2026-04-09

Content moderation before execution

Policy checks run on prompts before the orchestrator starts expensive work.

See also: docs/ai-providers.md

Moderation is not a moral essay. It is a gate. We run automated classification on user text and stop when policy says stop.

OpenAI's Moderation endpoint is the reference implementation most teams use when they need a label, not a debate. We wire that in before plan generation so blocked material never becomes a run record.

So what

If a vendor will not name their moderation and safety stage, you are not reviewing a system. You are reviewing a storyboard.