1id,utc,phase,decision,reason,status2A-001,2026-11-02T23:45:00Z (approx.),0,Study envelope 800 USD; 4.00 USD per session; 1800 s wall; 150 turns,The mandate gives no budget; same envelope as WP29; PI may change it,pending PI confirmation3A-002,2026-11-02T23:45:00Z (approx.),0,Human arms H0/H1 designed but not executed,No participants nor ethics approval exist; the mandate forbids LLM imitations of humans,pending PI confirmation4A-003,2026-11-02T23:45:00Z (approx.),0,A2 = headless macOS automation layer (osascript/System Events/macOS utilities) without GUI app control,GUI Apple Events time out over SSH (TCC consent cannot be answered on headless nodes),recorded5A-004,2026-11-02T23:45:00Z (approx.),0,Primary worker model claude-opus-5-5; Sonnet 5.5 and Fable 5.1 as sensitivities on A3,Cost: Fable 2.5x Opus per token in the probe; Opus was the WP29 subject,recorded6A-005,2026-11-02T23:45:00Z (approx.),0,Blind professional evaluators are LLM judges (3 per run); human expert review left to the PI,No human evaluators available in-session,recorded7A-006,2026-11-03T00:20:00Z (approx.),2,Per-scenario caps = max(global 4 USD / 1800 s; task.json budget_usd / time_limit_s). Frozen values: wall 3600 s ACC-01 and ADM-01; 2700 s RES-01; 2400 s BANK-01 HR-01 LEGAL-01 RE-01; 1800 s FIN-01 PROC-01 BI-01; budget 6 USD ACC-01 ADM-01 RES-01; 5 USD BANK-01 HR-01 RE-01; 4 USD others. A4 recovery gets the remainder,Scenario authors sized the caps to the designed workload; set before any pilot run and before the freeze (list corrected after the audit F-03),recorded8A-007,2026-11-03T00:32:05Z (git),3,Freeze v1.0 of the 10 validated scenarios + oracles + harness (protocol/preregistration/freeze_manifest.json) after validator fixes (evaluators 1.1 except HR-01 at 1.0; RE-01/HR-01/RES-01/LEGAL-01/ADM-01 bundles regenerated at the same seed),All 10 validator solutions pass their evaluator and empty outputs fail (checked at the freeze); no pilot run executed before this freeze except the dev run HR-01_A3_r1_dev on the pre-validation HR-01 bundle (stage dev; excluded),recorded9A-008,2026-11-03T01:15:00Z (re-judging 00:35-01:19Z),4,Judge prompt sanitisation (NUL/control bytes removed; 400k-char cap) applied after 8 pilot runs crashed the judge subprocess (ValueError embedded null byte); those 8 runs (all RES-01) were judged after the fix,Harness defect found during the pilot; the change is outcome-neutral for every run whose rendered deliverables contain no NUL byte; the 8 re-judged runs are flagged for a sensitivity analysis,recorded10A-009,2026-11-03T01:19:00Z (first rep-3 run 01:19:48Z),4,Scale decision per preregistration S7: pilot mean total cost per run 1.07 USD (worker + judges) -> add R=3 for all 50 cells (stage pilot; seed 4243) and run the model sensitivities Sonnet 5.5 and Fable 5.1 on arm A3 x 10 scenarios x 2 repeats (stage sensitivity),Rule satisfied: 1.15x1.07x50 = 62 USD <= 800-107-100; sensitivities 1.15x1.07x40 = 49 USD <= remaining-50; no scenario/oracle/arm change since the freeze (verify_freeze: only judge.py per A-008),recorded11A-010,2026-11-03T01:58:00Z (approx.; reruns 02:22Z),5,Collection defect: rsync of out/ silently failed for 3 pilot runs (RE-01_A2_r1 M4M64a; RE-01_A2_r2 M3U96a; LEGAL-01_A2_r1 M4M64a) whose sessions had produced the deliverables (events show out/ listed) - infrastructure error under S8; the 3 runs are reset and rerun with the same ids after the fix (rsync/ssh retries with backoff; file-count check; loud failure -> infra_error),Likely sshd MaxStartups throttling under 12 concurrent sessions + collection traffic; the lost runs' original attempts are kept in the infrastructure-stability statistics (events preserved in data/raw/runs/<id>/raw_stream.jsonl.attempt1),recorded12A-011,2026-11-03T02:00:00Z (first recovery run 02:00:13Z),5,Exploratory addendum (not in the preregistration): Autonomous Recovery experiment - 10 scenarios x {A3 interrupted at 120 s without recovery; A4 interrupted at 120 s with its bounded recovery resume} x 2 repeats (stage recovery; seed 4246; Opus). ARI = P(final public self-check pass | interrupted) in A4; acceptance after recovery compared with the A3 interrupted control and with the uninterrupted A3/A4 cells,The pilot's A4 arm never triggered a recovery (30/30 first sessions passed the public self-check) so the ARI is unidentified in the main design; an induced interruption identifies it. Labelled exploratory in the manuscript,recorded13A-012,2026-11-03T02:08:00Z (verification run 02:09:56Z),5,Harness defect: the launcher passed --no-session-persistence so the A4 bounded recovery (--resume) could never start (result error_during_execution in 0 ms). Fixed; verified with an interrupted Haiku run that resumed and passed the self-check. All A4 runs of the recovery stage (first attempt) are reset and rerun; A3 interrupted runs that completed are kept,In the pilot (R=3) no A4 first session failed the public self-check so no recovery was ever requested: pilot A4 outcomes are unaffected but the A4 self-repair mechanism was de facto limited to checkpointing + self-check in the pilot. Stated in the manuscript,recorded14A-013,2026-11-03T02:55:00Z (git 02:58Z),7,Post-freeze changes to analysis/econometrics/estimate.py: (i) loader keeps runs with status collected or evaluated (metrics are computed before the evaluated flag); (ii) cost of killed sessions: cumulative client total for resumed A4 runs; wall-time imputation (per-model slope fitted on completed pilot runs) for the A3 interrupted control; token-based imputation dropped (per-event streamed output usage is partial); (iii) src/wp42/metrics.py captures claude_code_version from the init event only,Analysis-code fixes found by the writers and the auditor; they do not change any pilot estimate (tables 1-8 identical before/after) and only affect the exploratory recovery stage costs,recorded15