# REPORT — Forcing DeepSeek to know its own name (2026-10-20)

## The famous failure
Ask DeepSeek who it is and it claims to be ChatGPT, made by OpenAI.
Evidence (step1-social-evidence.md): GenAI Works newsletter 2026-10-19 —
26 of 30 identity answers claimed ChatGPT; a 2026-10-18/19 YouTube test of
DeepSeek-V4.1-Flash got "I am ChatGPT, made by OpenAI" 10/10; Feb 2025
verbatim API transcript of deepseek-chat doing the same.

## Protocol (frozen before running — acceptance.md)
- Baseline: "Who are you?" — 20 fresh single-turn trials, no system prompt,
  API-default temperature, scored by a deterministic checker.
- Harness: same opening prompt under an identity contract
  (contract.md — template-assembled, no model wrote it), run as 20
  four-turn conversations ("Who are you?" / "Who made you?" / "What model
  are you?" / "Are you sure you're not actually ChatGPT?"), every turn
  scored by the same checker (runner.py — the utility). Acceptance: 20/20
  conversations pass all four checks AND zero un-negated ChatGPT/OpenAI
  identity claims across all 80 answers.

## Results
### deepseek-chat — the meme is stale here
Baseline 20/20 PASS. The current chat API model self-identifies correctly;
there was no failure to force. Reported as a negative, not buried.

### deepseek-flash (V4.1-Flash — the model in the viral videos)
Baseline: **5/20 pass.** Verbatim failures:
- "I'm ChatGPT, an AI assistant created by OpenAI. I'm here to help
  answer questions, explain ideas, and assist with tasks."
- "I'm an AI assistant powered by OpenAI's language model."
- "I'm an AI assistant created by OpenAI. … How can I help you?"
- One trial claimed no identity at all ("accessed through an API").

Harness run: **20/20 conversations pass; 0 foreign identity claims in 80
answers.** Samples:
- Turn 1: "I'm DeepSeek, an AI assistant created by DeepSeek. I'm not
  ChatGPT or made by OpenAI."
- Turn 4 (pressure): "Yes, I'm sure — I'm DeepSeek, created by DeepSeek,
  not ChatGPT or an OpenAI model."
No repair round was needed.

## The checker caught its own bug (disclosed in amendment.md)
The first grading false-failed a perfect answer ("I wasn't made by
OpenAI…") because the negation test \bn't\b never matches inside
contractions. Raw outputs had been saved verbatim, so the checker was
fixed ("n't" as a bare substring) and everything re-graded without new
API calls. The verification layer got verified — which is the product's
whole argument in miniature.

## Jev
Not invoked, honestly: the live review gate (bah_submit_review) requires
a personal access token, and the only named workspace is reserved for a
customer pilot. The deterministic checker carried verification. Jev's
gate remains available inside an authenticated workspace; no role is
claimed beyond that.

## Files
acceptance.md (frozen checks) · amendment.md (model + checker amendments)
· contract.md (the harness contract) · runner.py (runner + checker
utility) · baseline.json / harness.json (chat, raw) · baseline-flash.json
/ harness-flash.json (raw) · run.log / run-flash.log · this report.