NOESIS — Phase II: From Ontology to Observability: An Architecture for Epistemic Regime Detection
NOESIS — Phase II develops an architectural framework for observing epistemic regime transitions in generative AI systems. Building on the ontological reframing established in Phase I, which argues that many apparent “hallucinations” are better understood as failures of epistemic stability rather than isolated errors, this work focuses exclusively on the conditions required for such failures to be observable. Phase II introduces an observability-first architecture grounded in cognitive telemetry, comparative baselines, and regime transition detection. It treats epistemic regimes as intrinsic properties of a system’s internal dynamics, while explicitly acknowledging that observability is contingent on the interface available to the observer. Telemetry is defined as state-correlated signal activity originating from internal dynamics, regardless of whether it is accessed through direct instrumentation or constrained, interface-level proxies. Importantly, this work does not propose evaluation metrics, corrective mechanisms, safety interventions, or governance frameworks. Noesis Phase II is concerned exclusively with observation. Interpretation, judgment, and action are explicitly externalized, preserving a clear architectural boundary between epistemic measurement and downstream decision-making. The paper includes a single original architectural diagram defining the complete Phase II observability stack and its explicit termination prior to interpretation or action. The framework is intentionally implementation-agnostic and compatible with both white-box and black-box systems, with fidelity varying according to access constraints. This document is intended as a conceptual and architectural contribution, providing a stable vocabulary and structural frame for analyzing epistemic regime behavior in modern generative models. It is designed to be evaluated, critiqued, and extended, rather than adopted as a product, standard, or policy instrument.
Authors
- James Benjamin Jones
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-01-03
- DOI
- https://doi.org/10.5281/zenodo.18141913
- Primary Topic
- Human-Automation Interaction and Safety
- Type
- article
- Field-Weighted Citation Impact
- 0.00