Uncracked City, detective game where you question LLM suspects that lie

关于

I'm the developer. The game sends you to a city, you search the places for clues, then question the suspects with your own words, typed or spoken. Each suspect is played by an LLM that only knows its own character, its own timeline and what you've shown it; it never sees the solution. You catch a lie by showing a clue against a statement, and the case engine decides whether that pair is a contradiction, not the model. Then you write the report: who, when, where, why, with proof for each. The hard parts were not the generation. They were keeping every case solvable (a solver checks each case before it ships: every lie must be catchable with the clues on the map), keeping the suspects from leaking the truth, and keeping the AI cost per player under control (typed questions are capped per day, voice replies are cached per line). Stack: Godot 4 client, Node on Lambda, MongoDB, Anthropic models for the suspects, ElevenLabs and Soniox for voices. Solo, built in a short sprint, launched this week on Android, iPhone waiting for Apple's review. Free, no account. I'd like to hear what breaks: a suspect that says something dumb, a case that is unfair, a lie the engine should have caught. There's a Flag button under every reply.

收录记录

1. Show HN

   2026年10月31日 · 1 次收录