The inattentional gap in task conditioned AI models that omit otherwise reportable safety critical signals (reproducibility deposit)

Code, per-item frozen results, prompts, and documentation for the manuscript "The inattentional gap in task conditioned AI models that omit otherwise reportable safety critical signals" (Shin, Kang, Min, Yang & Lee; PolymathMinds Lab, Asan, Republic of Korea). All reported statistics reproduce offline from the frozen outputs and per-item result files; API scripts are supplied separately for prospective reruns, subject to model availability and vendor updates. Stimulus images are third-party-derived and withheld from public redistribution (SOURCE_PROVENANCE_STATEMENT.md); dual SHA-256 manifests (frozen original / distribution copy) anchor the frozen set, and images are available from the corresponding author. Figure 3 shows a rights-verified reconstruction of one frozen stimulus identity (CC0 source included). v1.5 (2026-08-02): dual-hash stimulus manifest, all-author metadata, removal of an obsolete re-adjudication script (frozen record specified_task_full_results.json is authoritative), added item-level bootstrap CI script, updated Figures 1/3/6 with nearest-neighbor works (MixRea, arXiv 2605.20128; Hidden in Plain Sight, arXiv 2506.00258) on the literature map. Earlier versions: v1.4 disclosure alignment; v1.3 Figure 3 rights closure; v1.2 rights closure and matched analysis; v1.1 personal-data remediation; v1.0 original (restricted).