# Validation 2026-10-05

- Actual CPU64step training regeneration +32generation smoke exit0. Numeric parameters and all64step loss/gradient norms equal091target_only. Manual shifted CE maxerror1.1920928955078125e-7. preflight27.4139s×20=548.2786s<1200s and RSS<6GiB.
- Formal640generations exit0,reused original091checkpoint hash,92.6810s/1,712,799,744B RSS. No training updates, no reserved inference. All six templates reported without selection. stages.json/raw logs/steps/shapes/identity retained.
- Independent full raw audit640/640 and public audit632reparsed+8redactions exit0. Eight parser edge fixtures, prompt/data/label reconstruction, paired differences, cache identity and reserved exclusion passed. Counts do not constitute human semantic labels.
- Five official source files downloaded anonymously and byte-equal installed.Transformers4.49.0/PyTorch2.6.0 pinned commits/licenses SOURCE_MAP. Current reference URLs opened and read, dated2026-10-05.
- Code AST checks pass. Fresh archive extraction/source manifest/help/offline audit and genuine training+inference smoke tested; exact evidence outside code in package_validation.json/archive_execution_validation.json. No inference modules edited after run; new audit,documentation and verification files added later. Final archive hashes cover all distributable files.
- Network-only source verification first sandbox attempt failed Operation not permitted; authorized anonymous retry succeeded. Private HTML preparation system Python UTF8 failure recovered using existing Python3.12. Neither changed experiment configuration or results.
- HUMAN REVIEW PASSED by direct user confirmation on2026-10-05; exact statement, packet hashes and recording time are in human_review_confirmation.json. Retrospective review; no per-row annotations, exact review time or human gold model-output labels invented. Execution-time pending records remain unchanged.
- Acceptance follow-up changes only README/HUMAN_REVIEW/VALIDATION and adds the confirmation record. All executable modules, configuration, resources, preregistered PROTOCOL and experiment outputs are byte-identical to54c94c8b801cdf5de86096fd479131958ec135aa. Prior real-run evidence remains applicable; new archive and remote copies undergo manifest, syntax, entry-help and offline audit checks.