Can disagreement in reasoning traces signal legitimate value conflicts?

Multi-agent systems almost universally treat disagreement as a defect to drive out — via voting, extra debate rounds, or fault-tolerant aggregation. That objective is right for instrumental tasks where disagreement signals noise, and wrong for value-laden ones where disagreement is a stable property of the problem. The move here is to abstract reasoning traces plus binary decisions into a symbolic layer: crossing reasoning similarity with conclusion agreement yields four states — convergent agreement, divergent agreement, convergent disagreement, divergent disagreement — that drive defeasible routing rules. The key state is convergent disagreement: agents share a description of the case yet differ in value prioritization. In a factual task that looks like inconsistency to resolve; in a normative one (content moderation is the paradigm case) it more plausibly marks a legitimately contested situation, and collapsing it into one automatic decision hides the contest. Escalation becomes a rational meta-action under normative uncertainty, not an automation failure.

This is the constructive counterpart to several failure-mode notes. Why do multi-agent LLM systems converge without genuine deliberation? shows consensus-seeking actively destroying information; trace-disagreement routing is what you'd build to use that information instead of suppressing it. It sharpens When does debate actually improve reasoning accuracy? by supplying the missing discriminator — a vote cannot tell interpretive disagreement (agents read the case differently) from evaluative disagreement (agents weigh shared considerations differently), but joint trace+decision comparison can. And it answers the diagnosis in Does agent confidence actually signal competence in deliberation?: route on the structure of reasoning, not on confidence, and manufactured consensus has nothing to feed on.

The honest limit: the framework is argued and instantiated conceptually in moderation, not yet validated at scale, and reliably separating interpretive from evaluative disagreement from imperfect traces is itself hard — traces can be post-hoc rationalizations. But the reframing is the contribution: the design question shifts from "how do we make agents agree?" to "what does the shape of their disagreement imply about the next action?"