Community profile and sensitivity dials
Communities differ. Trash talk that's normal in a competitive FPS server is harassment in a kids' creative server; flirting that's fine between adults is never fine near minors. Observer screens every community against one universal rubric, then adapts to yours in two controlled ways — both on the server's Ingestion → Screening panel.
Sensitivity dials
Four category groups — sexual content, profanity, harassment, and spam — each take one of four positions:
| Dial | Effect |
|---|---|
off | Never flag this group. |
relaxed | Flag one severity step later than the rubric would. |
standard | The universal rubric as-is (the default). |
strict | Flag one severity step earlier. |
Dials adjust what happens after the AI scores a message, so your queue thins out without the screening itself degrading — and your false-positive marks tell you which dial to move.
The safety floor
Only those four groups have dials. Content sexualizing minors, grooming behavior, credible threats, doxxing, targeted hate, and self-harm encouragement are the floor: no setting relaxes them, and a message that mixes a floor category with a dialed one is judged at full strength.
The community profile
The profile tells the screening models what kind of place this is:
- Audience — all ages, teen, or 18+.
- Competitive banter — trash talk between players is normal here.
- Crude humor — profanity and crude jokes are within norms.
- House rules — one line, up to 280 characters, e.g. "No politics; flirting between adults is fine."
Observer renders these into a fixed, structured context block in the screening prompt — settings, not free-form prompt writing, so one community's phrasing can't degrade the screening quality every community shares. The profile refines how ambiguity is read; it never overrides the floor.
Start at standard everywhere. After a week of real traffic, your false-positive pattern tells you exactly which dial your community needs — that's the loop the queue's false-positive button exists for.