Results, findings & evidence

The verdict, the matrix, AI-drafted findings with evidence, and your free re-run.

Results are layered so the first thing you read is the conclusion and everything under it is the proof. Each layer cites the one below; nothing asks to be taken on faith.

The layers, top to bottom

layer	what it tells you
Verdict	One plain sentence about what agents can and cannot do with your product, with severity counts.
Executive summary	One reviewed paragraph plus the 1–3 moves that turn the matrix green. AI-drafted, human-reviewed, evidence-cited.
The matrix	Every scenario × agent cell: pass rate over 5 runs with a 95% Wilson interval, labeled solid / flaky / broken. Click a cell for the runs behind it.
Findings	The diagnosed issues — each severity-ranked, evidence-linked, with repro steps and a suggested fix.
Per-run detail	Timelines and trace replay for every individual run; divergences between agents highlighted.
Evidence	Downloadable artifacts: traces, wire logs, timelines, the report, the plan. Every claim above resolves to something here.

Findings — AI-drafted, human-reviewed, evidence-bound

Findings are drafted by a model from the recorded evidence, then reviewed by us before you see them: your page shows published findings plus an honest count of any still in review. Three properties hold for every finding:

Labeled AI-drafted — always, on every surface that shows them.
Evidence-linked — each finding cites specific events in specific runs; the evidence chips jump you into the trace replay at the exact moment.
Honest about confidence — a finding without resolvable evidence is demoted to a hypothesis and displayed as one, never dressed up as a fact.

Flaky is a first-class verdict

A scenario that passes 3 of 5 runs is not a 60% success story — it's unreliable behaviour your users will hit. Cells that both pass and fail are labeled flaky and diagnosed as such; cross-agent divergences (works in Claude Code, fails in Cursor) are called out separately, because they usually indicate ambiguity in your interface rather than a bug in one client.

Deliverables

Alongside the page itself: report.md (the full written report), summary.json (machine-readable totals and cells), plan.json (exactly what ran), per-run traces and timelines, and the live event log. All downloads are credential-scrubbed before they reach storage (how).

The free re-run

Every audit includes one free re-run of its failing scenarios — after you've shipped fixes, run exactly the failed subset again from the results page, or let your coding agent do it through the Fix Loop. Beyond the free re-run, monitoring re-runs the plan on a schedule.