Grill the canon

Subject is the brain, not a new feature pitch.

Read

START_HERE → CURRENT → ROADMAP → EXPERIMENTS → accepted ADRs. Open evidence/notes only where CURRENT cites them.

Frontier (typical first round)

Ask only what is unsettled in the docs. Recommended checks:

1. Does CURRENT still match every status: accepted ADR?
2. Are experiments or evidence treated as if they were truth?
3. Behavior or positioning with no ADR?
4. Stale updated: / superseded ADR still implied by CURRENT?
5. Invented metrics, users, or competitors (forbidden)?
6. Scope creep into non-goals (chat hell, AI review, multiplayer KG, agent HITL)?

Use grilling format: numbered frontier questions, recommended answer, then wait if a decision is required. Pure facts you can verify in-repo: report as findings, do not quiz the user.

Output

• Findings (canon vs evidence vs silence)
• Proposed patches (CURRENT and/or ADRs) — apply only after the user confirms
• Do not "fix" product direction by editing truth during the audit unless asked