A 267-word specification preamble cut security defects in all five frontier models, where instruction files had failed

arXiv 2609.23270·high signalXBlueskyLinkedInCopy link

A pre-registered study ran 50 realistic backend tasks (money arithmetic, time handling, idempotency, access control) through five frontier models from five vendors, twice each: bare, and preceded by a fixed 267-word filled specification frame. The frame reduced defects in every model (0.16 to 0.70 fewer findings per task, every Holm-adjusted sign test significant), won 95 of the 100 differing pairs, and cut Bandit medium-or-high issues from 53 to 11. All 500 outputs, prompts, AST checkers and the pre-registration are published with a DOI, so a team can re-derive the result rather than trust it.