Testing as Environment: The EC-EpC Gap Framework for Assessing Epistemic Boundary Alignment in Large Language Models
Year: 2026
Type: preprint
Abstract: This preprint presents the EC-EpC Gap framework, a dual-axis approach for assessing divergence between a model's observable task-performance boundary, termed Epistemic Capability (EC), and its expressed epistemic boundary claims, termed Epistemic Claim (EpC). The framework is designed for restricted-access LLM evaluation: it does not require model weights, hidden states, internal representations, or proprietary telemetry. It introduces the Testing-as-Environment principle, which treats interaction sequences as structured behavioral environments that can shape, elicit, and reveal epistemic boundary behavior. The framework distinguishes overclaiming, underclaiming, calibrated uncertainty, and interaction-induced epistemic drift. This release establishes priority of core definitions and methodology and serves as an operational bridge between the EC-EpC line, ABS structural integrity work, and the broader Testing-as-Environment research direction. It is a conceptual and pilot-level disclosure, not a complete experimental report or a claim of finalized statistical validity.
Author: wang, sangzi
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7162070299