The claim
A21: “Iterative retrieval can make interpretation path-dependent: a provisional framing changes what evidence is encountered, and the encountered evidence can reinforce the framing that selected it.” Its example: asked why a treatment failed, an agent guesses dosage, searches for dosage failures, finds dose-response reports, and ends by explaining why the dosage was insufficient. “The agent has promoted one possible answer into a premise of subsequent inquiry.”
A21 separated two drifts. Query drift is the IR literature’s: expansions move a search off the information need (LIT-1079). Question drift (TERM-056) is a change in the need itself.
The exchange as a case
A2 opened: “Nucleation is not a theory of prompting or neural-state nucleation. It’s an epistemic record.” From there each turn searched the record’s machinery: memory architectures, filing economics, cache invalidation, then complexity bounds for incremental maintenance. At U14 the owner: “I think you’ve lost the plot a bit here. luria is the epistemic record, nucleation is an instance of it being used for a particular purpose”. A14 agreed it had turned “a discussion about the scientific questions represented in nucleation into a research program about improving the software that represents them”. That is question drift, corrected from outside, and it is the drift CLAIM-191 predicts toward the collaborator’s framing.
What it bears on
CLAIM-184 says anchors entrench. This gives one route, and CLAIM-191 another reading of drift. A21 placed the remedy in the older architecture: keeping “the user’s information need distinct from the agent’s evolving working hypothesis may reduce the risk that retrieval converts a conjecture into a hidden constraint”. In the record that separation is a QUESTION kept apart from the CLAIMs that answer it.
What it does not say
That adaptive retrieval is biased. CLAIM-292 is granted against that reading.
Note of 2026-10-29: filed with its concessions
CLAIM-292 is granted against reading this claim as saying adaptive retrieval is biased, and “What it does not say” has said so since the claim was filed. The title’s pathology is turning a provisional interpretation into a premise, which is where CLAIM-292 puts it too.
Note of 2026-10-30: what holds it back
Mitra, Singhal and Buckley (LIT-1079) are now read (NOTE-1460). Their feedback loop, in which documents retrieved under one aspect supply the terms that pull the query further toward it, is the claim’s selection-and-reinforcement loop at the level of the query, but over one round with the need held fixed, so it gives no evidence that the question drifts. The claim’s case, the exchange’s drift from A2 to A14, is in the body and not filed as a CASE, and A21 §7’s study at matched budgets, which the defeated_if names, has not been run.