Disagreement is useful collective state only when positions, reasons, evidence, scope, and responsibility remain legible; shared data, models, incentives, or identity can still correlate errors.

A visible difference in conclusions becomes informative only when the collective can determine what is being disputed, which evidence and assumptions produced each position, where each claim applies, and who remains responsible for acting or revising. Otherwise disagreement may be noise, duplicated confusion, strategic avoidance, or stylistic variation. Even attributed positions may not provide independent search when participants share the same sources, base model, incentives, evaluators, or identity-forming context. This claim assumes that provenance and rationale can be preserved at tolerable cost and that reviewers can use them. It does not require equal weight for every objection, permanent disagreement, or rejection of consensus. Architecturally, it motivates exact-version agent positions, rationalized challenge links, maintained tensions, disclosed shared dependencies, and behavioral tests of error independence rather than inferring plurality from agent count or model names.