Merge guard

Consensus-by-merge is the classic failure of multi-agent deliberation. Three sharp drafts go in, and what comes out is the average — technically containing everyone's ideas, committed to none of them, and worse than any single input. It feels collaborative, which is why it survives review.

The test

A merge is valid only if you can complete this sentence:

The merged design beats draft A, draft B, and draft C on <named criterion>, because <specific reason>.

The criterion must be one of the criteria named in the brief before drafting. Not a new criterion invented to justify the merge.

If you cannot complete that sentence for every input draft, the merge is invalid. Pick one draft as-is and record why.

Invalid merges

These are the shapes to reject:

• "Takes the best of both" — with no criterion on which the combination wins. This is the average wearing a compliment.
• Additive scope — merged design does A's thing and B's thing. Usually this is two designs shipped together, not one better design.
• Conflict-avoidant — the merge exists because rejecting a draft felt harsh. Picking one draft is not a judgment on its author, and its author does not know which model wrote the others anyway.
• Criterion swap — A wins on latency, B wins on simplicity, so the merge is declared to win on "overall." "Overall" is not a criterion.

Valid merges

A real merge usually looks like: A's structure with B's failure handling, where B's failure handling closes a hole A has, on the reliability criterion that was named up front. One draft is the base; another contributes a specific, nameable improvement.

Recording

Whichever way it goes, the Decision Record states:

• Which draft won, or what the merge was
• The criterion that decided it
• What was rejected and why — this is the part people skip and the part that is worth reading later