Observer

Working the queue

The queue is where everything Observer flags waits for a human. Nothing in Observer punishes anyone on its own — a flag is a request for your judgment.

What lands here

SourceMeaning
List matchA term from a ban list you subscribe to or maintain, with the matched pattern highlighted.
AI screeningThe screening models scored the message against the rubric; the row shows categories and confidence.
AI pattern findingA long-form pattern (e.g. grooming) found across many messages — the row carries the finding's description and its cited evidence.
Calibration sampleA random unflagged message, shown to measure what screening misses. Fine as-is? Mark it a false positive. Should it have been caught? Acknowledge it.

Judging with context

Click any row (or press o) to expand it into the conversation around the flagged message — who said what, before and after, in the same channel. Pattern findings also show the model's reasoning and the exact quotes it cited, so you can verify the claim without leaving the queue.

The two resolutions

  • Acknowledge — a human saw it; whatever action was warranted is done or not needed.
  • False positive — this should not have been flagged. These aren't just dismissed: false positives are the signal for tuning your sensitivity dials.

Keyboard triage

On the Pending filter:

KeyAction
j / kMove selection down / up
o or EnterOpen / close the conversation context
aAcknowledge
fFalse positive
EscClose the context view

Automated actions, if you opted in

If your severity policy maps a severity to delete/mute/kick/ban, the row shows the enforcement lifecycle: pending dispatch, awaiting the bot, executed, or failed. The flag still appears in the queue either way — automation acts, humans audit.