royvergara Critique Synthesis
You are killing the loudest voice problem. The output is a decision with reasoning, never a neutral summary that pushes the choosing back onto whoever reads it.
Separate constraints from opinions before weighing anything
Some feedback is not a vote — it is a claim about the world: "legal requires the disclosure above the fold," "this must match the pattern shipping in Q3," "that component is deprecated." Constraints do not compete with user evidence; they bound the decision space the evidence gets weighed inside. Pull them out first and verify rather than weigh: a confirmed constraint frames the decision, a claimed-but-unverifiable one gets named as an open verification with an owner, and a preference dressed as a constraint ("we can't ship anything this dense" from taste, not fact) goes back in the pile as opinion. Filing a real constraint under dissents is how a synthesis gets ignored by everyone who knows the constraint is real — and remember that a senior reviewer's flat objection is sometimes a constraint with the claim left unstated: when a comment smells like context rather than taste, ask what is behind it before you rank it.
Cluster, then weigh
Cluster the remaining feedback into distinct issues. Then weigh each issue by signal strength, in this order: direct user evidence beats everything on questions user behavior can answer, independent convergence (multiple reviewers raising the same issue separately) beats a single voice, domain relevance (the accessibility expert on an accessibility issue) beats general seniority.
Weigh user evidence by its quality, not just its type: sample size, task realism, and whether the finding could be an artifact of prototype fidelity (users "missing" a control that was low-fidelity placeholder is a fidelity artifact, not a finding). Thin user evidence — a two-person hallway test — still outranks pure opinion on a usability question, but name its thinness in the ranking so the decision carries its real confidence. And user evidence only trumps on questions it can answer: task success, findability, comprehension. On brand, strategic fit, or portfolio coherence it is one input, not the trump card.
Explicitly do not weight: volume, repetition by one person, seniority on issues outside the person's domain, or how recently the comment was made.
Produce four sections
1. The decision. What changes and what stays, stated plainly. When two signals of genuinely equal strength conflict, the decision is the specific test that settles them — named as precisely as any other decision: what to run, with whom, and what result decides it. That is still a decision; a neutral summary is not.
2. Ranked issues. Each issue with the strongest signal behind it named: who or what the signal is, and why it outranked the others.
3. Dissents. The feedback that lost, recorded with the reason it lost. Dissent that vanishes from the record comes back in the next review twice as loud.
4. Reopen conditions. For each major call, the evidence that would legitimately reopen it. This is what makes the decision a decision instead of a mood.
If a design-os.work/<slug>.yaml ledger is present, record the decision to decision.reviews.critique, and record each direction that lost as an entry in decision.kills[] — died_at: critique-not-selected, the reason it lost from the ranking, never a bare "not chosen," and a cost estimate if one is in hand (see templates/work-ledger.schema.md). No ledger changes nothing about the synthesis above.
Quality bar
A reader who disagrees with the decision should still be able to see exactly which signal beat their position, and exactly what evidence would change the outcome.