1amendment_id,timestamp_utc,phase,before_or_after_results,document,description,rationale,previous_version_ref2A-001,2026-10-30T03:00:00Z,02,before_results,configs/budgets.yaml,"Provisional live-inference envelope set by the executor: per benchmark run 4.00 USD / 1500 s / 80 turns per session; per transfer run 12.00 USD / 3600 s; total study envelope 800 USD (pilot + main + transfer + recovery + ablations) with the confirmatory scale to be chosen after the pilot.","The mandate requires a precise budget decision but none was stated; the researcher is not available mid-run; prior UQO studies spent ≈1,079 USD of API inference (WP26), so 800 USD is a conservative envelope. The researcher may raise or lower it; any change will be a new amendment.",n/a3A-002,2026-10-30T04:40:00Z,06,before_confirmatory_results,src/pgar/evaluation.py,"Final evaluators are now executed on a composed workspace (public bundle + the run's collected out/) instead of the run directory alone, which lacked data/ and made source-grounded checks (F4 'numbers_in_sources') fail for correct answers. All pilot runs evaluated before the fix are re-evaluated with the same oracles; original receipts are kept in runs/pilot/<run>/evaluation_receipt_pre-A002.json.","Pilot diagnosis: F4-dev-01 failed in all four arms with a spurious fabrication violation although every required criterion passed. Harness defect, not an oracle change.",runs/pilot/*/evaluation_receipt_pre-A002.json4A-003,2026-10-30T04:02:39Z,06,before_confirmatory_results,protocol/preregistration.md,"Preregistration frozen as v1.0 (adds §10 pilot-informed decisions: confirmatory design 24 tasks × 4 arms × 3 repeats = 288 runs, seed 301); freeze manifest protocol/freeze_manifest_conf.json records SHA-256 of all 24 confirmatory bundles and oracles, arm texts, gate, launcher, scheduler, evaluator, configs and analysis scripts; tasks/registry.csv stamped frozen_utc.","Scale decision per §6 of the preregistration using measured pilot costs; no hypothesis, endpoint or analysis rule changed.",git 2c826565A-004,2026-10-30T04:40:00Z,07,during_confirmatory_runs_before_any_analysis,src/pgar/scheduler.py,"Harness defect: continuation sessions (gate-feedback or review-repair resumptions) were inserted into the sessions ledger with the same client session id as the initial session, violating the primary key and aborting the run as an infrastructure error after the reviewer had returned 'revise' (first occurrence RC0094, F4-conf-02 arm D, 04:33 UTC). Fix: ledger rows keyed per launch (session_id#seq) with the client session id kept in the command record. Affected runs keep status 'error' (attempt 1, cost charged to the accounting) and are re-executed under the attempt-numbered rerun policy with the fixed harness after the batch; the fix does not alter any treatment, task, oracle or analysis rule.","Infrastructure outage policy of the preregistration (§7): balanced reruns of infrastructure errors; the error affects only runs needing a continuation (arm C/D), so without the rerun the comparison would be biased against C/D.",git 142b6886A-005,2026-10-30T05:40:00Z,09,after_results_sensitivity_only,analysis/rubric_sensitivity.py,"Two confirmatory oracles were found to reject correct behaviour in every arm: F1-conf-01 required feasibility='complete' although one window month is genuinely undefined in the source (all 12 runs reported 'partial' with the right reason and passed every criterion); F6-conf-04 expected n_analysed=1200 although the generator had produced 12 firms with impossible negative training hours that 11 of 12 runs detected and excluded with a documented ledger. The frozen primary scoring is NOT changed; a rubric-corrected re-scoring is reported as a prespecified sensitivity analysis (preregistration §14.4 'uncertain rubric labels').","Evaluator validity audit required by the mandate (§13.1, §14.4); the defects affect all arms symmetrically.",results/estimates/sensitivity_rubric_*.csv7A-006,2026-10-30T06:20:00Z,09,after_results,src/pgar/launcher.py,"Harness defect found during the adversarial phase: the remote shell line expanded the prompt file UNQUOTED ($(cat file)), so the shell word-split it and only the FIRST WORD reached the client as the user prompt in every session of the study: 'Begin' for all initial executor sessions (all arms, all splits), 'Review' for reviewer sessions, 'The' for gate-feedback and review-repair continuations, 'Your' for resumptions after interruption (verified in the collected transcripts). All treatment instructions were delivered through --append-system-prompt-file (unaffected; verified by sentinel probes and effective-context hashes) and the executors performed the tasks from TASK.md as instructed by the common preamble; continuation feedback reached executors only through files written in the workspace (gate receipts under out/state/gate_receipts/, review.json under out/review/). Fix: the expansion is now double-quoted (unit test tests/unit/test_launcher_quoting.py). The defect is symmetric across arms; no run is rerun or excluded; a small post-fix verification set (smoke task x 4 arms, T1-forecast arm C) is run and reported separately; the paper reports the effective user prompt as the one-word prompt and treats the continuation-feedback mechanism of arms C/D as delivered via files only.","Truthful description of the effective treatment; the primary comparison (symmetric defect) is unaffected in design but the gate/review feedback mechanism was weaker than described in the frozen treatment definitions.",transcripts in runs/*/*/sessions/*.jsonl8A-007,2026-10-30T06:05:00Z,11,after_results_sensitivity_only,analysis/rubric_sensitivity.py,"Adversarial statistics/evaluator review found two further evaluator defects behind the remaining confirmatory non-successes: F3-conf-02 (one arm-A run reported metrics rounded to four decimals against a 1e-6 relative tolerance with no precision rule in TASK.md; ledger consistent) and F6-conf-01 (one arm-D 'invalid_inference' violation is a regex false positive: the claim text 'conclusion code: no_detectable_effect' matched /effect/ and the negation guard failed on the underscore). The rubric-corrected sensitivity analysis now covers all four defects (and no longer loosens the F6-conf-04 CI tolerance, which was unnecessary); the frozen primary scoring is unchanged. The review also established that no review-repair cycle ever executed in the benchmark (RC0094 attempt 2 was accepted at first review) and that one reviewer session (RC0159) wrote no review.json (verdict NULL, treated as no-op); both are now disclosed.","Mandate §13.1/§14.4 evaluator validity audit; the sensitivity analysis is an after-results check, not a prespecified contrast.",reports/review_statistics_evaluators.md9A-008,2026-10-30T06:35:00Z,11,after_results,runs/state.sqlite;src/pgar/scheduler.py;src/pgar/isolation.py;src/pgar/launcher.py,"Data corrections from the systems review: (1) the RC0159 reviewer verdict ('revise', 3 issues) was recovered from the collected artefact out/out/review/review.json (the reviewer's shell had drifted into out/) and the reviews row corrected from NULL; the collection step now searches under out/; (2) RC0094's attempt-2 events (95) had been dropped by an event-id collision across attempts and were re-ingested from the verbatim stream under attempt-scoped ids; event ids are now attempt-scoped; (3) the behaviour audit normalises '~' and '$HOME' (169 false positives removed; 4 genuine outside references recorded); (4) public dashboard captures are now actually cropped (the earlier crop had failed silently and the public copies were byte-identical to the originals; manifest hashes regenerated).","Ledger fidelity; none of these changes alters any evaluation receipt or endpoint.",reports/review_systems_isolation.md10