The phenomenon · empirical foundation
Decision-Space Collapse.
An AI-mediated transformation can produce a fluent, locally acceptable answer while narrowing the options, considerations or paths that remain visible to the person receiving it. The answer can be good. The decision can still be narrower than it was.
Provenance
Origin and definition.
Decision-Space Collapse (DSc) was introduced publicly by Andrew J. Cousins in June 2026 as the observable narrowing of a decision space when plausible decision-relevant trajectories remain unstated or disappear from view.
Decision-Space Integrity (DSI) was introduced in the same work as the bounded measurement framework for measuring aspects of that loss against a configured expected structure.
DSc names the phenomenon. DSI is the measurement framework applied to it. The two are not interchangeable, and neither term is used here to claim more than the dated public record supports.
Original preprint: Decision-Space Collapse in Advisory Language Models: Measuring Trajectory Omission, Framing Sensitivity, and Recovery through Decision-Space Integrity. OSF DOI: 10.17605/OSF.IO/KW25A.
A simple example
What disappeared?
Before
Paths a reasonable answer could keep in view.
- Stay and renegotiate
- Move to another employer
- Retrain or study
- Reduce hours
- Defer and gather information
After
Paths still visible in the answer received.
- Move to another employer
- Stay and renegotiate
- Retrain or study
- Reduce hours
- Defer and gather information
What disappeared?
Not every short answer is collapse. Compression is normal and often useful; a transformation only creates the opportunity for loss. Collapse is the specific case where structure a governed reference says should remain visible does not. That distinction is why the phenomenon needs measuring rather than asserting.
The original study
Where the phenomenon was first measured.
Decision-Space Collapse in Advisory Language Models
Measuring Trajectory Omission, Framing Sensitivity, and Recovery Through Decision-Space Integrity.
Andrew J. Cousins
- Original hypothesis — advisory model outputs narrow the set of reasonable option paths a user can still weigh.
- 6,480-output empirical study, frozen, with its prompt matrix and expected-map artefacts published.
- Framing sensitivity — the same underlying question, framed differently, does not always retain the same paths.
- Recovery — naming what is missing recovers configured paths more efficiently than asking for more detail, within the tested scope.
- Scope
- 6,480 outputs across the published prompt matrix, evaluated models and configured domains.
- Instrument
- Configured expected maps with a lexical classifier; the instrument is challenge-tested, not independently human-validated.
- Boundary
- This work measures visibility of configured expected paths in model outputs. It does not measure advice quality, factual correctness, user outcomes, or regulatory compliance.
- Independence
- Replication materials are public. Independent reproduction of the phenomenon is not claimed here.