All field notes
    Field notes

    Day 3 of 5: review

    Day 3 of 5. Review.

    The model that wrote the code doesn't get to mark its own homework. Every plan is reviewed before any code is written and every change is reviewed after, both times by a model from a different lab to the one that did the work. Different training, different blind spots. And it's every change, including the dull ones.

    One catch from the last fortnight, with every test green.

    Our own merge tool would have accepted a passing test run from someone else's copy of the code, or one weeks old, as proof that main was fine. The reviewer found it by asking where that evidence came from.

    What keeps review useful is one rule. A finding blocks the merge only if real data or a normal user action can reach it. Anything else goes in as a written follow-up so it doesn't get lost.

    AI review raises the floor a lot. The diff still gets read before anything merges.

    Tomorrow, verification, and why green checks aren't the finish line.

    If you've put a second model on code review, what was the first thing it caught that the authoring model missed?

    No model marks its own homework. The plan is reviewed before any code: By a model from a different lab. Different training, different blind spots. Every change is reviewed after: Including the dull ones. Ask where the evidence came from: That question caught our merge tool trusting a weeks-old test run from someone else's copy of the code. Block only what real data can reach: Everything else goes in as a written follow-up. The diff still gets read: Before anything merges. The catch came with every test green. The question was where the evidence came from.
    Download card (PNG)