The Verifier Is Where Values Live: Governing Evaluator Revision in Open-Ended AI

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

When an AI system revises the reward model, rubric or judge that trains it, some revisions only correct mistakes; others change what the system is trained to value. This paper separates the two, shows why recorded judgments and coarse harm rules cannot settle every new case, and proposes a standard for accountable evaluator revision. The standard requires a common promotion gate, a fixed safety floor, standing for affected parties, independent corroboration and external authority. It adds a conservative no-loss rule for changes a system proposes itself. Theory and position preprint; no experiments.