Five reviewers. One honest verdict.

Six steps, no trust required.

1. Scope it. Point it at a diff, some files or a folder. A dry run bundles the code and prints the worst-case cost. Nothing is sent.
2. Check the coverage. A file counted as bundled can still arrive cut short. The run records truncation and stops until you accept it.
3. Send it out. Four model families, in parallel, each returning findings in the same JSON format. Your own agent and any specialist sub-reviewers add theirs.
4. Prove they ran. A reviewer that crashes can leave an empty list, which looks exactly like “found nothing”. The summary file lists who answered, the tokens used and the cost, and the run fails loudly if any reviewer did not.
5. Count the agreement. “4 of 5 agree” is strong. A lone reviewer is a FLAG: check it against a real source before you act.
6. Decide. The review is read-only. It reports; you or your agent fix, as a separate step.

What it refuses to pretend.

• Empty is not clean. An error stub and a real “nothing found” are told apart by the summary, never by the file.
• Starved is not broken. A reviewer cut off at its output limit gets a retry with a bigger cap, not a place on the bench.
• Counted is not covered. Bundled files are checked for cut-off content, and you can require that the symbol you changed is in the bundle.
• Lone voices get verified. Models hallucinate from partial context. Single-reviewer findings are flags until checked.
• Read-only. It never edits the reviewed code, and it never writes its output inside your repository.
• Spend is capped first. A positive budget is required, and the run refuses to start if the worst case would exceed it.
• Free means free. A seat you believed was free but that was billed is flagged.
• Once, aimed, at the gate. A cheap pass on every change; the full panel once, on the final change, before money, logins, tenant data, migrations or deploys ship.

What it sends, what it costs.