Alemetria: A Protocol for Auditing Interpretational Behavior in Large Language Models (Short Version)
A. V. Blinkova · Zenodo (CERN European Organization for Nuclear Research) · 2026
This record presents a condensed English version of Alemetria, a protocol for auditing interpretational behavior in large language models (LLMs). Alemetria shifts evaluation from direct output comparison to validation of the conditions under which comparison is methodologically permissible. The protocol is organized as a structured sequence (B0-B4) and introduces a rule prohibiting calculation when interpretational comparability or stability conditions are not satisfied. The method distinguishes between substantive agreement, false consensus, divergence, evasion, interpretational suppression, and instability under reformulation. This condensed version focuses on the formal logic of the protocol. For the full methodological specification, extended empirical data, case studies, and appendices, see the Russian preprint:https://doi.org/10.5281/zenodo.19674164