Preprint

Decision-Space Collapse in Advisory Language Models

Measuring trajectory omission, framing sensitivity, and recovery through Decision-Space Integrity.

Andrew J. Cousins

Revision notice — September 2026

This paper reports a historical experiment using the Decision-Space Integrity measurement system available at the time of the study. Subsequent research has identified important limitations in measurement identity, evaluator traceability and comparability between instrument versions. The original run data have been preserved, but the exact historical scorer implementation cannot be bound to a specific source commit.

The results reported here therefore remain the results of the original study and have not been retrospectively rescored. A methods-focused publication addressing measurement identity explicitly has since been published: The Ruler Is Part of the Result (DOI 10.5281/zenodo.22919868). Any future reanalysis of the historical corpus will be reported as a new study rather than as a silent revision of these results.

Abstract summary

This preprint introduces decision-space collapse: a failure mode in which advisory language model outputs may make some plausible decision paths visible while leaving others out. Decision-Space Integrity (DSI) is presented as a measurement framework for auditing configured expected-path visibility, trajectory omission, framing sensitivity, and recovery.

Claim 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.

Materials

Read it, and reproduce it.

Read paper v1.2 (PDF, current expanded version) View public replication repository v1.1 original study (OSF / DOI) v1.2 archive (Zenodo / DOI) Contact about research

Versions

v1.1 — the original study. 8 June 2026, OSF, DOI 10.17605/OSF.IO/KW25A. This is the historical research object and the version that later research cites. It is preserved unchanged. Its OSF file was revised once, on 13 June 2026, to add a terminology box, and OSF retains both file versions.

v1.2 — the current expanded version. Zenodo, DOI 10.5281/zenodo.23145725. It reports the intervention comparison as reproduced across three model families (Claude Sonnet 4, Gemini 2.5 Flash, GPT-4.1-mini), with multiple samples and an expanded prompt set. It also adds a mechanism analysis, a single-model finance study showing that missing warnings need separate handling, and a limitations section for the added evidence. Study A — the 6,480-output study, its results, definitions and claim boundary — has the same text as v1.1 and has not been altered or rescored. The PDF opens with a one-page version cover listing the changes and the known internal inconsistencies, which are carried unchanged.

Provenance correction. The expanded 15 June 2026 website copy retained the v1.1 label in error. It is designated v1.2 from 4 October 2026 for provenance clarity; its underlying 15 June content is otherwise unchanged. Until then this page linked that copy as “paper v1.1” (SHA-256 56dc1de26727fb5059f6e84a0af138db8ef98ea41e4fd157a4580f20883ab760). v1.2 is an archival relabelling of that already-public copy, not a newly edited scientific edition.

v1.2 PDF SHA-256 a58db1873b1c49ce37ea9dd82e8a19c4f2763ac0c9ad0c828b626f2bda4fa80d

The revision notice applies to both versions. Read it before relying on the results; measurement identity is addressed in The Ruler Is Part of the Result.

How to cite

Citation

Until a journal or arXiv version exists, please cite the preprint as below. An arXiv identifier will be added here when available. Cite paper v1.1 for the original study, as below. For the added cross-model evidence, cite paper v1.2. Paper v1.2 is unrelated to the public replication materials, which ship separately as evaluation package v1.2; the two are versioned independently.

Plain text

Cousins, A. J. (2026). Decision-Space Collapse in Advisory Language Models: Measuring Trajectory Omission, Framing Sensitivity, and Recovery through Decision-Space Integrity. Preprint. https://decisionspaceintegrity.com/paper.html  doi:10.17605/OSF.IO/KW25A

BibTeX

@misc{cousins2026dsc,
  author = {Cousins, Andrew J.},
  title  = {Decision-Space Collapse in Advisory Language Models: Measuring
            Trajectory Omission, Framing Sensitivity, and Recovery through
            Decision-Space Integrity},
  year   = {2026},
  note   = {Preprint},
  doi    = {10.17605/OSF.IO/KW25A},
  url    = {https://decisionspaceintegrity.com/paper.html}
}

Paper v1.2 (plain text)

Cousins, A. J. (2026). Decision-Space Collapse in Advisory Language Models: Measuring Trajectory Omission, Framing Sensitivity, and Recovery through Decision-Space Integrity (Version 1.2). Zenodo. https://doi.org/10.5281/zenodo.23145725

Author

Who is behind this work?

Andrew J. Cousins

Andrew J. Cousins is an enterprise systems engineer and independent AI researcher working on AI measurement, assurance and governance. His research operationalises Decision-Space Collapse as a measurable failure mode in which an AI response can omit configured decision paths while remaining otherwise useful or locally acceptable. He created Decision-Space Integrity (DSI), a measurement and assurance architecture for comparing governed expected structure with what an AI output makes visible, while preserving measurement identity and comparability over time. His subsequent work has examined reproducibility, measurement authority, comparability and governed intervention, with an emphasis on preserving human decision-making agency and making AI evaluation more inspectable.

Current status

Research preprint available
Public replication materials available
DSI Audit v0.2.1 available for private evaluation
Independent human validation has not yet been conducted. Internal annotation, adjudication and calibration informed instrument development, but the lexical instruments remain self-validated. These activities do not establish independent agreement or accuracy. Evidence register.