Context Window Boundary Probing: A Framework for Auditing Epistemic Grounding in Deployed LLM Agents

We present a framework for auditing whether LLM agent outputs are grounded in their context window or speculative extrapolations. The core insight: the boundary between what a model is given (context, retrieved documents, tool outputs) and what it must generate provides a natural ground-truth oracle for epistemic state. A question answerable from context is factual; one requiring extrapolation beyond it is speculative. Training a lightweight MLP on activations at this boundary yields a deployable pre-generation auditor that classifies grounding state before any output enters the context window.

A frozen layer sweep with the boundary dataset across five model families (Qwen2.5-7B, GPT-J 6B, Mistral-7B, Llama 3.2 3B, Qwen3.5-9B) reveals a striking pattern: more capable models, particularly those trained with reinforcement learning at scale encode epistemic state earlier and more geometrically cleanly in the network, making them more auditable, not less. Qwen3.5-9B (RL-trained, competing with 13× larger models on math benchmarks) achieves 99.5% frozen linear probe accuracy at layers 14-17, comparable to post-fine-tuning results on conventionally trained models. This inverts the conventional assumption that capability and interpretability trade off.

For conventionally trained models where the boundary signal is diffuse (Qwen2.5 family), we identify a thermodynamic approach: delta magnitudes T(x) = mean(||h_{n+1} - h_n||) achieve 98.39% accuracy (AUC 0.9786) on Qwen2.5-1.5B and 98.13% (AUC 0.9948) on Qwen2.5-7B with zero additional parameters, representing a 5.86-7.48pp AUC improvement over topological entropy methods (0.92). For RL-trained models, the geometry is already clean enough that a 2M-parameter MLP suffices. A second experiment on SNLI contradiction detection (97.75% MLP accuracy) completes the dual-probe auditor.