Beadle Community Management Agent

• Observe every message through thin platform adapters (Telegram + Discord).
• Classify with deterministic signals (member age, first links, rate bursts, duplicate payloads, domain lists) plus Jev for semantic ambiguity (solicitation, question shape, impersonation, harmful-link intent).
• Gate with a pure reducer: hard vetoes quarantine; uncertainty asks a human; only clean, confident cases can auto-resolve.
• Escalate uncertain cases into a visible quarantine queue (Notion) with the reason and confidence attached.
• Resolve approved actions through the Swytchcode execution boundary (Discord post, Notion ledger write, Resend digest).
• Learn from human resolutions: every override becomes a labeled example; later, similar cases cite it — "because you taught me", never a black box.

The demo moment: a genuine question gets a confident answer while a polished scam hits the visible policy block — same second, opposite outcomes, judgment and confidence on screen.