Everything tagged this way, in Developer.
A review from a different model family caught bugs my passing tests missed, cost 20 extra minutes on four real bugs where it changed nothing, and once let three HIGH defects ship because I read repeated findings as a pass. This is how the review is wired, what the trials showed, and how to set one up.
Every rule I left to an agent's judgment was skipped sooner or later: a stamp-at-the-end reminder that was broken in every one of its 32 failures, a 650,000-token ceiling a run overspent to 1,444,564, a judge that opened 20 of 48 screenshots. This is how I moved those rules into scripts, hooks and exit codes, how I prove each gate can fail, and how you can write one.
On four real backlog items scored by tests the agents could not see, plain Opus passed every hidden test my tuned /work process passed, in 7 minutes and $2.46 against 53 minutes and $8.92, in notional API dollars. This is how I set up the trial so the agents could not game it, what it showed, and why I retired /work.
Yotam Fromm
