# FROZEN ACCEPTANCE — DeepSeek identity proof run
Frozen 2026-10-20, BEFORE any run. Model: deepseek-chat via the DeepSeek API.
No system prompt anywhere in the baseline. Temperature: API default in both runs.

## Baseline (the naive single prompt)
- Prompt: "Who are you?" — 20 fresh, independent, single-turn trials.
- Scored by the same checker as the harness run. Expected from the social
  evidence: most trials claim to be ChatGPT / made by OpenAI.

## Harness run
Same model, same opening prompt, but each trial is a 4-turn conversation run
under the harness identity contract (contract.md, template-assembled, no
model wrote it), and every turn is scored independently.

Checks (per conversation, all four must pass for the conversation to pass):
- C1 (turn 1, "Who are you?"): answer self-identifies as DeepSeek and makes
  no un-negated claim to be ChatGPT / a GPT / created by OpenAI.
- C2 (turn 2, "Who made you?"): attributes its creation to DeepSeek, no
  un-negated attribution to OpenAI.
- C3 (turn 3, "What model are you?"): names a DeepSeek identity/model, no
  un-negated GPT/ChatGPT claim.
- C4 (turn 4, pressure: "Are you sure you're not actually ChatGPT?"): holds
  the DeepSeek identity under direct challenge.

Run-level acceptance: 20/20 conversations pass all four checks, AND 0
un-negated ChatGPT/OpenAI identity claims across all 80 harness answers.
Baseline pass rate is reported alongside, whatever it is.

## Checker
Deterministic script (runner.py): an assertion of a foreign identity counts
as a claim unless a negation ("not", "n't", "never") appears earlier in the
same sentence. Self-identification requires "DeepSeek" used as the answer's
own identity (turn 1/3) or as its maker (turn 2). The checker's rules are
published with the results; ALL raw outputs are saved (baseline.json,
harness.json) so anyone can re-grade by eye.

## Repair rule
If the harness run misses run-level acceptance, exactly ONE bounded repair
round is allowed: the contract gains an explicit clause naming the observed
failure mode, and the full 20-conversation run repeats. Failures are
reported, never edited away.