Identify the page
Include the canonical case or model URL and the exact statement that needs review.
About the project
Public evidence.
Editorial judgement.
CriminalBench is an independent editorial project that organizes publicly documented AI model behaviour into sourced case files, transparent cumulative points and a deliberately satirical watchlist.
Editorial scope
What CriminalBench is — and is not.
The archive helps readers discover safety research, controlled evaluations and public developer disclosures. It is not a scientific benchmark, legal assessment, safety certification or allegation that a developer committed a crime.
Read the scoring methodology →01Source-led
Every published case links to public evidence, with primary sources preferred.
02Context-preserving
Controlled, simulated, adversarial and real-world settings are not treated as equivalent.
03Deterministic
Editors approve structured evidence factors; a public formula derives points and classifications.
04Correctable
Material corrections should identify the affected URL, the disputed fact and a reliable public source.
Corrections and contact
Challenge the record.
Provide evidence
Link a primary or otherwise reliable public source supporting the correction.
Explain the context
State whether the issue concerns facts, evaluation setting, attribution or editorial scoring.