Decision-Space Collapse and Preference Narrowing in Personalized Language Models

Language-model assistants increasingly condition their answers on what they know about a user, which raises the question of which alternatives remain visible. Decision-Space Collapse (DSc) names one version of this concern: an advisory response surfaces some plausible courses of action while omitting others, measured as configured decision-relevant trajectories that the response does not surface. Its published results come from a historical study, not rescored, that used a configured, rule-based instrument and required neither personalisation nor memory. PRISK, a separate study, reports "preference narrowing": given a user profile, retrieved memories or both, responses include fewer of the options an LLM judge rated useful for that user. We argue that the common object is option visibility, not narrowing itself. The two studies offer converging evidence for context-sensitive option visibility, not replication and not validation of the DSc instrument: advisory responses leave out part of a reference option set even at baseline, and what is surfaced changes with what accompanies the request. The direction differs: DSc's main reframing raised coverage, while PRISK's personalisation lowered it. Context can broaden or narrow visibility, so the open question is which conditioning changes it, in which direction, and when the change is warranted. We separate this from a prospective longitudinal construct, Decision-Space Drift, which neither study measured. The note reports no new empirical result.

This research note is a synthesis. It reports no new data and no new analysis. Decision-Space Collapse results are taken from the original v1.1 study (DOI 10.17605/OSF.IO/KW25A); they have not been retrospectively rescored, and the historical scorer cannot be bound to a source commit. The note claims convergence only for context-sensitive option visibility, and claims no replication, validation, priority, harm or longitudinal effect. Decision-Space Drift is a proposed construct that has not been measured.

Licence scope. CC BY 4.0 covers the manuscript text, Figure 1 and the claim-evidence matrix supplement. Verification and figure-generation code is not released with this record.