Skip to main content
root / home / blog / ea-002-architecture-review-board.md
2026-11-15 11m_READ

Leave Architecture Review With a Decision, Not Homework

A review board earns its place by what leaves the room, not by what it approves. Three seats, four questions, four legal endings and the rules for writing advice back.

#architecture #design-review #decision-making #engineering-leadership

Here's what I learned the hard way: an architecture review is worth exactly what leaves the room, and for a while ours left nothing. Items sat waiting on dates that had already passed, and a model drafted a reply in the room's name promising things nobody had said.

By the end of this post you'll have the method I use now (who sits in the room, the four questions to ask, the four ways a session may end, and how the advice gets written back), so whoever brings you their work leaves with a decision, not homework.

A mole holding a stray item in the gap between two noticeboards, where neither one keeps it.

The intake list

It started on our own intake list. An item still scheduled for a past date drops off the agenda view, because the agenda only looks forward, and the decision log shows only items that reached decided, so an item that's neither vanishes from both. 12 days past its date, it was invisible to anyone not holding the spreadsheet open. That same week the log stopped updating while two forums came and went, and the agenda sat empty with an item scheduled for it. No view was wrong, and the record of decisions stopped moving anyway. Nobody said no, so nobody owned the no. Nobody said yes, so nobody was building.

That's the first way a forum fails: it leaves nothing behind. The documents in front of the room are three of the seven from Part 1: the request for comments (RFC), the design doc (SAD) and the decision record (ADR). Each item arrives as one of them, and the room's job is to put its advice back into that document.

What it cost us was the lesson the room taught: bring a document here and it can vanish without a trace. So people brought nothing, and decisions moved to smaller, warmer rooms (a chat, a hallway, a DM) where nothing gets recorded. Nobody sets out to route around the forum. The forum teaches them how.

The follow-up reply

My first draft of a reply on one of our wide threads came out of a model. It read beautifully, which was the problem: it committed me to a compliance review nobody had asked for, cited a regulation nobody had mentioned, and named gaps nobody in the room had raised.

None of it came from the session, and none of it could be checked from the room: verifying a control means looking at systems, not reading a thread. It arrived as commitments in my name anyway.

I caught it before it went out because I know our environment well enough to flinch. Without that flinch you'd send it, and a compliance-shaped workstream with no decision behind it gets planned, estimated and stitched into a roadmap.

That's the second way a forum fails, and the one I fear more: the forum's authority leaving the room attached to text the room never agreed to. It read like output, and output gets copied: one paragraph lands in a brief, a deck and a ticket, and it's a fact in three documents before anyone checks its source.

You're probably thinking: isn't the fix a review board, the gate we all hate, with fewer chairs, and how can three people see a whole estate? It isn't a gate, because the room approves nothing: every item ends one of four ways, and none of them is wait. And it doesn't need to see the estate, because everyone else reads the document async and the write-back puts the result where anyone can open it.

So here's what you do, in the order a session runs: seat the room, ask the four questions, end on one of the four endings, and write the advice back.

The three seats

Three seats, and that's where I'd start. Get the guest list wrong and you're running theatre, not a review.

The author. They're the customer, and if they leave with objections and no decision, you've handed them homework, not a review.

Whoever owns the neighbouring system. If the decision moves their queue, schema, budget or on-call rotation, they belong in the room: an apology afterwards is a different meeting, and a worse one.

One practitioner running something comparable in production. Someone with the scars. A second architect with opinions belongs in a separate review, not in this room.

For every seat after those three, run one test on your own invite list: will this person feel this decision within a quarter? If not, they read the document async and comment, which helps more and costs less.

The four review questions

Ask these in order, every time, and you'll leave with a decision, not a critique.

What breaks if we do nothing? For a decision, ask this first: it turns critique into a demand. For an advice request, ask what the asker has already tried, because advice on an untried idea is guesswork. If nothing breaks and nothing's been tried, close the item.

If this room didn't exist, what would you do? You learn what the author intends, which is often fine, and which of your comments are load-bearing and which are decoration.

What's the smallest decision we can make today? Most items are three decisions of different urgency bundled into one document (a schema change, a vendor call, a rollback plan), so take the one that unblocks this week and park the rest.

What risk are we taking, and who owns it? Say it aloud and write it into the document: a named risk is a decision, and a risk that only exists in the room is a rumour.

The four endings

A session isn't over until the author holds one of these four, and you're the one who checks.

A decision or advice, written into their document. Minutes and recap emails drift; the next engineer reads the artifact.

A default plus a deadline. If nobody decides by the date, the default happens: silence becomes a decision, and the author can act on it next morning.

A named escalation: a person, a question and a date. It's complete when the session ends: it leaves with a route and an owner.

The fourth, and cheapest, is nothing needed: say so, close the item, and write that line into the document.

"Come back next sprint" is none of them. It's how a room declines to decide while still sounding reasonable.

How an item moves through a review session An item arrives, the convener checks it, the session asks four questions and decides what is needed. Every outcome is written back the same day, checked again, then the item closes or stays open. INTAKE Item arrives with a link and a date CHECK Convener check right people, right document? SESSION Session four questions, in order CHECK Does this need anything from the room? CHECK Can the room decide today? NOT YET Back to the author with the gap named not yet NOTHING NEEDED Nothing needed written into the document ADVICE Advice written back into the document YES Decision recorded as an ADR ANOTHER OWNER Named escalation person, question, date PARK IT Parked owner, date, default parked twice, decided next session, even when the answer is no. Otherwise the default fires on the date WRITE Write-back, same day every outcome lands in the document CHECK Convener check does the document reflect the call? NO Item stays open the author fixes the document CLOSE Item closed

The parking lot rules

Three rules stop a parked item going invisible.

Every parked item leaves with an owner, a date, and a default that fires when the date passes, because a park with no default is a no that nobody has to own.

Read the lot aloud at the top of every session (all of it, including the items that look dead). What's that cost you? Two minutes! A forgotten item can't hide once its name is said aloud with its author in the room.

Parked twice means decided now, even when the answer is no. Especially then: a no is still a decision, so the author stops waiting.

The write-back rules

The forum's output isn't a decision until it's an edit, so the rules for those edits are narrow.

Advice that stays in the room doesn't exist, and advice that leaves as prose stays advice. Edits to the documents already in front of the room are what give it standing. A paragraph nobody in the room agreed to has none, however plausible it reads.

A decision becomes a decision record. In the author's repo, in the author's voice, living with the code of the service it decides.

A caveat becomes an inline comment. Next to the paragraph it changes, so it can't outlive the paragraph.

"You should also consider X" becomes a bullet in open questions. It was never a decision and it mustn't look like one.

[ IMPORTANT ]

Two tests decide whether advice gets written back: someone in the room said it, by name, and it lands in the author's own document. If it fails either, it has no standing.

Who does the write-back? The author, same day if possible, while it's fresh. Everyone else leaves the document alone, including the convener, whoever maintains the list and the loudest opinion in the room: the document is theirs, and so is the commitment.

Two rules held up best. Only advice from people who spoke in the session gets written back, by name, so hearsay never becomes a decision. And an accepted decision record is never edited: the next change is a new record, and the old one keeps its date.

Then the convener reads the document before the item closes: if it doesn't reflect the call, you're not done. That check is the whole quality gate (one person, reading one document), and the last thing between the room and a record nobody trusts.

The one-page sheet

This is what I hand people for an item that arrives as a decision, and it fits on a page!

Intake: what / who is asking / what decision they need / by when / link
Four questions: what breaks if we do nothing · if this room did not exist, what would you do · what is the smallest decision we can make today · what risk are we taking, and who owns it
Four endings: decide · default plus deadline · named escalation (person, question, date) · nothing needed
Write-back: what was advised, who spoke, what was NOT decided (a decision becomes an ADR)

The intake line keeps any item without a link the room can open off your agenda, and a repo the room can read counts as open.

For decisions that are expensive to undo, the sheet gets one more question: what does this cost to change later.

The closing check

Both failures come from one mistake: treating the room, not the documents people carried in, as the deliverable.

Part 1 was the vocabulary: seven document types, one question each. This part is the room where they get argued over.

So here's the check I run at the end of every session: name what changed, and in which document. If something changed that nobody agreed to, take it out before the item closes. If nothing changed, write one of the four endings into the document before anyone leaves.

What you can do this week

Two things, and neither needs permission.

  1. Audit your intake list for the state gap, not view by view: anything scheduled for a past date, and anything in triage that never got scheduled, is invisible in both. Put them on the next agenda by name.
  2. Take the parking lot out of your head and into a table with four columns (item, owner, date, default), then read it aloud at the top of your next session. Ask for the first item's default and watch how fast the room picks a date.

A forum earns its place by what leaves the room, not by what it approves.

Next up, Part 3: making this work visible. A decision recorded in our own documents is still invisible to the people who weren't in the room, and a function that's one person and one recurring meeting doesn't scale by meeting more.

Alexandre Castro

Principal Software Engineer, AI Platform, at Chamberlain Group. 15+ years shipping systems that handle real money.

Connect on LinkedIn