A preregistered trial found structured spec contracts don't beat narrative prose.

arXiv 2608.10314 had two LLM snapshots translate five theoretical accounts into code under structured-contract versus prose formats, producing 320 programs. Both primary hypotheses returned NOT_SUPPORTED, only 19 of 108 criterion evaluations passed, and cross-model format identifiability sat near chance (AUC 0.469-0.523). It's a null result, so it bounds rather than kills the practice, but it suggests the payoff people credit to structure may come from the added content instead.