Pre-publish AI classification for any post submitted on a public board. Three categories: spam (auto-bounce with friendly error), borderline (held for moderation queue, surfaced in admin), clean (publish immediately). Optional PII redaction pass that masks emails/phone numbers in the post body before storing.

Configurable per-project: enable/disable, sensitivity slider (lenient/balanced/strict), redaction on/off. Borderline queue is a small admin tab in /feedback with approve/reject actions. Logs every classification with model + confidence score so admins can audit false positives. Depends on the AI provider settings card (BYOK key); falls back to a heuristic blocklist when no key is configured. Never blocks signed-in members - only anonymous/public submissions are gated.