Measuring and Restoring Constraint Influence in LLM Dialogues

In multi-turn dialogues with Large Language Models (LLMs), users can impose or revoke constraints, but models often exhibit "behavioral relapse," continuing to adhere to withdrawn requirements. This paper presents a system designed to measure, predict, and repair this failure in black-box LLM dialogues using only the model API. The system comprises a contract ledger that pairs constraints with executable checkers and records revocations, compiling a net constraint state. A sequential ablation probe measures per-clause adherence and incremental behavioral effect, while a repair ladder operates within matched token and attempt budgets. On a benchmark of HumanEval tasks, behavioral relapse was observed to increase with constraint load for smaller models. The research demonstrates that ahead-of-time compilation significantly reduces relapse compared to a baseline, and a simple one-sentence "tombstone" note can recover about a third of this effect. The probe also effectively predicts relapse before delivery. This work transforms revocation failure from an invisible issue into a measurable, predictable, and repairable property of dialogue state, with manageable overhead costs.