How Lucy thinks

Lucy is a gatekeeper, not an autopilot. She scores risk, freezes evidence, and waits for a human. When she talks about gates and wards, that is flavor. The decisions stay yours.

Four principles

1. Never auto-ban. Detection and CAPTCHA can mute and hold. Only an admin (or a human guild ban you already chose to enact) removes someone. A score never becomes a ban by itself.
2. Evidence first. Every automated hold explains itself: score, signals, and the triggering message when there is one. You should never wonder why Lucy acted.
3. Trust-aware Release. Releasing someone from quarantine unmutes them and marks them trusted. Watch unmutes them but keeps early scrutiny. Ban is the only hard exit.
4. Reputation is an extractor, not an actor. Cross-group standing is one more detection signal. Only human-confirmed actions write reputation events. The signal is hard-capped at +25 points, so reputation alone cannot quarantine. There is no positive reputation to farm. A single group (or guild cluster) tops out at Noted.

Persona intensity

Tone is a dial under General in /settings. It changes how Lucy speaks, not how moderation works.

Intensity	Feel
professional	Plain ops copy
subtle	Light sparkle
full (default)	Wards, gates, keys, and the occasional star joke

What she deliberately does not do

Lucy stays out of several common moderation bot features on purpose. Missing them is a product boundary, not a gap to paper over:

• Mass purges and bulk cleanup sweeps
• Antiflood and antiraid lock modes
• Channel locks and timed lockdowns
• Word blocklists and content filters
• Warn history management commands (/warns, /rmwarn)
• Temporary ban or kick variants beyond mute durations
• Moderator notes and custom filter languages
• Full internationalization of bot copy

She focuses on join risk, early-message risk, human verdicts, reports, and optional federation. For setup steps, see Getting started. For how scoring works, see Bot detection.