Research data and replay code for Representation-Dependent Scores in IFC Editing Evaluation: Controlled Audit and Bounded Revision Checks

Version 1.2.0 provides research data and study-authored replay code for a controlled, version-specific audit of a public IFC editing evaluator. It is not a manuscript preprint or a claim of journal publication. The predecessor v1.0.0 and v1.1.0 records retain their original files. Executable coverage: the complete task11 witness; the original R1 integrity-submetric audit of 15 configurations in 11 related input files; the unchanged recovered historical provenance-aware property correction and its 27 synthetic tests, retaining 8 PASS and 7 UNKNOWN; eight calibration conditions and their 48 component readings; and all four full-mesh calibration representation pairs. The complete original task-level replay is task11; the broader R1 replay is a scoped submetric audit. Shared inputs and helper logic are not independent engineering projects, and UNKNOWN is not counted as a repaired configuration. The bounded 2026-10-30 revision checks add one public editing-API path, 12 new near-boundary conditions plus two shared controls, an independent analytic STEP-parser check for three core products across four files, and explicit support-domain and source/candidate argument-position checks. Nonorthogonal-axis and parent-chain limitations, unit-scale and floating-point boundary behavior, and unsuccessful support cases are retained. These checks do not establish general IFC canonicalization, natural LLM failure prevalence, a leaderboard correction, or the placement signature’s unique incremental discrimination. Reproduction requires separately acquired fixed third-party dependencies and the pinned Python 3.11.9 / IfcOpenShell 0.8.5 / NumPy 2.4.4 runtime. Fixed URLs and SHA256 identities are supplied. The same-machine execution in new directories and alternative software paths is not an independent-team or independent-environment replication. R2, natural model editing and consumer-effect pipelines are not fully replayed. A sanitized public projection of the revision freeze removes personal absolute paths while retaining the original freeze identity; it is not a byte-identical original receipt. The saved-run checker applies to a newly generated complete local run, not the partial public snapshot without IFCs and the original private-path freeze. Excluded: third-party source or derived IFC files, upstream evaluator/helper source, third-party binaries, model/API logs, credentials, and manuscript/figure submission files. Study-authored Python code is MIT; study-derived data and documentation are CC BY 4.0. This is a file-specific allocation, not a choice of either license for every file. Historical copyright notices are preserved. See RIGHTS.md and FILE_INVENTORY.csv. Third-party dependencies retain their own terms. OpenAI Codex assisted implementation, extraction, documentation and internal checks. It is not an author or independent human reviewer; this version does not claim a newly completed personal human-review checklist. The fixed BIBIMBAP source snapshot is not asserted to be the historical execution version of published benchmark scores. The audited work is referenced by DOI 10.35490/EC3.2026.271.