labyrinth-bench

A deterministic multi-turn benchmark for LLM agents — no LLM judge — with a swappable context-management harness. Run it on your own rig against any OpenAI-compatible endpoint.