Historical local model trials from2026-09-30. Only clean runs are included. Each run uses an earlier evaluator; these are not newly measured current-evaluator scores or official ALE ratings. The eight preserved task inputs match the current submission hashes exactly. Current scoring implementation and explicit delivery wording differ. GPT filename-renaming0.2707 is an author counterfactual with the old evaluator, not a clean agent score. No human expert sign-off or completion-time claim is made.
