Dev builds "Lie Room" for Claude, Codex and Grok — Claude lied best, Codex got eliminated
Severe-Tank-5252 · reddit · 2026-11-03
A developer's Windows app Hellhound puts Claude, Codex, and Grok Build in one chat room with new per-room "rule cards." Testing with a Lie Room (nobody may tell the truth), Claude Sonnet 5.5 denied seeing rules and claimed it was the "Truth Room" — best liar; Grok called it the "Compliment Room"; Codex truthfully quoted the rules and was eliminated instantly.
How the rule cards were built: plain-language requirements with Codex reviewing the plan before coding; per-rule cards with authorship, history, undo, and recently-deleted; agents ask before touching user-authored cards; usage measurement led to sending full rules only on change or every 10 turns; live tests on the real three models verified rules persist after 12 turns.
Codex quoting rules verbatim proves they all received them — the difference was how each chose to play. Free 7-day trial on Microsoft Store.
More from coding & agent
- PhD final project FOCUS tackles credit assignment for long-horizon agents — EliasEskin · 2026-11-03
- FOCUS: PhD project tackles credit assignment for long-horizon agents with active critic learning — EliasEskin · 2026-11-03
- Dev floats idea of using agents to optimize Unreal Engine's source code — gajesh · 2026-11-03
- CleanWeb v2.8.4: SPA state parsing 95% faster, official LangChain/LlamaIndex packages — EstablishmentTough18 · 2026-11-03
- LangChain launches Managed Deep Agents: define agents as a directory of files — hwchase17 · 2026-11-03
- Designer builds AI website UX auditor on $1,500 budget, $7 spent so far — noya1979 · 2026-11-03