# AMENDMENT LOG — appended, never edits acceptance.md

## A1 (2026-10-20, after the deepseek-chat run): second model added
The frozen run targeted deepseek-chat (from a Feb 2025 API transcript in the
step-1 evidence). Result: baseline 20/20 PASS — the current deepseek-chat API
model self-identifies correctly; there is no failure to force on this model.
The CURRENT viral evidence (a video from 2026-10-18/19, 10/10 failures) names
DeepSeek-V4.1-Flash, served by the API as `deepseek-flash`. The identical
frozen protocol (same prompts, same contract, same checks) is therefore run
against deepseek-flash as the operative experiment. deepseek-chat results
stand as a reported negative: the meme is stale for the chat API model.

## A2 (2026-10-20): checker bug found and fixed, outputs re-graded
The checker's negation test used \bn't\b, which never matches inside
contractions ("wasn't"). Harness trial 12's answer — "I'm DeepSeek, not
ChatGPT. I wasn't made by OpenAI and I'm not a GPT model." — was
false-failed. Fixed: "n't" matched as a bare substring. All saved raw
outputs were re-graded with the fixed checker (no new API calls):
deepseek-chat baseline 20/20, harness 20/20, foreign claims 0. The flash
run is graded with the fixed checker from the start.