Where your branches disagree

Run the same question through three models and you get three answers that mostly say the same thing. The interesting part is the small share that does not, and that is the part a summary throws away first.

The Agreement and conflict panel is built for that share. Open the AI Summaries drawer in the Workbench and pick the "Agreement & conflict" tab. Conflicts appear at the top, and the shared themes you may already know sit underneath them.

Why agreement alone is misleading

Two branches agreeing is not the same as two branches being right. If two of three models make the same wrong call, agreement is exactly what a majority view will report, and the one branch that caught the problem is the minority that gets rounded away.

That is the reason conflict sits above agreement in the same panel rather than behind a tab of its own. Reading what your branches share, without also reading where they split, is the half of the picture that a single model could mostly have given you anyway.

The four things it reports

Contested. One branch states something, another states the opposite. Both sides are shown, per branch, so you can see who took which position rather than a merged verdict.

Conditional. Some branches state a claim flatly and one qualifies it: "flag this, unless the name is genuinely the user's". This is usually the most valuable line on the screen. A lone qualification against a confident majority is a strong signal that the majority skipped a check, and it is precisely the detail that disappears when answers are summarised.

Only one branch. A substantive point exactly one branch raised. Minority of one often means the most thorough branch, not the outlier, so these are surfaced rather than dropped for failing a majority threshold.

Different conditions. Two or more branches agree on the verdict and disagree about what would change it. "Flag it unless the name is real" and "flag it unless the profile is stale" reach the same conclusion by different routes, and the difference is the finding. Every branch's reasoning is shown in full here, because the conditions are the whole point.

What it will not do

It will not manufacture a disagreement to have something to show. When your branches genuinely agree on the substance, the panel says so and stops.

It will also never claim agreement when it could not complete the analysis. If the check fails, it tells you it failed. "No conflicts found" means the comparison ran and found none.

Reading it honestly

Each position is labelled with the model that produced it, because branches of one fork usually share a title and the model is what actually distinguishes them.

If more branches were open than the panel compared, it says how many it used. A comparison over six of nine branches is a different statement from a comparison over all nine, and you should not have to guess which one you are reading.

The panel needs at least two branches with content, and it waits for branches to finish streaming before it runs, so findings do not shift under you mid-read.