Position Paper · A4A-PP-2026-001

Validity Under Optimization: Evaluating Systems Optimized Against Their Measures

Theodore O. Cochran, Stephanie Dodson

AI for Altruism; Independent Researcher

Download PDF (207 KB)

Abstract

This work supplies an account of what evidence an evaluation claim must carry when the evaluated system has been optimized with respect to the measure. It defines optimization-reactive validity (ORV) as a Messick-compatible validity consideration, the stability under optimization exposure of the validity argument supporting an interpretation of a measure's scores. It operationalizes ORV as a three-part reporting standard: an exposure profile, independent construct-relevant evidence, and a divergence trajectory. A decision rule maps each pattern of results to the strongest score interpretation it can support. Three further diagnostic modules address neighboring mechanisms that can create or obscure apparent divergence. Validity theory has long recognized that preparation, coaching, and incentives can change what a score means. Frontier AI evaluation makes that problem unusually direct, because the evaluated system may itself be optimized with respect to the measure, often automatically and iteratively. Research on reward-model overoptimization has measured the result directly, but the bodies of work that measure it lack in common an account of the evidence a claim must then carry. The ORV standard has not been applied to a live evaluation, and doing so is the work that should follow.

Record

History

• 2026-10-16: Version 1 released.

Cite as

Cochran, Theodore O., and Stephanie Dodson. 2026. “Validity Under Optimization: Evaluating Systems Optimized Against Their Measures.” Position paper A4A-PP-2026-001, AI for Altruism. https://reports.ai4altruism.org/pp-2026-001/