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" (Kwan Soo Shin et al., PolymathMinds Lab, Asan, Republic of Korea). Every reported statistic reproduces offline from the frozen per-item JSON files; API rerun scripts are provided for extension. Stimulus images are third-party-derived and withheld from bulk redistribution (per-file rights records were not logged at acquisition; SOURCE_PROVENANCE_STATEMENT.md). SHA-256 manifests anchor the frozen set; images are available from the corresponding author. One disclosed exception: the Figure 3 source identity was verified post hoc as byte-identical to a CC0 Wikimedia Commons original, and Figure 3 shows a rights-verified reconstruction of that frozen stimulus identity (CC0 source included). v1.4 (2026-08-02): disclosure alignment for the Figure 3 stimulus lineage (cxr_4/cxr_8 = CC0 original, post-hoc checksum verification recorded). v1.3: Figure 3 rights closure and removal of unverifiable rights assertions. v1.2: rights closure and matched signal-detection analysis. v1.1: personal-data remediation and added control experiments. v1.0: original deposit (restricted).