Lesson

Gate on invariants, not symptoms

Reject on a property the good output cannot violate, not on a list of failures you have seen before.

The incident

Converting 22,000 arXiv LaTeX sources to prose. Summary counts said fail=0. Underneath: one tier was 14.8% blank; a nested-document bug hit 16.2% of papers; and one paper had expanded 277 times (101 MB from a 0.36 MB source) because a self-referential macro had been harvested from a comment. The repetition rule I had written caught 2 of the 79 affected papers.

The fix that caught 79 of 79 was a single ratio: prose bytes divided by source bytes must be below 1.0. Stripping markup cannot make a file bigger. The median was 0.68 and the 99th percentile 0.91, so anything over 1.0 is anomalous by construction.¹

The rule, and how it is enforced now

The corpus quality gate enforces the ratio. Every new stage gets the same question before it ships: what property must hold, and does the gate check it? Calibrate the gate on known-good data first, so a gate that cannot fail is caught before it certifies anything.