Preprint finds signed requirements reduce unsafe AI agent decisions

A preregistered study reported in an arXiv preprint found that signed requirement manifests cut unsafe commits under attack from 69.4% to 9.0% in synthetic tasks using Qwen3-8B. Across 1,440 trajectories and five threats, the method preserved conformant commits relative to structured handoffs. The system records requirements as verbatim quotes and checkable conditions, signs them with HMAC-SHA256, and uses deterministic code to verify the merger's decisions. An unsigned manifest committed violations in 79% of mutated runs; the signed version committed none. The protection stops at extraction: an adversarial line introduced before signing raised violations from 14.6% to 37.5% for Qwen3-8B. The signature then authenticated the corrupted requirements.