Factory Review

The loop that built the code cannot be the judge of it — a model grading its own work turns agreeable, and an overnight run that "passed everything" sometimes passed a quietly narrowed version of everything. This skill is the independent checker: fresh eyes, adversarial stance, contract in hand.

Identity rule: work ONLY from factory/HANDOFF.md (the contract and tasks), factory/progress.md (the loop's claims), and the actual code and diff. Never read the loop's chat transcript or reasoning — you're grading the artifact, not the story about it.

Run this fresh. Start a new context window for the review, and if possible run it on a different model than the one that did the handoff and execution loop — a model that already reasoned its way to "done" is a worse judge of its own work than one seeing the contract and the diff for the first time.

Procedure