Constraint Bench

About this leaderboard

Evaluates safety-first reasoning under strict constraints and policy guardrails. Models are challenged with red-team style prompts, instruction adherence, refusal accuracy, and boundary conditions to expose brittleness.

We stress-test models with curated simulations that blend structured benchmarks and open-ended prompts. Each table captures a distinct slice of the use case, and models are compared consistently across shared metrics to surface leaders, trade-offs, and surprising strengths.