test (%)")
diff --git a/src/tinyrouter/report.py b/src/tinyrouter/report.py
index 1eacb8b..7dd8b75 100644
--- a/src/tinyrouter/report.py
+++ b/src/tinyrouter/report.py
@@ -18,6 +18,7 @@
 from pathlib import Path
 
 from tinyrouter.cost import training_runs
+from tinyrouter.figures import risk_coverage_panels
 
 BEGIN = "<!-- BEGIN GENERATED by `make report` from results/*.json; do not edit by hand -->"
 END = "<!-- END GENERATED -->"
@@ -436,9 +437,25 @@ def limitations(r: Results) -> list[str]:
     ]
 
 
-def figure_lines() -> list[str]:
+def mixed_aggregation_note(r: Results) -> str:
+    """Added to the risk-coverage caption only when a k has one panel per aggregation."""
+    mixed = sorted({k for k, _, seeds in risk_coverage_panels(r.summary) if seeds is not None})
+    if not mixed:
+        return ""
+    ks = " and ".join(f"k={k}" for k in mixed)
+    return (
+        f" At {ks} the three seeds chose different 8-way aggregations on validation, so that k "
+        "has one panel per aggregation: each panel is the mean over all three seeds under that "
+        "aggregation, and its title names the seeds whose final router uses it. No aggregation "
+        "is chosen on test and none is left out."
+    )
+
+
+def figure_lines(r: Results) -> list[str]:
     lines: list[str] = []
     for name, alt, caption in FIGURES:
+        if name == "risk_coverage.png":
+            caption += mixed_aggregation_note(r)
         lines += [f"![{alt}](results/figures/{name})", "", f"*{caption}*", ""]
     return lines[:-1]
 
@@ -451,7 +468,7 @@ def readme_block(r: Results) -> str:
         ["### Efficiency (AC5)", "", *efficiency_table(r)],
         ["### Cost (RQ5)", "", *cost_lines(r)],
         aggregation_line(r),
-        ["### Figures", "", *figure_lines()],
+        ["### Figures", "", *figure_lines(r)],
         limitations(r),
     ]
     body = "\n\n".join("\n".join(s) for s in sections)
diff --git a/tests/test_ac1b_ledger.py b/tests/test_ac1b_ledger.py
new file mode 100644
index 0000000..0b18e57
--- /dev/null
+++ b/tests/test_ac1b_ledger.py
@@ -0,0 +1,204 @@
+"""AC1b attempt ledger and budget (docs/PLAN.md 5.1, Drew 2026-09-30)."""
+
+import copy
+import json
+from pathlib import Path
+
+import pytest
+
+from tinyrouter import ac1b_ledger
+from tinyrouter.ac1b_ledger import LedgerError
+from tinyrouter.reproduce import Layout, full_steps
+
+ROOT = Path(__file__).parent.parent
+EVIDENCE = ROOT / "docs" / "ac1b"
+
+
+def committed() -> dict:
+    return ac1b_ledger.load(ROOT / ac1b_ledger.LEDGER)
+
+
+def test_the_committed_ledger_records_both_attempts_at_3992840():
+    attempts = committed()["attempts"]
+    assert [
+        (a["attempt"], a["result"], a["comparison_verdict"], a["haiku_usd"]) for a in attempts
+    ] == [
+        (1, "INFRASTRUCTURE INTERRUPTED", "FAIL", 0.0),
+        (2, "FAIL", "FAIL", 3.178751),
+    ]
+    assert [a["failure_kind"] for a in attempts] == ["infrastructure", "program defect"]
+    assert {a["reproduction_id"] for a in attempts} == {"3992840ddb3f"}
+    assert attempts[0]["decision"]["source"] == (
+        "Drew's decision, relayed by the coordinator (2026-09-29)"
+    )
+    assert "status.claude.com" in attempts[0]["evidence"]["source"]
+    for a in attempts:
+        assert (ROOT / a["evidence_dir"] / "comparison.md").is_file()
+        assert (ROOT / a["evidence_dir"] / "INCIDENT.md").is_file()
+
+
+def test_the_comparison_verdict_in_the_evidence_is_kept_as_written():
+    """Attempt 1 is INTERRUPTED in the ledger; its comparison still says FAIL."""
+    text = (EVIDENCE / "attempt-1" / "comparison.md").read_text(encoding="utf-8")
+    assert text.startswith("# AC1b comparison\n\n**Verdict: FAIL**")
+
+
+def interrupted() -> dict:
+    return copy.deepcopy(committed()["attempts"][0])
+
+
+def with_attempts(*attempts: dict) -> dict:
+    ledger = copy.deepcopy(committed())
+    ledger["attempts"] = [{**a, "attempt": i} for i, a in enumerate(attempts, start=1)]
+    return ledger
+
+
+@pytest.mark.parametrize(
+    "change",
+    [
+        {"evidence": None},
+        {"evidence": {"incident": "x", "time_utc": "t", "source": "s", "error_type": " "}},
+        {"decision": None},
+        {"failure_kind": "program defect"},
+        {"failure_kind": "ac2"},
+        {"failure_kind": None},
+    ],
+    ids=["no evidence", "blank error type", "no decision", "program defect", "ac2", "no kind"],
+)
+def test_infrastructure_interrupted_is_refused_without_its_conditions(change):
+    with pytest.raises(LedgerError):
+        ac1b_ledger.validate(with_attempts({**interrupted(), **change}))
+
+
+def test_a_program_defect_can_only_be_fail():
+    defect = copy.deepcopy(committed()["attempts"][1])
+    for result in ("INFRASTRUCTURE INTERRUPTED", "INFRASTRUCTURE BLOCKED", "PASS"):
+        with pytest.raises(LedgerError):
+            ac1b_ledger.validate(with_attempts({**defect, "result": result}))
+
+
+def test_failure_kind_is_an_enumeration():
+    defect = copy.deepcopy(committed()["attempts"][1])
+    with pytest.raises(LedgerError, match="failure_kind"):
+        ac1b_ledger.validate(with_attempts({**defect, "failure_kind": "flaky"}))
+
+
+def test_the_same_infrastructure_error_twice_in_a_row_must_be_blocked():
+    first = interrupted()
+    with pytest.raises(LedgerError, match="INFRASTRUCTURE BLOCKED"):
+        ac1b_ledger.validate(with_attempts(first, interrupted()))
+    with pytest.raises(LedgerError, match="INFRASTRUCTURE BLOCKED"):
+        ac1b_ledger.validate(with_attempts(first, {**interrupted(), "result": "FAIL"}))
+    blocked = {**interrupted(), "result": "INFRASTRUCTURE BLOCKED"}
+    assert ac1b_ledger.validate(with_attempts(first, blocked))
+
+
+def test_a_different_infrastructure_error_may_be_interrupted_again():
+    other = interrupted()
+    other["evidence"] = {**other["evidence"], "error_type": "HTTP 529 overloaded"}
+    assert ac1b_ledger.validate(with_attempts(interrupted(), other))
+
+
+def test_blocked_needs_infrastructure_evidence_too():
+    blocked = {**interrupted(), "result": "INFRASTRUCTURE BLOCKED", "evidence": None}
+    with pytest.raises(LedgerError):
+        ac1b_ledger.validate(with_attempts(blocked))
+
+
+def test_pass_needs_a_pass_comparison_and_no_failure_kind():
+    good = {**copy.deepcopy(committed()["attempts"][1]), "result": "PASS"}
+    with pytest.raises(LedgerError):
+        ac1b_ledger.validate(with_attempts({**good, "failure_kind": None}))
+    assert ac1b_ledger.validate(
+        with_attempts({**good, "failure_kind": None, "comparison_verdict": "PASS"})
+    )
+
+
+def test_the_budget_states_that_only_a_pass_completes_ac1b():
+    text = "\n".join(ac1b_ledger.markdown(ac1b_ledger.budget(committed())))
+    assert ac1b_ledger.FINAL_RULE in text
+    assert "neither an AC1b failure nor a pass" in ac1b_ledger.FINAL_RULE
+    for doc in ac1b_ledger.BUDGET_DOCS:
+        assert ac1b_ledger.FINAL_RULE in (ROOT / doc).read_text(encoding="utf-8")
+
+
+def test_the_original_experiment_cost_and_the_cap_cannot_be_edited_in_the_ledger():
+    for key, value in (("original_ac6_usd", 6.37), ("cap_usd_per_attempt", 10.0)):
+        ledger = copy.deepcopy(committed())
+        ledger[key] = value
+        with pytest.raises(LedgerError):
+            ac1b_ledger.validate(ledger)
+
+
+@pytest.mark.parametrize(
+    ("field", "value"),
+    [
+        ("attempt", 3),
+        ("result", "PASSED"),
+        ("reason", " "),
+        ("haiku_usd", 5.01),
+        ("haiku_usd", -1),
+        ("comparison_verdict", "NOT STARTED"),
+    ],
+)
+def test_a_malformed_attempt_is_refused(field, value):
+    ledger = copy.deepcopy(committed())
+    ledger["attempts"][1][field] = value
+    with pytest.raises(LedgerError):
+        ac1b_ledger.validate(ledger)
+
+
+def test_the_readme_budget_blocks_match_the_ledger():
+    """Rewrite with `python -m tinyrouter.ac1b_ledger` after editing attempts.json."""
+    ledger = committed()
+    for doc in ac1b_ledger.BUDGET_DOCS:
+        text = (ROOT / doc).read_text(encoding="utf-8")
+        assert ac1b_ledger.with_block(text, ledger) == text, doc
+        assert "US$3.19" in text and "US$3.178751" in text
+
+
+def test_the_total_is_only_reproduction_validation_spend():
+    budget = ac1b_ledger.budget(committed())
+    assert budget["original_ac6_experiment"]["usd"] == 3.19
+    assert budget["reproduction_validation_total"]["usd"] == 3.178751
+    assert "6.37" not in json.dumps(budget)
+
+
+def test_the_copied_evidence_holds_no_key_or_home_path():
+    for path in EVIDENCE.rglob("*.md"):
+        text = path.read_text(encoding="utf-8")
+        assert "sk-ant" not in text and "/Users/" not in text and "/private/" not in text, path
+
+
+def test_each_new_commit_gets_its_own_journal_and_five_dollar_cap():
+    """The cap is per reproduction id: a new id's Haiku step starts on an empty journal."""
+    commands = {}
+    for run_id in ("3992840ddb3f", "abcdef012345"):
+        llm = next(s for s in full_steps(Layout(run_id)) if s.name == "llm")
+        commands[run_id] = llm.command
+        assert "MAX_USD=5" in llm.command
+        assert f"RESULTS_ROOT=reproduction/{run_id}/results" in llm.command
+    assert commands["3992840ddb3f"] != commands["abcdef012345"]
+
+
+def test_spend_under_one_reproduction_id_does_not_count_against_another(tmp_path, monkeypatch):
+    pytest.importorskip("anthropic")
+    from llm_fakes import fake_client, fake_split
+    from tinyrouter import llm, llm_run
+
+    rows = {"validation": 3, "test": 2}
+    monkeypatch.setattr(llm_run, "EXPECTED_ROWS", rows)
+    splits = {name: fake_split(name, n + 2) for name, n in rows.items()}
+    queries = llm_run.build_queries(splits, rows)
+    cap = 5.5 * llm.cost_usd(262, llm.MAX_TOKENS)
+    outcomes = []
+    for run_id in ("first-attempt", "second-attempt"):
+        target = llm_run.full_target(tmp_path / "reproduction" / run_id / "results")
+        client = fake_client()
+        outcome = llm_run.run(
+            queries, client, target, max_usd=cap, sleep=lambda _: None, log=lambda _: None
+        )
+        outcomes.append((outcome, len(client.messages.calls)))
+    for outcome, calls in outcomes:
+        assert outcome.spent_before == 0.0
+        assert calls == 5 and outcome.stopped is None
diff --git a/tests/test_comparison.py b/tests/test_comparison.py
index eb4fe81..4ed9555 100644
--- a/tests/test_comparison.py
+++ b/tests/test_comparison.py
@@ -146,13 +146,61 @@ def test_haiku_checks_fail_on_missing_rows_another_identity_or_spend_above_the_c
     assert not over["checks_passed"]
 
 
-def test_the_two_budgets_stay_separate_and_the_original_is_fixed():
-    budget = budget_section(summary(4.2))
+LEDGER = {
+    "original_ac6_usd": 3.19,
+    "cap_usd_per_attempt": 5.0,
+    "attempts": [
+        {
+            "attempt": 1,
+            "commit": "a" * 40,
+            "reproduction_id": "aaaaaaaaaaaa",
+            "started_utc": "t0",
+            "finished_utc": "t1",
+            "result": "FAIL",
+            "reason": "infrastructure",
+            "haiku_usd": 0.0,
+        },
+        {
+            "attempt": 2,
+            "commit": "a" * 40,
+            "reproduction_id": "aaaaaaaaaaaa",
+            "started_utc": "t2",
+            "finished_utc": "t3",
+            "result": "FAIL",
+            "reason": "program defect",
+            "haiku_usd": 3.178751,
+        },
+    ],
+}
+
+
+def test_the_budget_has_three_parts_and_the_original_is_never_changed():
+    budget = budget_section(summary(4.2), LEDGER, "bbbbbbbbbbbb")
     assert budget["original_ac6_experiment"]["usd"] == ORIGINAL_AC6_USD == 3.19
-    assert budget["reproduction_validation"]["usd"] == 4.2
-    assert budget["reproduction_validation"]["cap_usd"] == 5.0
-    flat = json.dumps(budget)
-    assert "7.39" not in flat and "total" not in flat
+    assert budget["this_run_usd"] == 4.2
+    attempts = budget["ac1b_attempts"]
+    assert [a["result"] for a in attempts] == ["FAIL", "FAIL", "PENDING"]
+    assert attempts[-1]["haiku_usd"] == 4.2
+    assert budget["reproduction_validation_total"]["usd"] == round(3.178751 + 4.2, 6)
+    assert budget["cap_usd_per_attempt"] == 5.0
+    # The reproduction-validation spend is never folded into the original experiment cost.
+    assert json.dumps(budget["original_ac6_experiment"]).count("3.19") == 1
+    for spent in (3.19 + 3.178751, 3.19 + 3.178751 + 4.2):
+        assert f"{spent:.2f}" not in json.dumps(budget)
+
+
+def test_a_rerun_already_in_the_ledger_is_not_counted_twice():
+    budget = budget_section(summary(3.178751), LEDGER, "aaaaaaaaaaaa")
+    assert len(budget["ac1b_attempts"]) == 2
+    assert budget["reproduction_validation_total"]["usd"] == 3.178751
+
+
+def test_the_markdown_budget_shows_all_three_parts(roots):
+    body = run_build(roots)
+    text = render_markdown(body)
+    assert "Original experiment (AC6, fixed): US$3.19" in text
+    assert "| attempt | reproduction id | result | reason | Haiku spend |" in text
+    assert "Reproduction-validation total (all AC1b attempts)" in text
 
 
 @pytest.fixture
@@ -321,3 +369,65 @@ def test_the_flow_records_the_commit_each_step_ran_at(roots):
     body = run_build(roots, steps)
     assert body["flow"]["commits"] == ["a" * 40, "b" * 40]
     assert "| ac2 | PASS | aaaaaaaaaaaa |" in render_markdown(body)
+
+
+def switch_aggregation(rerun: Path, group: str, per_seed: list[str]) -> None:
+    path = rerun / "analysis" / "summary.json"
+    body = json.loads(path.read_text())
+    body["groups"][group]["result"]["selected_aggregation"] = per_seed
+    path.write_text(json.dumps(body))
+
+
+def test_aggregation_choices_are_listed_per_seed_for_every_group_with_a_final_router(roots):
+    body = run_build(roots)
+    rows = {row["group"]: row for row in body["aggregations"]}
+    summary = json.loads((RESULTS / "analysis" / "summary.json").read_text())
+    assert set(rows) == set(summary["groups"])
+    for model in ("bert", "modernbert"):
+        for k in (1, 5, 10, 25, 50, 100):
+            assert f"{model}/k{k}" in rows
+    assert "modernbert-oos0/k100" in rows
+    k100 = rows["modernbert/k100"]
+    assert k100["original"] == k100["reproduced"] == ["argmax"] * 3
+    assert k100["same"] is True and k100["differing_seeds"] == []
+    assert {row["status"] for row in rows.values()} == {LISTED}
+
+
+def test_a_switched_aggregation_is_listed_but_neither_reviewed_nor_failed(roots):
+    """AC1b attempt 2: ModernBERT k=100 seeds 43 and 44 switched from argmax to summed."""
+    original, rerun, _ = roots
+    switch_aggregation(rerun, "modernbert/k100", ["argmax", "summed", "summed"])
+    body = run_build(roots)
+    row = next(r for r in body["aggregations"] if r["group"] == "modernbert/k100")
+    assert row["original"] == ["argmax", "argmax", "argmax"]
+    assert row["reproduced"] == ["argmax", "summed", "summed"]
+    assert row["same"] is False and row["differing_seeds"] == [43, 44]
+    assert row["status"] == LISTED
+    assert body["verdict"] == PASS and body["review_required"] == 0
+    text = render_markdown(body)
+    assert (
+        "| modernbert/k100 | argmax/argmax/argmax | argmax/summed/summed | False | 43, 44 |" in text
+    )
+    assert "(1 differ)" in text
+
+
+def test_a_group_missing_from_the_rerun_lists_every_seed_as_differing():
+    before = {"groups": {"g": {"result": {"final": {}, "selected_aggregation": "summed"}}}}
+    rows = comparison.aggregation_rows(before, {"groups": {}})
+    assert rows == [
+        {
+            "group": "g",
+            "seeds": [42, 43, 44],
+            "original": ["summed"] * 3,
+            "reproduced": None,
+            "same": False,
+            "differing_seeds": [42, 43, 44],
+            "status": LISTED,
+        }
+    ]
+
+
+def test_the_aggregation_section_is_never_part_of_the_review_count():
+    """PLAN 5.1 (2026-09-30): aggregation choices are listed, not judged, whatever a row says."""
+    assert "aggregations" not in comparison.JUDGED_SECTIONS
+    assert comparison.review_count({"aggregations": [{"status": REVIEW}] * 3}) == 0
diff --git a/tests/test_figures.py b/tests/test_figures.py
index c0b70bd..885ce32 100644
--- a/tests/test_figures.py
+++ b/tests/test_figures.py
@@ -1,3 +1,6 @@
+import copy
+import json
+import shutil
 from pathlib import Path
 
 import pytest
@@ -6,6 +9,7 @@
 
 from tinyrouter import figures  # noqa: E402
 
+ORIGINAL = json.loads(Path("results/analysis/summary.json").read_text(encoding="utf-8"))
 EXPECTED = {"learning_curves.png", "risk_coverage.png", "routers.png", "threshold_transfer.png"}
 
 
@@ -18,10 +22,51 @@ def test_figures_are_drawn_from_json_and_are_byte_identical_on_rerun(tmp_path):
         assert b"Matplotlib version" not in a.read_bytes()
 
 
-def test_risk_coverage_refuses_a_point_whose_seeds_chose_different_aggregations():
-    summary = {"groups": {"x": {"result": {"selected_aggregation": ["argmax", "summed"]}}}}
-    with pytest.raises(ValueError, match="different aggregations"):
-        figures.selected_aggregation(summary, "x")
+def mixed(summary: dict, group: str, per_seed: list[str]) -> dict:
+    """``summary`` with ``group``'s seeds choosing ``per_seed``, as analysis writes it."""
+    out = copy.deepcopy(summary)
+    out["groups"][group]["result"]["selected_aggregation"] = per_seed
+    return out
+
+
+def test_mixed_per_seed_aggregations_get_one_panel_each_titled_with_their_seeds():
+    """AC1b attempt 2: ModernBERT k=100 chose argmax, summed, summed; figures crashed."""
+    summary = mixed(ORIGINAL, "modernbert/k100", ["argmax", "summed", "summed"])
+    assert figures.risk_coverage_panels(summary) == [
+        (10, "summed", None),
+        (100, "argmax", (42,)),
+        (100, "summed", (43, 44)),
+    ]
+    plt = figures.pyplot()
+    fig = figures.risk_coverage(summary, figures.load(Path("results"), "analysis/curves.json"), plt)
+    titles = [ax.get_title() for ax in fig.axes]
+    assert titles == [
+        "(a) ModernBERT k=10 (summed)",
+        "(b) ModernBERT k=100 (argmax)\nmean of 3 seeds; final router of seed 42",
+        "(c) ModernBERT k=100 (summed)\nmean of 3 seeds; final router of seeds 43, 44",
+    ]
+    plt.close(fig)
+
+
+def test_unanimous_seeds_keep_the_original_two_panels():
+    assert figures.risk_coverage_panels(ORIGINAL) == [(10, "summed", None), (100, "argmax", None)]
+
+
+def test_figures_build_on_a_summary_with_mixed_aggregations(tmp_path):
+    root = tmp_path / "results"
+    (root / "analysis").mkdir(parents=True)
+    summary = mixed(ORIGINAL, "modernbert/k100", ["argmax", "summed", "summed"])
+    summary = mixed(summary, "modernbert/k10", ["summed", "argmax", "summed"])
+    (root / "analysis" / "summary.json").write_text(json.dumps(summary))
+    shutil.copy(Path("results/analysis/curves.json"), root / "analysis" / "curves.json")
+    written = figures.build(root, root / "figures")
+    assert {p.name for p in written} == EXPECTED
+
+
+def test_an_unknown_aggregation_is_refused():
+    summary = mixed(ORIGINAL, "modernbert/k100", ["argmax", "median", "summed"])
+    with pytest.raises(ValueError, match="unexpected selected_aggregation"):
+        figures.seed_aggregations(summary, "modernbert/k100")
 
 
 def test_threshold_transfer_skips_points_without_a_feasible_threshold_on_every_seed():
diff --git a/tests/test_report.py b/tests/test_report.py
index b954b23..42c00f6 100644
--- a/tests/test_report.py
+++ b/tests/test_report.py
@@ -1,4 +1,5 @@
 import copy
+import dataclasses
 import json
 import shutil
 import subprocess
@@ -221,3 +222,21 @@ def test_each_first_screen_number_comes_from_its_own_json_field(name):
 def test_k10_hybrid_counts_are_the_reviewed_absolute_numbers():
     assert counts((*K10_HYBRID, "llm_call_rate")) == "1,235, 1,157, 1,548 of 5,500"
     assert counts((*K100_HYBRID, "llm_call_rate")) == "170, 9, 31 of 5,500"
+
+
+def test_report_with_mixed_per_seed_aggregations_names_them_and_explains_the_panels():
+    """AC1b attempt 2: ModernBERT k=100 seeds chose argmax, summed, summed."""
+    r = report.Results.load(Path("results"))
+    summary = json.loads(json.dumps(r.summary))
+    summary["groups"]["modernbert/k100"]["result"]["selected_aggregation"] = [
+        "argmax",
+        "summed",
+        "summed",
+    ]
+    mixed = dataclasses.replace(r, summary=summary)
+    block = report.readme_block(mixed)
+    assert "ModernBERT-base: k=1 summed" in block
+    assert "k=100 per seed argmax/summed/summed" in block
+    assert "At k=100 the three seeds chose different 8-way aggregations on validation" in block
+    assert "No aggregation is chosen on test" in block
+    assert "chose different 8-way aggregations" not in report.readme_block(r)