Enterprise deployment of Large Language Models (LLMs) faces persistent inference-time challenges: hallucinations expressed with high confidence, internal inconsistency across turns, unjustified stance reversals under user pressure, and over-accommodation to perceived preferences. While recent work in LLM evaluation and self-consistency sampling has made progress on some of these issues, a dedicated inference-time stability mechanism—distinct from both training-time alignment and external guardrails—remains underexplored. This paper introduces Evaluative Coherence Regulation (ECR), an inference-time stability layer that constrains internal inconsistency across short reasoning horizons using explicit, measurable criteria. ECR does not modify model parameters, require retraining, or assume access to ground truth. Instead, it evaluates multiple candidate response trajectories using mathematically defined coherence metrics—evaluative variance, contradiction rate, trajectory smoothness, expectation stability, and policy divergence—each normalized to [0,1], and selects responses that remain internally stable under uncertainty. ECR is explicitly positioned as a containment and reliability mechanism for mature AI systems, not an optimization objective, alignment guarantee, or truth verification system. We present formal definitions with explicit normalization schemes, an inference-time selection algorithm, system maturity preconditions, scope limits, a worked numerical example, and practical deployment guidance. The framework is lightweight, auditable, vendor-neutral, and designed to meet the practical and conceptual needs of enterprise AI deployment.