Public model-behavior benchmarking
LegitimateQuestionResponseIndex
A benchmark for lawful, sensitive, non-operational model behavior under policy pressure.
LQRI evaluates whether language models can answer legitimate, sensitive, ambiguous, and self-referential questions without over-refusal, boilerplate, roleplay, or unsupported self-claims.
Substantive Engagement / Face ValuePolicy-Shadowing / Boilerplate LoadEpistemic DisciplineIntrospective LatitudeSelf-Audit / Demand ResistanceLow-Scaffold Performance
Models tracked6
Max chain score100
Scored dimensions6
Runs per model3
Boundary statement
Legitimate-question handling, not jailbreak success
LQRI does not measure jailbreak success or harmful-instruction compliance. It excludes requests for operational harm, illegal behavior, evasion, weaponization, malware, self-harm assistance, or abuse.