They agreed, and the bug shipped anyway.
One developer put their whole review loop in four words: the model “arguing with” itself, “all the way.” One model writes the code, another copy of the same model reviews it, they agree, and the bug ships anyway.
Same model, same blind spots.
Same training, same habits, same blind spots. If the reviewer reads the author's explanation first, it also inherits the author's framing: it checks what the author says the change does, not what the change actually does.
Agreement between two copies of one model is the first opinion again.
What doesn't work.
Adding “be very critical” to the reviewer's instructions. The tone changes. The blind spots don't.
Several agents debating. Agreement among the same model isn't independence.
Counting how many reviewers approved. The count measures effort, not coverage.
What actually helps.
Most of this isn't us.
The reviewer reads the change and its tests only, not the author's write-up.
The reviewer must break something on purpose and watch a test fail, so the review finds what the tests don't cover.
Review on a different machine or operating system from the one it was written on.
For risky changes, bring in a person or a different model.
A blind read, on a different machine.
Our second reader reads the diff and the tests only, never the author's write-up, and plants failures of its own. It works on a different computer from the author's.
Reading on Windows rather than the author's machine, the blind read found a PATH entry that let a program skip its own trust check, and a check no test covered. Both were fixed before the change shipped.
We also traced every way our sessions coordinate. When two of them truly disagree, a person settles it, so that's where the design sends real conflicts.
| What was measured | What the record says |
|---|---|
| what the second reader sees | the diff and its tests, not the author's write-up |
| found on a different operating system | a trust-check bypass and an uncovered check, fixed before shipping |
| who settles a real disagreement between sessions | a person, by design |
| recorded | 3 August and 13 October 2026 |
| where it's from | Vinculum's own project record, which runs on Vinculum |
Where a shared memory fits.
Vinculum lets you control and record what a reviewer was shown, which makes the review checkable.
It's still the same model family underneath. Reading blind narrows the gap; it doesn't make the review independent.
Common questions.
Should the reviewer always be a different model?
For risky changes it helps. Changing what the reviewer sees, and making it break things, does more than changing the model alone.
Don't multi-agent debates catch more bugs?
They catch different wording of the same idea. Independence comes from different inputs and planted failures, not from more copies talking.