問題文
A service sends each prompt and response to AI Defense for evaluation. Violations appear in the event log, yet unsafe prompts still reach the model. Which statement from the documentation describes this arrangement?
選択肢
- Evaluation happens when the application posts a prompt to the inspection API, so the decision about letting it reach the model stays inside the application itself.
- Evaluation happens as the prompt travels through a proxy that the inspection API brings up, so interception is carried out before the model ever sees it.
- Evaluation happens inside the model endpoint that the inspection API registers, so the vendor of the model is the party that refuses unsafe content.
- Evaluation happens once the model has answered with the inspection API logging the exchange, so the block applies to the following prompt rather than this one.