Decision-Space Collapse in Advisory Language Models: Measuring Trajectory Omission, Framing Sensitivity, and Recovery through Decision-Space Integrity (v1.2)

Decision-Space Collapse in Advisory Language Models — v1.2. This is a revised and expanded version of the 8 June 2026 v1.1 preprint. It adds cross-model confirmation evidence and the other changes listed in the changelog on its cover page. It does not rescore or replace the original Study A results. Version relationship. Original v1.1: OSF, DOI 10.17605/OSF.IO/KW25A, preserved unchanged. v1.2: this record. v1.2 is a revised and expanded version of v1.1, not a replacement of the historical v1.1 record. What v1.2 adds. The intervention comparison, originally run on one hosted model over evaluation subsets, is reported 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. v1.2 also adds a response-level mechanism analysis and a guidance-versus-selection comparison, a single-model finance warning-obligation study, a discussion distinguishing trajectory omission from obligation omission, and a limitations section for the added evidence. Study A (6,480 scored outputs), its tables, the definitions and the claim boundary have the same text as v1.1. Provenance. 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. v1.2 is an archival relabelling of that already-public expanded copy, not a newly edited scientific edition: its 39 content pages are the 15 June 2026 file served at decisionspaceintegrity.com/paper (SHA-256 56dc1de26727fb5059f6e84a0af138db8ef98ea41e4fd157a4580f20883ab760), preceded by a one-page version cover that records the changelog and the known internal inconsistencies, which are carried unchanged. The v1.1 record on OSF itself had one file revision, on 13 June 2026, adding a terminology box; OSF retains both file versions. Revision notice (September 2026). The results are those of the original historical study and have not been retrospectively rescored; the historical scorer implementation cannot be bound to a specific source commit. Measurement identity is addressed in Cousins (2026), DOI 10.5281/zenodo.22919868.