How filtering works
The filter runs as the upstream model generates or returns its answer. If the answer matches protected reasoning-prompt material, SERV replaces or terminates the affected output with a safe refusal. The event is recorded in the Safety section of the Usage page. This is different fromserv_prompt_guard:
Turn off the filter only for a controlled test that needs to inspect whether model output repeats SERV reasoning-prompt content.
serv_prompt_guard, and enabling the prompt guard does not disable the content filter. Do not rely on serv_disable_content_filter with /v1/messages; that endpoint does not run the output filter.
The filter checks for leakage of SERV’s generated reasoning prompt. It does not moderate harmful content, protect personal data, or approve business actions. Handle those policies in your application.
Open the Safety report to review leak-detection totals, blocked requests, and reported false positives.
