At the end of a run, one block sits under the answer. Read it before the answer — it tells you how much to trust what you are about to read. It is the same on the Mac, the web and the iPhone.
Row by row
- Outcome. Every reviewer agreed, a majority did, or there was no agreement. A seat that could not answer or said neither yes nor no never counts as agreement.
- Routed. When the app itself decided how to run your question — sent it to one model because it was a short question, answered directly because nothing was built or changed, or brought in reviewers because files changed — this row says so and why.
- Checked. How many reviewers there were, how many commands they ran, and how many passed. A command that failed as it was meant to — a probe that found nothing listening, or the failure you asked to see — is counted as failed as expected, not as a failure.
- Evidence. One row for every check that ran, folded to a single line with a small green arrow beside it; click the arrow to open it. The line says who ran the checks (for example "5 checks — Claude ran 4, ChatGPT 4, Grok 1") and the commit and machine they ran on. Each row says what the check was for, who ran it — "Claude, rerun by ChatGPT" when a reviewer repeated it — and, beneath, the exact command, its exit code and the line of output that matters. Its state is one of Checked, Reproduced (it failed, and the failure is real), Blocked by missing setup (a tool, dependency, key or service was missing), Untested hypothesis, or Ran, result not known. The exit codes are measured by the app, not taken from a model's account, and a command a seat's own tools refused is shown as refused, never as checked.
- Gaps. What nobody verified, shown even when everyone agreed: each model's own "Not verified:" lines, setup that was missing, and a verdict given without checking anything. The same gap said by several models appears once. When nothing reached a deployed version, a quiet line underneath says the work was checked on this machine only.
The run tree
The Summary panel also holds the run as a tree: each round, the seats' turns in it (role, model, time, money), each turn's tool calls with what they came back with, a refusal marked, and the discussions a turn delegated to. It is built from what the run recorded — the receipt's rows and the tool rows — never from a seat's own account of what it did, and it is folded until you open it, on the web, the Mac and the phone. When a run failed, the Summary's Why row may end with a likely cause: one sentence the cheapest seat wrote afterwards from the failure record and the last few tool calls, marked as its reading, never as the fact.