Lesson

Verify by attacking the gate

Before trusting a gate, make it fail: tamper a byte, skip a stage, forge the evidence, and watch it refuse.

The incident

A tag-hygiene change to a retrieval router scored 15 of 15 on the set it had been tuned on. Blind, on a set it had not seen, it scored 15 of 23. The gate that certified it had never been asked to fail. The ranking change built on top of it was withdrawn.

The same week, a promotion gate for model checkpoints was checked by feeding it a checkpoint with one byte flipped, a run with the contamination stage skipped, and a forged evidence file. It refused all three. That is what a passing gate looks like.

The rule, and how it is enforced now

Every gate ships with its attack tests in the same suite: mutation tests on the controls, a known-bad input that must be rejected, and a check that the evidence file cannot be forged without detection. A sub-agent that reports “all green” is asked to show the red it produced.