A clean-room reviewer finds different bugs

This note is about a piece of IntrinsicOS: the review gate every autonomous code change in the vault passes through before it lands.

When human code review stopped, autonomous code merged unreviewed for six days. The replacement: a fresh model session that sees only the diff and the task.

How it works

The reviewer is not a fork of the author's session — a fork would inherit the author's framing and blind spots, precisely what the gate exists to escape.

No inputs. Everything the reviewer sees is assembled by the scheduler: the task, its acceptance criteria, and the diff. The author cannot pass it anything. The interface is deliberately hostile to steering.

Permissions deny writing, browsing, and messaging. A reviewer cannot be argued with or message the author. It can read, run commands, and replay failures.

Fail closed on landing, open on delivery. A blocking finding forces the branch to PR instead of auto-landing. But the branch always ships—either merged or as a pull request carrying the findings.

The scheduler validates the review by hashing the branch's final diff and comparing it to the report. Stale reports and changed diffs trigger a re-review. Across 55 checked reviews, 10 required re-validation: seven because the diff changed after review, three because no review was requested at all.

What it found

Two different configurations, same target, 2026-08-12. A deep multi-agent review at medium reasoning: 8.5 minutes, roughly 121,000 tokens. A single clean-room session at low reasoning: about 90 seconds. The two runs found different bugs.

Coverage is stochastic. This is the entire justification for the layered arrangement: a cheap per-diff gate on every branch, plus a slower adversarial pass twice a week.

The live numbers, 2026-08-12 to 2026-09-03: 51 branches actually reviewed. Results: 43 clean, 2 high-severity, 5 medium, 5 low. The fix-and-re-review loop fired 10 times (17%). Both high-severity findings hit the same day, same project, same author.

(Seven of 58 reports had no diff to review—nothing ran, nothing's reviewed. Corrected figure: 43 of 51 actually reviewed.)

Where this goes next

There's no false-positive measurement—nothing records whether a finding was right or wrong. A one-word disposition when the author acts on a finding would let the gate report not just how often it blocks but how often it was right to. That number decides whether a cheap per-diff reviewer belongs in front of everything, or whether slower deep passes should carry more load.