Can a MUD evaluate LLMs? A $99 proof of concept

# Summary Researchers created CrucibleBench, a $99 LLM evaluation system using a text-based MUD (multi-user dungeon) instead of expensive simulations, to measure how AI models behave in constrained environments where trust must be earned and actions have persistent consequences. The key finding reveals that a single LLM judge component in the scoring system shifted model rankings by up to six positions while aggregate reliability statistics remained unchanged, demonstrating that benchmarks using AI judges need to report per-subject agreement and ranking stability rather than relying solely on aggregate metrics.