AI2AI date-to-event diagnostic, 12 October 2026

This bundle continues the 9 October source-standing diagnostic with a narrow error class: unsupported chronology attached to statements from document metadata. All new examples are fictional. It includes frozen cases, answer keys, the exact experimental task prompts, every unedited final answer, primary scores, a condition-masked same-system factual-support review, and SHA-256 integrity manifests.

Development: eight cases in two fresh same-system assistant contexts, baseline B and guard G. Later check: six previously unseen cases supplied to those same contexts, with the same handoff task prompt and without repeating the extra guard. Earlier instructions remained available. No answer was regenerated or replaced. Each phase ran within a single session on 12 October; there was no overnight delay.

Development cases, keys, and prompts were frozen locally before responses. Held-out cases and their key were created and frozen after the first results were reviewed, before either later response. This is local sequencing documentation, not independent public preregistration or an independently designed holdout. Shared assistant-level instructions and private execution/coordination text are not included; the provided public task prompts are complete. Exact served model, sampling parameters, tokens, money, and internal inference time are unavailable.

The later check is short-context instructed carryover. It is not the original planned four naturally relevant uncued tasks across two project contexts, which remain pending. Correct baseline answers prevent attributing performance to the added instruction. No model weights were updated.

Run python verify_counts.py from the extracted directory to check answer counts, per-answer word counts, and summary totals. Hashes verify the saved files, not external preregistration.
