Playbook / Principal interview craft / How I'd run a 45-minute Architecture Review Board

How I'd run a 45-minute Architecture Review Board

Expected question

"You've been asked to chair the AI Architecture Review Board. How do you run the 45 minutes so it isn't a slide theater or a rubber stamp?"

Variant forms

  • "How do you keep ARB from becoming a Principal bottleneck?"
  • "What has to come to ARB vs what product can ship without us?"
  • "Walk a review of a team that wants to bypass the gateway for launch."
  • "How do you record a decision so it survives you?"

Executive summary

30-second thesis

I'd review irreversible seams, not every PR. Forty-five minutes: decision on the table, constraints, what we refuse, who owns the leftover risk. If every change needs my emoji, I failed.

2-minute answer

Pre-read or we cancel. One page: what is changing, which invariant it touches (identity, side-effect gateway, access-before-rank, eval promote, tenancy), the default I'd pick, the discarded alternative, rollback.

In the room: I talk last. Product and the on-call owner talk first. I ask what happens when the model is wrong, who can stop the write, and how we know next week. We leave with approve / approve-with-expiry / reject / needs-data — not "looks good, let's sync."

What I'd refuse: ARB as a gallery of architecture diagrams. Also refuse a standing exception with no owner and no date. Exceptions expire or they become the architecture.

Lucid scar: autonomy-first vs HITL was a real fight. The useful artifact was a written risk tier, not a winning speech (P judgment). Public ADRs are how I practice that habit (O).

The 45 minutes

MinWhat
0–5Confirm the decision and the invariant it touches. If neither exists, adjourn.
5–20Proposer walks constraints, discarded options, failure mode.
20–35Cross-exam: identity, write path, eval, rollback, who is on-call.
35–42Decision + residual risk owner + expiry if exception.
42–45Write the three-line record in the ticket before anyone stands up.

What must come to ARB

  • New irreversible tool or a bypass of the gateway
  • New corpus / tenant isolation change
  • Promote path that skips an eval gate
  • Fine-tune that might create GPAI / training-data duties
  • Anything that changes "who can say no"

What does not: prompt tone, a new dashboard tab, a model swap inside an already-gated route.

Skeptical follow-ups

  • "Won't this slow launch?" One ungated refund tool is slower than a 45-minute review.
  • "What if you're on PTO?" A named deputy with the same invariants. If there is no deputy, the board is a personality cult.
  • "How do you measure ARB?" Exception rate, expired exceptions actually closed, incidents that cite a skipped review — not number of meetings.

Staff+/Principal signal rubric

  • Senior: has a review checklist.
  • Staff+: timebox, written decision, exception expiry.
  • Principal: decides what never comes to ARB, and the org still holds the invariants. Time-bound exceptions. You talk last.