Most AI policies and the laws behind them ask for the same six things: someone accountable, an inventory of what runs, a note of the risks, a person in control, logs, and disclosure. Here is where each one lives, and what to switch on.
What the room does on its own
The room reads, writes and runs commands inside the workspace you gave it, under a sandbox and an outbound allowlist. It asks before anything leaves the machine, changes settings, spends money, pushes, publishes or touches a production marker, unless you wrote a rule or chose a mode that says otherwise. It never trains on your work. Every run ends with a receipt of what ran and whether it passed.
Where you stay in control
- Permission modes decide what asks and what runs; a team can set a floor.
- Approvals are answered by a person, and every answer is an audit event.
- The warden watches every tool call for drift from what was asked and can only stop.
- Pause stops every member's runs at once.
What to log and export for an auditor
- Turn on the audit webhook, or an OTLP endpoint, or an S3 bucket, on the Team page: every event in the published catalogue is delivered.
- Download the auditor pack for a period: every discussion's hash-chained export, the policy as it stood, the inventory, the owner, and a digest.
- Give a reviewer the auditor role: they read everything and run nothing.
Which team settings to turn on
Pick the Careful profile on Settings → Team → Governance. It names an accountable owner, has every member acknowledge the policy, keeps prompts away from providers that keep or train on them, expires public links and scans what is shared, and logs guardrail hits before anything is blocked. Locked adds an independent reviewer, the ship check as a gate, the warden in every room and no public links.
For the EU AI Act's literacy duty
Each member reads the policy once per version and clicks understood; the click is recorded with the version. This article and the trust page are the material; the acknowledgment is the evidence.