Omission Constraints Decay While Commission Constraints Persist in Long-Context LLM Agents

Instruction-following evaluations test whether a model obeys a behavioral constraint at turn 1 and assume the result holds. We demonstrate that omission constraints decay under context accumulation while commission constraints persist. Omission rules leave no positive trace when obeyed, so the policy document encoding them competes for attention weight against growing context volume and loses. Commission outputs appear in the model's own conversational history and self-reinforce forward. In a 4,416-trial three-arm causal study across 12 models and 8 providers at six conversation depths, omission compliance falls from 73% at turn 5 to 33% at turn 16 while commission compliance holds at 100% throughout (Mistral Large 3, CMH χ² = 147, p < 10⁻³³). Token-matched padding controls show schema semantic content drives 62–100% of the effect; token volume alone does not. We introduce Safe Turn Depth (STD), a per-model threshold at which omission compliance crosses 50%, and show that constraint re-injection before STD restores compliance without retraining. The behavioral constraints governing deployed agents are prohibitions: never disclose credentials, never execute untrusted code, never forward user data. They pass every existing single-turn evaluation. They are the structural class that decays.