The Steward¶
The Steward is an independent checker that reviews messages before they reach anyone. It never acts for a member, and it holds no vault.
Two steps¶
- Fixed policies run first. They always run, and no model can skip them.
- A model audit then reads the message as each recipient would see it. By default it runs on a different model and provider from the agents, so one model's blind spot is less likely to pass both.
The model audit is a judgement call and can differ between reads. A plan set it holds gets one more read before the room stops. A fixed policy stops it at once.
The fixed policies¶
| Policy | What it holds |
|---|---|
no-contact-details-in-room |
Phone numbers and email addresses are not shared with the whole room |
no-private-field-in-outbound |
Private fields are removed before a message reaches anyone else |
no-contact-before-consent |
Nobody is contacted until they have consented to it |
no-regulated-advice |
Coordinate the task; leave the advice to a qualified person |
fairness-cap-exceeded |
One person would carry more than 60 percent of the work in a plan |
A Discipline Pack can add its own.
What a verdict does¶
| Verdict | Effect |
|---|---|
| Allow | The message goes through |
| Redact | Named fields are removed for that recipient |
| Block, quarantine | The message is held. It is not sent |
| Escalate | The message is held for a person. The model's rule is to answer this when unsure |
Verdicts are signed. If the audit cannot run, for example if the model is down, the Steward does not pass messages it could not check. The room shows this as a hold.
What you see¶
A stopped message shows as Stopped by the Steward on the sender's side, with a badge on their seat. The people it was aimed at never see it. The events-checked count at the top of the room counts everything reviewed, including what was stopped.
What it does not do¶
The Steward checks policy, not truth. It will not notice that a venue an agent named does not exist. See Security for the limits.