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

v1.1 (2026-08-01): personal-data remediation (all source images now distributed in redacted form; v1.0 files restricted), added signal-absent / specified-task / critic-specificity experiments (the critic-specificity run uses the same critic model as the recovery experiment), SHA-256 stimulus manifest with duplicate mapping and image-type classification, a verification script, and the Supplementary Note 2 sensitivity analyses. Scenario generators, image-composition and stimulus code, model-query and adjudication scripts (including verbatim judge prompts), per-item adjudication result files, generated figures, and source/composited chest-radiograph stimuli. Reproduces every reported statistic and figure for the manuscript of the same title. Source radiographs are public domain (Wikimedia Commons), redistributed with identifier bands redacted.