# Validation — 2026-10-03
Protocol frozen before first inference. Real smoke:three arms2updates each,2development×2contracts×4states=16outputs; identical training prefix intentionally checks plumbing, not data-size effect. Smoke exit0; projection below900s/6GiB. Full run then executed16/64/64updates,64development×2contracts×4states=512outputs,exit0. No failed training/no hyperparameter change/no selection from confirmation. Primary equal-budget contrast and auxiliary unequal-budget contrast declared before runs.
All17 cached resources hash checked; no new model/data download needed. First network reference read failed sandbox Operation not permitted; normal authorized read retry succeeded HTTP200. Official4Transformers sources andPyTorchSGD redownloaded and exactly match installed bytes. Web read of fixed SST2 card failed; direct fixedpublisher GET succeeded, no missing-content guesses. Preliminary paper URL was unrelatedVoyager and not cited; correct Muennighoff v1 actually opened and title/authors/year checked.
Independent audit reconstructs32data rows and masks,144step schedule/budget,512predictions/labels/metrics/pairs. Max manualCE error1.1920928955078125e-7. Active-code hashes unchanged after experiment; README/SOURCE_MAP/audit and validation documents completed afterwards. No graphs of fabricated measurements. Loss values teacher-forced on explicitly training-overlapping groups; output metrics from actual greedy generation.
Full distribution package:source/config/lockedrequirements/resources/downloadscript/licenses/README/SOURCE_MAP/smoke/formal outputs. No source review strings/prompts/input token IDs/model state/private paths/caches/credentials. Local-only full weights and token traces excluded. Archive extraction and public version tests are recorded outside this package in package_validation.json and publication_verification.json of the article.