architecture-review-board ยท diff
git:20260718.ea9bc00 to git:20260728.fac4544
29 added, 28 removed. Audit A to A.
---
name: architecture-review-board
description: Run an Architecture Review Board that filters proposals by blast radius, reviews them through fixed lenses, issues binding decisions recorded as ADRs, and honors an appeals path. Use when a technical choice crosses team boundaries and needs a durable, accountable decision instead of hallway consensus.
---
# Architecture review board
An ARB exists to make the expensive, hard-to-reverse decisions once, with the
right people present, and to write down why. Its failure modes are a matched
pair: too loose and it stamps whatever arrives, too heavy and it becomes a
tollbooth every team learns to route around. A board worth keeping reviews only
what matters, decides with real authority, and leaves a record the next
architect can trust.
## Method
- 1. **Set a submission bar keyed to blast radius.** Require review when a proposal
- adds a new datastore or language, crosses team boundaries, changes a shared
- contract, or opens a security or compliance surface. Apply the one-way-door
- test: reversible, two-way-door choices stay with the team. Everything else is
- noise the board should decline at intake.
- 2. **Demand a written proposal, not a calendar hold.** The submission is a design
- doc or ADR draft stating the problem, the options weighed, the recommendation,
- and the trade-off accepted. A board that reviews slides reviews charisma.
- Reject anything with a single option: one option is a decision already made,
- dressed as a question.
- 3. **Run a shepherd pre-review before spending the board.** One reviewer reads the
- proposal cold, confirms it clears the bar and answers the obvious questions,
- and either advances it or returns it. This keeps the full board out of triage
- and the author out of a public ambush over a missing section.
+ 1. **Set a submission bar keyed to blast radius.** Require review when a
+ proposal adds a new datastore or language, crosses team boundaries, changes a
+ shared contract, or opens a security or compliance surface. Apply the
+ one-way-door test: reversible, two-way-door choices stay with the team.
+ Everything else is noise the board should decline at intake.
+ 2. **Demand a written proposal, not a calendar hold.** The submission is a
+ design doc or ADR draft stating the problem, the options weighed, the
+ recommendation, and the trade-off accepted. A board that reviews slides
+ reviews charisma. Reject anything with a single option: one option is a
+ decision already made, dressed as a question.
+ 3. **Run a shepherd pre-review before spending the board.** One reviewer reads
+ the proposal cold, confirms it clears the bar and answers the obvious
+ questions, and either advances it or returns it. This keeps the full board
+ out of triage and the author out of a public ambush over a missing section.
4. **Judge through fixed lenses so outcomes stay consistent.** Score every
proposal on reversibility, build-versus-buy, operability and on-call cost,
security and data handling, total cost of ownership, and fit with the stated
tech strategy. Named lenses stop the decision from turning on whoever spoke
loudest that afternoon.
5. **Issue a binding decision in one of four shapes.** Approve, approve with
conditions, reject with reasons, or defer pending named information. "We have
concerns" is not a decision. Record the outcome as a dated ADR with the
deciders, the rationale, and the constraints, and link it from the
repositories it governs.
- 6. **Publish an appeals path and honor it.** An author who disagrees escalates to
- a named authority, a principal engineer or a CTO-level review, inside a set
- window. A board with no appeal breeds quiet workarounds; a real appeal is what
- makes the board's authority legitimate rather than resented.
- 7. **Time-box the loop and measure your own latency.** Commit to a decision within
- a fixed number of working days of a complete submission, and track that
- turnaround like any other target. Nothing teaches teams to stop consulting the
- ARB faster than making the ARB the slowest step in the project.
+ 6. **Publish an appeals path and honor it.** An author who disagrees escalates
+ to a named authority, a principal engineer or a CTO-level review, inside a
+ set window. A board with no appeal breeds quiet workarounds; a real appeal
+ is what makes the board's authority legitimate rather than resented.
+ 7. **Time-box the loop and measure your own latency.** Commit to a
+ decision within a fixed number of working days of a complete
+ submission, and track that turnaround like any other target. Nothing
+ teaches teams to stop consulting the ARB faster than making the ARB the
+ slowest step in the project.
## Checks
- Would a reversible, single-team choice be turned away at intake, or does the
board review whatever lands on it?
- Does every decision exist as a dated ADR with rationale, not a verbal yes in
someone's memory?
- Can an author who disagrees escalate through a named, bounded appeals path?
- - Is the board's own decision latency measured, or is it quietly the slowest step
- in every project?
+ - Is the board's own decision latency measured, or is it quietly the slowest
+ step in every project?
## Boundaries
The ARB decides architecture that crosses boundaries or resists reversal; it
does not review routine, team-local design, which belongs to normal code and
- design review. It rules on the proposal, not on implementation quality or release
- readiness. The depth of any one proposal is the design-doc skill's job, and where
- your organization already runs an RFC or ADR convention, adopt its template
- rather than standing up a parallel one.
+ design review. It rules on the proposal, not on implementation quality or
+ release readiness. The depth of any one proposal is the design-doc skill's
+ job, and where your organization already runs an RFC or ADR convention,
+ adopt its template rather than standing up a parallel one.