The claim
A21 §4: “adaptive retrieval is not inherently irrational or biased. In an ideal sequential Bayesian design, search policy can depend on current beliefs, and inference can account correctly for how observations were selected.” The pathology lies elsewhere: treating selected evidence as representative, ignoring explanations the queries excluded, treating conjectures as constraints, or stopping because one’s own results agree with each other.
And the measure: “Neither measure alone certifies bias. We’d need an external standard of task relevance or correctness.” That is CLAIM-190’s point about the record’s own metrics, applied to search.
Why it is granted and Active
It is a standard result of Bayesian design that the selection rule does not bias inference when the likelihood accounts for it, and it cuts against the strong reading of CLAIM-301. A20 also said it of a single retrieval: “Cue-dependent retrieval does not imply bias by itself.”
What it does not say
Not that real agents account for their selection. Whether they do is QUESTION-107.