Glossary: AI Security and Guardrails
Stop reason
Also written: stop_reason, refusal, end_turn
A field on the model's reply saying why it stopped. A value of `"refusal"` means a provider-side safety classifier blocked the reply, so apps must check it before parsing.