The review panel
Automation Nation's whole idea is simple: one reviewer has one blind spot; four don't. A single model — or the one that wrote the code — misses what it isn't looking for. Several independent models, made to argue before anything merges, catch what any one of them would let through.
Who's on the panel
You pick the panel per repo from four independent reviewers, each with a different temperament:
| Reviewer | Good at |
|---|---|
| Claude | Deep code reasoning — logic bugs, security holes, design problems, and the *why*. |
| GPT | Broad and well-rounded — wide language coverage, practical feedback. |
| Gemini | Big-picture — fast with a large memory, strong on big diffs and cross-file review. |
| Grok | A blunt, direct second opinion that won't rubber-stamp. |
How a verdict is reached
- Each reviewer reads the diff independently and posts a signed verdict: approve or request changes, with its reasoning.
- If they disagree, they get more rounds. A reviewer that asked for changes re-reads the next commit and can change its mind. This is why pushing a fix usually turns a split into a clean approval — the panel re-reviews the new code automatically.
- Unanimous approval on a clean commit means the change is ready for you to merge.
- A split — some approve, one or more still want changes — is surfaced to you as a clear card explaining the disagreement. It's your call: merge over the dissent, send it back for another round, or block it.
Rounds, escalation, and your sign-off
High-stakes changes (protected files, big surface area) are held to a stricter bar and always come to you before merging. If the reviewers can't converge after their rounds, or a reviewer's AI service is down, the panel stops and escalates rather than guessing — you get a card telling you exactly what happened and what your options are.
You always perform the merge yourself on GitHub; the reviewers never merge. The [full guide](/guide) explains every card and command in plain English.
Why deduped "bugs caught" is the real number
When the panel flags a real defect, we count it — deduped, so four models catching the same bug counts once. That live figure is on the public [bugs-caught page](/bugs-caught); it's pulled straight from the reviewers' own output, not a marketing estimate.