A 267-word specification preamble cut security defects in all five frontier models.

Pre-registered study, 50 realistic backend tasks covering money arithmetic, time handling, idempotency and access control, run bare and again behind a fixed filled specification frame. Defects fell in every model by 0.16 to 0.70 findings per task, every Holm-adjusted sign test significant, winning 95 of the 100 differing pairs. Bandit medium-or-high issues dropped from 53 to 11. All 500 outputs, prompts, AST checkers and the pre-registration are published with a DOI, so you can re-derive it rather than trust it. The interesting contrast is that this worked where instruction files had not. (arXiv 2609.23270)