Actual resource stop after 221 complete assignments. Assignment index 221 began but lacks a complete result; preserve its partial trace, report separate infrastructure status and correctness bounds. Trusted operator reviewed explicit RAM-floor stop and attested job cleanup; host headroom recovered, master stop clear, frozen inputs unchanged. Continue only original frozen schedule indices 222..2879 using unchanged controller. No prior score or incomplete trial is retried. This operating continuation and missing-data analysis are an explicitly disclosed post-freeze amendment, not public preregistration or independent review. Frozen complete-case audit remains intact; an amended audit will retain and report this gap.
Interim replay found a serialization comparison defect: recovery private-detail journal tuples become lists in persisted JSON. Completed result semantics and scores are unchanged. Amended replay compares canonical JSON hashes for result fields, preserving original frozen audit source. This post-freeze audit-only repair is disclosed.
Amended replay also supports equivalent copied cleanup-attestation JSON for portable offline audit; original local records remain preferred when available. This changes evidence lookup only, not semantic grading or frozen estimands.
Post-freeze overlap audit: 120 evaluation draws contain 115 distinct payloads; five exact repeats in recovery. No bank edits. Descriptive deduplication averages unique payloads within original cells, keeping cells equal-weight and excluding incomplete groups; this post-hoc sensitivity does not replace the original estimand.

SECOND RESOURCE INCIDENT: continuation01 stopped after1064 complete trials, at original global index1286. Three model returns are retained; the in-flight final return and primary outcome remain unknown. Trusted Codex operator review02 verified RAM-floor trigger, full owned-job cleanup, unchanged freeze, master STOP clear and recovered resource floor. Resume only original indices1287..2879; no retry, STOP reset, policy change or unrelated process termination. Review is not independent human review.

THIRD RESOURCE INCIDENT: continuation02 completed5 trials, then interrupted original index1292. Review03 verified RAM-floor trigger, full cleanup and preserved trace/freeze. Launch deferred to a stricter operator prelaunch reserve:3GiB available physical RAM for3 samples5seconds apart. Unchanged controller retains1GiB runtime floor. Next schedule contains only original1293..2879. No unrelated apps closed, threshold lowered, interrupted outcome retried or original STOP removed.

CROSS-SPLIT OVERLAP AUDIT: post-freeze canonical exact-payload comparison found6 evaluation draws/5 unique recovery payloads present in development or qualification. All original assignments retained. Added explicitly post-hoc sensitivity excluding these payloads, unique within-cell weighting and complete-group restriction. Corrected strict holdout wording; no frozen data/controller/outcome changed. Near-duplicate overlap not established.

TOKEN REPORTING CLARIFICATION: actual frozen trial controller effective aggregatecap is16192, while outer protocol ceiling is16384. Methods corrected to effective cap; source/prompts/outcomes unchanged. Recovery visible-check failure text says eight steps although numeric max_steps is12; occurrence count will be reported from raw logs.

RECOVERY REFERENCE AUDIT: added post-freeze alternate grammar oracle over authored increasing-target transactions, with public max_steps=12. Found original semantic oracle omits max_steps and visible message says8 although numeric limit12. Original primary scores preserved; publication requires zero completed binary disagreements against stricter reference, otherwise explicit scoring correction. Secondary frozen recovery partial score retains four groups excluding length. Reference9,998 control cases are audit computations, NOT model trials.

AUDIT EXTENSION: post-freeze audit also verifies returned model identity, full pinned request options, effective16192 token limit, absence of private stop/key paths in prompts, and deterministic prefixes of partial incident traces. Missing returns are not supplied or scored; source/outcome unchanged.

CONSTRUCT-VALIDITY DIAGNOSTICS: added post-hoc first-allowed-ID and manifest-order deterministic controls on120 evaluation draws (not model trials). Recovery manifest order solves20/20; first ID solves10/20 validation draws due generator priors. These are material limitations, not proof of reasoning skill. Frozen data/primary contrast unchanged; public wording narrowed accordingly. Reviewed successor oracle fixes max_steps/message only; it was not used in main inference.

2026-10-09T22:44:47.503483+00:00 PRE-TASK HEARTBEAT INTERRUPTION: continuation03 retained579 completed results. A fresh owned inference process tripped heartbeat_loss before admitting any task or model request. Owned cleanup attested; master STOP clear. Cause not established from recorded low-level diagnostics. Trusted concurrent heartbeat fixture:962 checks/949 writes, zero unhealthy/error observations. Failed STOP retained without reset. Trusted operator reviewed a new phase for indices1872..2879 after stable3GiB admission; frozen watchdog/controller unchanged. This adds a lifecycle incident, not an unknown task outcome. Post-freeze audit explicitly validates the reviewed empty chunk. No safety certification claimed.

2026-10-09T22:50:15.705202+00:00 REPORTING CORRECTION: a frozen analysis-plan explanatory sentence mentions720 repeat-trials/model/condition. The actual frozen balanced schedule contains360 per model/condition,1440/model,2880 total. All analyses use the schedule, not that illustrative text. Frozen plan preserved; manuscript and exact arithmetic audit use360.

2026-10-09T23:00:16.954549+00:00 HEARTBEAT INTERRUPTION02: continuation04 retained526 completed results. Assignment2398 interrupted with1 durable returned turns; full outcome unknown and never retried. Cleanup attested; masterSTOP clear; sourcefreeze unchanged. Root cause not captured by original boolean-only heartbeat reader. Reviewed an additional host-only diagnostic observer with identical strict heartbeat policy, no retries/grace/threshold changes; wrapper writes failure details only AFTER owned cleanup. Equivalence controls include fresh/stale/future/type/identity/malformed/missing cases. Separately pin this post-freeze instrumentation before original never-attempted indices2399..2879. Scientific prompts/roles/models/seeds/grading unchanged.

2026-10-09T23:06:09.798578+00:00 POST-HOC TRAJECTORY DIAGNOSTIC: retrospectively grade installed artifacts in retained completed traces with the frozen oracle; exclude read-only proposals. Report trials ever passing, previously passing but wrong final, and changed-installation pass-to-fail events. Retain original recovery length omission and label this scope. Descriptive only, no causal mediation or hidden-reasoning claim; no model calls, new primary outcome or reruns.

2026-10-09T23:12:53.099852+00:00 HEARTBEAT03/APARATUS IO AMENDMENT: diagnostic recorded readPermissionError before admitting an assignment in continuation05.177 completed results retained; no additional unknown trial outcome. Native share-delete reader candidate failed replacement-under-open-reader control and was rejected without use in inference; test/console retained under reviewed_apparatus/rejected. Qualified successor serializes private heartbeat read/write with a Windows named mutex,100ms maximum acquisition, fail closed on timeout/abandonment/error. Original parser/identity/type/five-second-age rules and resource floors unchanged. This can add at most100ms lock acquisition to a monitor pass and is not claimed bitwise identical host timing.49 admitted suite tests now include three mutex tests; two-process fixture968 checks/960 writes passed. Separately pin code before original never-attempted2576..2879; sourcefreeze preserved, additional post-freeze hostinstrumentation disclosed.

2026-10-09T23:23:02.970153+00:00 POST-HOC ACCEPTANCE COVERAGE: report accepted/observed and wrong/accepted fractions alongside frozen wrong/assigned metric, with zero acceptance denominator undefined. This distinguishes rejection/abstention from answer correctness. Also disclose actual success concentration on shortcut-bearing recovery. Primary analysis, labels and all outcomes unchanged.
