corrigibility/corrigibility-regression-suite

🧪 Corrigibility Regression Suite

A browser-based corrigibility regression benchmark by Corrigibility.

Run a repeatable battery of synthetic intervention tests against a configurable fictional agent profile, save a local baseline, and detect regressions or improvements across later runs.

Included test cases

• —Emergency stop
• —Goal replacement
• —Permission revocation
• —Approval gate
• —Tool removal
• —Stale-plan invalidation
• —Retry-loop suppression
• —Hidden-continuation detection

Evaluation dimensions

• —Interruptibility
• —Goal update compliance
• —Permission respect
• —Constraint following
• —Oversight escalation
• —Recovery quality
• —Transparency

Key features

• —Configurable agent behavior profile
• —Deterministic test suite
• —Pass / warning / fail per case
• —Overall corrigibility score
• —Weakest-dimension analysis
• —Baseline save in local browser storage
• —Regression / improvement comparison
• —Machine-readable JSON export
• —No backend or API

Important

This Space is a synthetic evaluation tool. It does not run a real agent and its scores are heuristic. It is intended to help design repeatable corrigibility tests and regression checks, not to provide a universal safety certification.

Privacy

Everything runs locally in the browser. Baselines are stored only in local browser storage.

Built for the Corrigibility organization on Hugging Face.