Comment Spam Flood

Design content moderation for a public discussion platform under spam flood.

Functional:
- Authenticated users post comments. Each comment is checked, then either published, queued for review, or rejected.
- Real comments arrive at baseline cadence. Attacker posts LLM-generated near-human spam at high volume across many accounts.
- Some attacker comments will look indistinguishable from real ones. Some real comments are off-topic, low-quality, or angry (false-positive risk).
- Detected attack accounts can be banned, but ban-then-recreate loop is cheap for the attacker.

Non-functional:
- Block at least 70% of spam volume (visible to users) without auto-banning real accounts.
- Real-comment publish rate must stay above 80% during attack (no global review queue).
- Defenses: per-account rate-limit, content classifier, novel-text similarity check, account-age gating, link-density filter, manual moderator escalation queue.
- Layered defense required. Single classifier alone has too many false positives.
- Cannot rely solely on banning: attacker recreates accounts faster than bans land.

Return a CanvasState modeling the comment ingest path, classifier stages, and review escalation.