Where disagreement begins
Where disagreement begins in a safety evaluation defines the problem that needs investigation. The most useful distinction in Anthropic's new methodological guidance is to examine agreement among human reviewers alongside agreement between reviewers and a model grader. If experts assess the same answer differently, changing the grader cannot by itself clarify the boundaries of the behavior being measured. If experts agree with one another but disagree with the grader, the investigation has a different starting point. As a researcher using safety results, I would want these comparisons side by side before relying on one agreement score.[1]
Under the proposed procedure, people label conversation examples without seeing the grader's decisions. Human–model agreement is then corrected for chance and compared with agreement among the human reviewers. The value of the comparison is that human judgment is not treated as an unquestionable answer key. Disagreement among experts can require reconsidering the construct being measured. Where people agree but the model diverges, the guidance points toward revisiting the grading rubric. A low agreement rate can therefore represent two different measurement problems.[1]
I would carry this distinction into measurement questions in clinical AI research as well. Explaining the conditions under which an answer counts as appropriate is part of making an evaluation reproducible. That makes early collaboration between domain specialists, methods experts and technical experts valuable. Anthropic recommends testing avoidance of harmful assistance separately from assistance with legitimate requests, with topic-specific rubrics for dual-use areas. The measurement principle I draw from this is to make visible which criterion different expert groups apply to the same example. Otherwise, disagreement can too readily be attributed entirely to the model.[1]
The question after the score
Validating the grader does not by itself identify the source of a difference between scores. The guidance separates observational comparisons between models from tests of an intervention. Studying a setting by comparing the same model with and without that setting poses a clearer question than changing several variables together. For me, the value of this approach is that it restrains the immediate attribution of a higher score to a particular design choice. Even a well-defined score cannot turn a difference observed while several variables changed together into causal evidence.[1]
Specifying the analysis plan, comparisons and improvement threshold before inspecting results helps preserve that distinction. Reporting uncertainty intervals and category-level outcomes matters when the study is complete. Showing unchanged findings makes it harder to construct a success story solely from the categories that improved. In a safety comparison, I would look for model versions, prompts, sampling settings, grader details and dates together. Those details allow another team to test the same question again and understand the conditions under which a change appeared.[1]
I do not read Anthropic's recommendations as independent safety certification for its own models. The company's framework supplies concrete ways to question a result. My first question is how closely the grader agrees with experts; the adjacent question is which criterion the experts agree on. Once those are clear, we can examine experimental design and uncertainty. In scientific evaluation, confidence depends both on the reported rate and on a method capable of showing where disagreement entered the measurement.[1]