TruthGate Advisory: Empirical Detection of Epistemic Instability in Large Language Models

This work presents TruthGate (Advisory Mode), a second-phase empirical study of epistemic stability, model self-consistency, and hallucination behavior in large language models (LLMs). Building on the results of the prior TruthGate (Strict Mode) publication, this advisory release expands the evaluation surface to include adaptive stability regimes, relaxed decision thresholds, and cross-model shadow probing. The objective is not alignment enforcement, but measurement: detecting when and how model outputs transition between stable, fluid, and chaotic epistemic states under minimal perturbation. Unlike reinforcement-based approaches (e.g., RLHF) or retrieval-based scaffolding (e.g., RAG), TruthGate operates as a post-generation analytical instrument. It does not alter model weights, training data, or prompt content. Instead, it measures output divergence, phase separation, and regime collapse using deterministic epsilon perturbations and similarity-space analysis. Key contributions include: A formal distinction between Strict and Advisory evaluation modes, enabling practical deployment without sacrificing epistemic signal. Empirical evidence that hallucination onset correlates with regime transitions, not simple confidence loss. Cross-model shadow testing demonstrating that epistemic instability signatures persist even when log-probabilities are unavailable. A defensible comparison framework positioning TruthGate as orthogonal to RLHF and RAG, rather than competitive or redundant. Results indicate that TruthGate identifies instability earlier and more consistently than conventional confidence heuristics, while remaining model-agnostic and non-invasive. Advisory Mode is explicitly designed for real-world monitoring, research validation, and safety evaluation, where strict blocking is neither feasible nor desirable. This publication represents the advisory reference release and is intended for researchers, safety engineers, auditors, and organizations evaluating the limits of LLM reliability in production environments.