Operational Identifiability of Internal Intention in AI Systems — A Measurement and Validation Framework (v2.0)
Omri Bankuti · Zenodo (CERN European Organization for Nuclear Research) · 2026
This work extends the formal operationalization of internal intention in AI systems (v1, DOI: 10.5281/zenodo.19376962) into a full measurement and validation framework. Version 2.0 introduces: identifiability with constructive exclusion and worst-case bounds; robust statistical guarantees (UCB, independence, multiple testing correction); intervention coverage with information-theoretic thresholds; elicitability as an optimization problem; external validation via independent proxy triad; false negative characterization and detection boundary (α*). This document defines the detection core and measurement theory. Companion work addresses construct validity and necessity. This work is also subject to the ORIGÓ License (see document).