What You Need to Know
An architectural decision is a choice about how a system is built. A trade-off is what that choice gives up to get what it gains. The owner, or accountable stakeholder, carries the consequence. The sponsor funds the work.
Take a chimney sweep firm adding Claude to answer booking enquiries and draft the report a sweep leaves after each visit. The owner is the sponsor, a developer builds the system, and the firm's insurer cares what the reports say.
The Complete Communication Has Seven Parts
A complete communication lets somebody else evaluate a decision instead of merely accepting it. It has seven parts:
- The decision.
- The options genuinely considered, including the one the audience expected.
- The criteria they were judged on, so the weighting is visible and arguable.
- The evidence. For a Claude system this usually means Domain 4 measurements, not assertions.
- What each option costs and exposes.
- A recommendation. Declining to recommend is abdication, not neutrality.
- What would change the recommendation: the measurement, the volume, the price change or the regulatory event that would flip it.
The seventh part separates an architect from an advocate. It turns a position into engineering that can be revisited without re-arguing the whole question. A decision record carries the same content in 6.4.
At the sweep firm, the decision might be that a sweep checks each draft report. The options include the one the owner expected: no check. The evidence is the measured error rate on past reports. The revisit condition names the rate that would drop the check.
Translation Changes the Lead, Never the Facts
Every room gets the same decision and the same facts. Only the lead and the level of mechanism change.
- Executives lead on business consequence and risk.
- Engineers lead on mechanism and constraint, because that is what they build against.
- Legal and compliance lead on the obligation and what residual risk survives the controls.
- The sponsor leads on cost, date and precisely what they are being asked to accept.
At the sweep firm, the owner, the developer and the insurer each hear the same error rate, led differently.
Two failures mirror each other: burying an executive in mechanism, and handing a specialist generalities they cannot act on. A third is worse. Simplifying past the point of truth is not translation, and the person who trusted it always discovers it later.
Say the Error Rate Out Loud
These systems have error rates, and the architect owns saying so before anyone signs. State the rate and the population it was measured on. State the handling path for the residual errors: who catches them, how quickly, at what cost. At the sweep firm, the sweep reading each draft is that path.
This makes a demonstration a communication risk rather than a communication tool. A demo is a sample from the favourable tail. An audience that has watched five correct answers has been shown a determinism that does not exist.
Uncertainty travels as a range, with the work that would narrow it. Anthropic's evaluation research supplies the register: report the interval alongside each score, not the score alone. When two candidates are compared, give their pairwise differences intervals too. Without them, a comparison may be presenting noise as a finding.
The Disagreement Pattern, and Its Exception
Sometimes an accountable stakeholder wants what the evidence contradicts. Measure, frame, recommend, and leave the decision where the accountability sits. The owner gets the quality difference as a number and its downstream cost, in the currency the owner controls. The owner also gets the exposure to the organisation, the saving, a clear recommendation, and then the decision.
Say the firm's owner wants booking replies live before the autumn rush, though some quotes come out wrong. The owner gets that rate, its cost and what the date saves, then decides. Three postures fail: silent non-compliance, compliance without comment, and escalation over the owner's head as an opening move.
The exception is different in kind. An owner's authority covers cost against quality and scope against date, because the owner can accept the consequence. It does not cover a legal, regulatory or safety obligation, which is nobody's preference to accept. There, say so plainly and bring in the function that owns the obligation, which is Domain 5's territory. Reports telling customers a fire is safe to light, with no sweep checking, are an unsupervised action with a safety consequence.
Read the Consequence, Not the Volume
Which case you face, in a project or a stem, is decided by what being wrong does. A margin, a date or a quality level says the owner decides with the full picture. An unlawful data path, or an unsupervised action with a safety consequence, says obligation. The loudness of the stakeholder is not evidence either way.
In the exam
Items on this objective reward completeness of reasoning rather than eloquence. Both translation failures appear as distractors.
Anthropic publishes nothing about how to brief a sponsor, so treat this objective as professional craft rather than product behaviour. Its authority is the exam guide's own description of the role: collaborating with cross-functional stakeholders and communicating architectural decisions effectively. The role's rooms include security, legal and executive participants.
Exam Traps
Implementing a decision you hold measured evidence against, without comment
Capitulation. The measurement existed and the accountable owner decided without it, which is exactly the advisory failure the credential is meant to exclude.
Building it your way quietly after the argument went the other way
Evasion, with or without a budget line attached. It removes the owner's ability to accept a consequence that is legitimately theirs, and when it surfaces it costs the trust that made the advice worth having.
Taking the disagreement to the owner's superior before the trade-off has been put to the owner
Obstruction as an opening move. It spends organisational capital on an argument the evidence had not yet lost. Escalation is the last move, and only where the owner's authority does not cover the consequence.
Answering 'how accurate will it be?' with a single confident number
A point estimate with no interval and no population is a promise nobody measured. The keyed option gives a range on a defined population plus the measurement that would narrow it.
Practice Scenario
A logistics operator's operations director has decided the route-exception assistant will ship without its confidence-banding and human-review step, because the review queue needs headcount she does not have. Evaluation on real exceptions puts most of the errors inside the band that step would have caught, and each undetected error triggers a re-delivery costing more than the saved headcount. She has not seen these results. Company policy requires a risk owner to see the evaluation evidence before accepting a risk, and staffing decisions stay with her. What should the architect do?
Build Exercise
Write One Architectural Decision as It Would Be Communicated
What you'll learn
- How to assemble a decision, its options, criteria, evidence and consequences into one communication
- How to write the revisit condition that keeps a decision from being re-argued from scratch
- How to translate for an executive and an engineer without changing the facts
- How to state uncertainty as a range with the measurement that would narrow it
- How to tell an owner-decides trade-off from an obligation the owner cannot accept
- Take the hardest trade-off in a system you know and write the decision, the options genuinely considered, the criteria they were judged on, the evidence, and what each option costs and exposes
Why: Completeness is what lets somebody else evaluate the decision rather than merely defer to it, and it is what the exam rewards over eloquence.
You should see: One page carrying five parts, with the evidence traceable to a measurement rather than an assertion.
- Write the recommendation, then the sentence most architects omit: what would change this recommendation, stated as a measurement, a volume, a price or a regulatory event
Why: A recommendation with no revisit condition has to be re-argued from scratch the next time anything moves.
You should see: A named recommendation plus one falsifiable condition that would reverse it.
- Produce the executive version and the engineering version of the same decision, then read them against each other
Why: Translation changes the lead and the mechanism; distortion changes the facts, and it is discovered later by whoever trusted it.
You should see: Two versions with different leads and identical facts.
- Write the accuracy answer as a range on a defined population, with the work that would narrow it, and note what you would need to measure to earn a single number instead
Why: A point estimate with no interval is a promise nobody measured, and it becomes an incident after launch rather than a conversation before it.
You should see: A band, the population it applies to, and a named piece of evaluation work with a duration.
- Classify the decision: owner-decides trade-off, or an obligation the owner has no authority to accept — and write the two sentences you would say in the second case, naming the function you would bring in
Why: The same evidence supports opposite moves, and the consequence of being wrong is what tells you which one.
You should see: A one-word classification with its justification, plus the two sentences and a named function.
Sources
- A Statistical Approach to Model Evaluations — Anthropic
- Building Effective AI Agents — Anthropic
- Claude Certified Architect Professional Exam Guide (v1.0) — Domain 6 — Anthropic