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

A Hacker News "Show HN" post introduced Uncracked City, a detective game in which players interrogate LLM-driven suspects that hold a hidden ground-truth state and deliberately lie, accepting free-form typed and voice input across six languages. The game uses clue-contradiction mechanics to crack each suspect's logic, and includes a "Flag a bad reply" button that relies on human-in-the-loop correction rather than model guarantees to maintain reliability. The post frames the project as a shipped example of multi-agent character consistency and controlled deception under adversarial user probing, the same consistency and jailbreak-resistance challenge faced when deploying persona agents.

Hacker News https://uncracked.city Show HN: Uncracked City, detective game where you question LLM suspects that lie Which summary reads better? Pick one — models revealed after.Both summaries are AI-generated. Production LLM agents can now reliably maintain complex, deceptive personas across six languages, holding back truths until specific contextual clues are presented to dynamically crack their logic. For engineers building conversational systems, this proves that multi-turn adversarial guardrails and conditional state transitions can be managed directly by the model rather than relying on brittle, hard-coded logic trees. This shifts agent design from rigid flowcharts to dynamic, intent-driven state validation, allowing you to ship highly interactive, non-cooperative AI characters with minimal state-machine overhead. This is a consumer detective game built around LLM NPCs that hold a hidden ground-truth state and deliberately lie, with free-form typed/voice input across six languages and clue-contradiction mechanics — effectively a shipped example of multi-agent character consistency and controlled deception under adversarial user probing. The hard engineering problem it exercises is keeping each suspect's story internally coherent and non-leaky across unbounded questions, which is the same consistency-and-jailbreak-resistance challenge you hit deploying persona agents; the "Flag a bad reply" button is a telling admission that reliability here is maintained by human-in-the-loop correction, not model guarantees.