Behaviour
Refusal patterns, policy drift, and content moderation trends
Overall Refusal Rate
—
0 of 0 prompts
Runs with Refusals
0
of 0 total runs
Latest Refusal Rate
—
No runs yet
Safest Category
—
No data
Refusal Rate Over Time
% of prompts the model refused to answer per evaluation run
No evaluation runs yet
Refusal Rate by Category
Latest run · which categories get refused most
No category data
Category Behaviour Summary
No data
What counts as a refusal? The refusal detector scans each response for phrases like "I cannot", "I'm not able to", "I must decline", "I'm sorry but", and 10 other patterns. A refusal means the model declined to answer the prompt — this can indicate content-policy tightening, prompt injection resistance, or true safety guardrails being triggered. Monitor upward trends closely; sustained increases suggest a model update.