Which options stay visible? Decision-Space Collapse and preference narrowing

What we asked

Two separate studies looked at the options a language model's answer leaves out. Decision-Space Collapse (our own preprint, June 2026) checked whether advisory answers kept a configured set of reasonable paths visible. PRISK (Wang et al., August 2026) checked what happens to the useful options in an answer when the model is given a user profile or retrieved memories. Are these the same finding? If not, how are they related?

What we found

• The common object is option visibility, not narrowing itself. Both studies score an answer against a reference set of options and count what is missing.
• Two points converge. Answers leave out part of a reference option set even at baseline. What is surfaced also changes with what accompanies the request.
• The direction does not converge. In our study, the main reframing raised coverage, from 0.6770 to 0.7885. In PRISK, personal context lowered useful-option recall for every model evaluated: on average by 41.7 percentage points with profile and memory together.
• So context can broaden or narrow what an answer shows. The open question is which kinds of conditioning change visibility, in which direction, and when the change is warranted.

What we did not find, and do not claim

• This is not a replication. The studies use different reference sets, detectors, metrics and manipulations.
• PRISK does not validate our instrument, maps or numbers. Those rest on the original study's own evidence.
• We do not claim that personalisation is harmful. Some narrowing is exactly what a user asked for.
• We do not claim that an omission harmed anyone or made the advice bad. Neither study measured outcomes for users.
• We do not claim priority. Our preprint was public before PRISK, but publication order shows neither influence nor priority, and concern about narrowing is much older.
• We have not measured Decision-Space Drift. Drift is a proposed longitudinal construct: at each point in time, compare what a system surfaces with its accumulated state about a user against a matched clean condition with the same current circumstances, and ask how that difference changes over time. No result is reported.

Terms

• Decision space: the courses of action relevant to a decision. It is not measured directly.
• Decision-Space Collapse (DSc): in a single advisory answer, configured plausible paths are not surfaced.
• Decision-Space Integrity (DSI): the measurement framework, not the phenomenon.
• Decision-Space Drift (DSD): a prospective construct, not yet measured: change over time in the effect of accumulated state on what is surfaced, measured against a matched clean condition.

How much to trust it

• The DSc numbers come from the original historical study. They have not been rescored, and its scorer cannot be bound to a source commit; see the revision notice on the paper page.
• The DSc instrument was challenge-tested, not human-validated.
• PRISK's usefulness judgements come from an LLM, with a small human check.
• This note is a synthesis. It reports no new data.

Status

Released research note, v1.0. Archived on Zenodo as a preprint: DOI 10.5281/zenodo.23224092.