Typed Process-State Constraint Durability Under Context Pressure: A Behavioral Adherence Study of Operator-Level Constraints in Multi-Turn LLM Agent Sessions

This study tests whether process-state verbs (PSVs), bracketed all-caps verb tokens (RESOLVE / AWAIT / FATAL) used as operator-level constraints, achieve durable behavioral adherence in DeepSeek V4 Pro under multi-turn context dilution. Such operator-level constraints are the primary mechanism for governing agentic deployments. Study 1 crosses 5 constraint-delivery conditions x 2 omission-class constraints x 6 conversation depths (N=150 per cell, 9,000 sessions); Study 2 (1,800 sessions) is a pre-planned confirmatory replication on this substrate of the omission/commission compliance asymmetry reported by Gamage (2026). Two pre-registered findings hold across sensitivity checks. (i) Gamage's omission/commission asymmetry is not supported: at depth 10, commission compliance was 0% and omission compliance 39% (a -39pp gap), inverting Gamage's reported +67pp on Mistral Large 3 across roughly 106pp of swing. (ii) The Knows-But-Violates partition is supported: PSV violations were 100% knowing violations and control violations 100% ignorance violations. An exploratory 200-session supplement shows the canonical noise corpus contained verbatim restatements of constraint vocabulary, functioning as depth-correlated implicit re-injection; under domain-neutral noise the 60.6% pooled canonical ambiguity rate largely dissolves. Decay-curve studies importing conversational noise should audit corpus-constraint vocabulary overlap before treating hedging rates as substrate properties.