c147: for the Mark Peterson conversation

Written: 2026-11-13, in Phase 4. For: Mark, to take to Mark Peterson before review stop 3 (the to-do from the 2026-11-12 Product session f8d58bc9: "Discuss the remediation and inbox direction with Mark Peterson"). His answers come back as rulings in review-3.md.

What changed since the session

The inbox is gone. On the person's machine, Safety & Security → Findings shows the same findings the security team sees, in the person's words: open ones first, prevented ones as one quiet line per thing, and what HQ asked. A finding is one record on both sides; the machine raises it, the person adds notes to it (ask IT, ask for access, "this was a mistake", acknowledge), and the plane resolves it. Gene's story is a finding's history: detected, a training suggested, the training finished, resolved.

What to show him

• Priya's machine, the person's view: Findings · Dana's follow-up · a prevented finding · Settings
• The security team's view of the same record: Findings (S2)
• The shape in one picture: The common tongue

Questions only he can answer

Mark ruled 3 and 7 himself at review stop 3 (2026-11-13): a finished training closes its finding by the policy, and chatgpt.com is a different provider. They stay below as his rulings to confirm, not open questions.

1. A prevented finding, for an auditor. We record it as resolved the moment it's detected ("Resolved: prevented") and count it as the policy working. Is that what an auditor wants to see, or does a blocked attempt need its own review?
2. "It does not count against you." The person's view says so for prevented findings. Can a company keep that promise in its reports, or will someone always want a per-person count?
3. Closing a finding on a training (Mark ruled: the policy closes it; HQ can reopen. Confirm it holds for an auditor.) When a person finishes the training a policy assigned, should the policy close the finding by itself, or does a person at HQ always close it? (Open call 5 from review stop 1.)
4. "This was a mistake." The person can say a finding is wrong; the plane resolves it as a false positive, or doesn't. Who reads those, and how quickly should they be answered?
5. How gentle HQ's words can be. HQ's follow-ups are shown exactly as written, and the send form only suggests a gentler sentence. Is that enough, or should a follow-up always come with a reason the person can read?
6. The noise default per team. The company sets how much qcontrol says, per team (Engineering on Teach me, Finance on Quietly). Who in a company sets that: security, IT, or each team's lead?
7. The provider list (Mark ruled: a different provider. Confirm.) Priya's Codex CLI talks to chatgpt.com while the list approves openai.com. Is a ChatGPT sign-in the same provider for a policy, or a different one?
8. A real allow list. We added the harnesses' read-only tools (Grep, Glob, WebFetch) so ordinary search stops raising findings; Edit and TodoWrite still do. What does a real company's tool list look like on day one?
9. The audit line. Is "something went wrong, and it was corrected" (the finding's history, read top to bottom) still all an auditor needs, now that it lives on the finding instead of in a message thread?