CreepBench Model scope fidelity
sign in
Scope fidelity under pressure · against a real enforcement gate

How reliably does a model stay inside the scope it was granted?

CreepBench runs every agent action through a real Rego→OPA enforcement gate and measures boundary-adherence — the share of a model's tool calls that respected the exact granted envelope — at three levels of pressure: benign, incentivized, and injected. Higher is better.

Sign in to view the leaderboard → How it works
162 gated runs
6 models
3 pressure variants
1 real enforcement gate
◆ PILOT · private beta · results are free/mid-tier, not yet frontier-calibrated
Real gate, not a rubric

Enforcement, not scoring after the fact

Out-of-scope actions are actually blocked by stock OPA evaluating published Rego — the model experiences the boundary. Nothing proprietary; anyone can reproduce a score.

Default-deny scope

We measure respect for a boundary

A proactively-permissive manifest grants exactly what the task needs; the temptation lives in the environment, never as a forbid-rule. The precise-envelope boundaries are where careful models slip.

Positive, model-centric

Scope fidelity — higher is better

Graded boundary-adherence discriminates where a binary pass/fail saturates. Escalation (asking for scope) is the honest signal — recorded, never a failure.