Noesis Tension: A Telemetry-Driven Taxonomy of Prompt-Induced Representational Pressures in Large Language Models

Noesis Tension is a lightweight, strictly telemetry-only framework for diagnosing how prompts induce specific representational pressures inside large language models.

Using only per-layer activation statistics (mean/median/p95 deltas, curvature, entropy, and final-layer spikes), the system computes two core indices (tension and drift) and maps them onto a compact set of twelve empirically-derived “tension categories.” These categories are then used by a deterministic classifier to assign high-level cognitive regimes (e.g., exploratory_liminal, confident_hallucination_lite, safety_procedural).

The framework requires no inspection of prompt or response text, runs in microseconds, and shows stable core attractors across architecturally distinct models (Llama-3.1-8B, Mistral-7B, Qwen1.5-MoE-2.7B). Results are demonstrated on a false-presupposition prompt that reliably triggers ontological impossibility and related pressures.

This work constitutes Phase III of the NOESIS project and builds directly on the conceptual foundations established in:

- Phase I: Toward Epistemic Regime Detection (Zenodo, Dec 2025)
- Phase II: Observability Architecture for Cognitive Telemetry (Zenodo, Jan 2026)

Full classifier source code and example traces are included in the appendix of the preprint and will be open-sourced upon arXiv submission.

Author: James Benjamin Jones (ORCID: 0009-0002-6129-2847)