Reading the Conflict Heatmap: A Reviewer's Guide

The Conflict Heatmap is the part of Qonera people remember from a demo: an answer where every claim carries a color, Green, Orange, Red, or Outlier, showing how three independent models agreed or diverged on it. But a visualization is only as useful as the decisions it drives. This is the practical guide: what each tag actually tells a reviewer, and what to do when you see it.

One framing first. The heatmap exists because the Multi Model Stress Test runs three models on the same question and the same vetted evidence, without letting them see each other, and a judge synthesises the result. The colors are the judge showing its working at the claim level, instead of hiding the disagreement inside one smooth answer. The heatmap is not a verdict on truth. It is a map of where the answer is sturdy and where it deserves your attention.

Green: aligned, so verify efficiently

A Green claim means the models independently converged. That is a meaningful signal: three systems that could not coordinate reached the same conclusion from the same evidence. It is not proof. Models can agree because the evidence is genuinely strong, or because the underlying source they all leaned on is itself wrong. The right reviewer move on Green is efficient verification: spot-check the citation, confirm the source is current, and move on. Green is where review speeds up, not where it stops.

Orange: partial agreement, so read the difference

Orange marks partial agreement: the models overlap on the substance but diverge on scope, emphasis, or a qualifier. One says the clause permits termination; another says it permits termination with conditions. Orange claims are where nuance lives, and the reviewer’s job is to read what the difference actually is, because the qualifier one model added is often the caveat the client needs to hear. Do not resolve Orange by picking the smoother sentence. Resolve it by checking which version the evidence supports.

Red: disagreement, so go to the source

Red means the models genuinely conflict, and this is the heatmap’s highest-value signal. Independent disagreement on shared evidence almost always means one of three things: the evidence is ambiguous, the evidence is contradictory, or the question was underspecified. All three demand a human. The move on Red is to open the per-claim citations and read the underlying passages yourself, because whatever confused three models will confuse the client’s advisors too, and you want to be the one who found it first.

Outlier: one dissenter, so ask why

An Outlier tag means one model broke from the other two. The instinct is to side with the majority, and often the majority is right. But the outlier is sometimes the only model that noticed the footnote, the superseded version, the exception clause. Treat an Outlier as a question rather than a vote: what did the dissenting model see? A minute spent on that question is either cheap reassurance or the catch of the engagement.

The heatmap is triage, and triage is the point

Read this way, the heatmap solves the real problem of AI review, which was never “is the answer good” but “where should limited human attention go.” Greens get efficient checks, Oranges get a careful read, Reds and Outliers get the reviewer’s full expertise and a trip to the sources. Every claim links to its evidence, so each of those moves takes clicks, not archaeology. And the review itself, what was flagged, what the reviewer decided, lands in the audit trail through the review and approval workflow.

Teams sometimes ask whether a mostly-Green heatmap means the work is done. It means the opposite of alarming, but the discipline holds: the colors direct the review, the evidence settles the questions, and a named person still signs before anything ships. The heatmap does not replace judgment. It makes sure judgment is spent where the answer is actually uncertain, which is what separates review from rereading.