From radiology benchmark scores to patient benefit: a completed exploratory stock-model comparison and a conditional impact model

Abstract

Completed exploratory comparison: Athena 191/210 (90.95%), stock Astra 193/210 (91.90%). Stock scored higher descriptively; this non-resource-matched evaluation does not establish a scaffold-specific causal effect, Harrison superiority, clinical safety or lives saved. Completes the remaining 100 stock answers after the graded/published 110-item snapshot. All original answers are preserved. No local-memory tools or conversation continuation were sent to the remote stock model; the local broker was not an OS sandbox. Independent software checks verify all 210 scores, not clinical truth. The live workflow was only partially operationalised, and identical underlying checkpoint/full-session compute remain unverified. AI-assisted, not peer reviewed or clinically validated. Public availability, prior exposure, question cueing, post-hoc ranked-key decoding, capacity/transport changes, partial-snapshot chronology and proprietary-method limitations are disclosed. Mortality scenarios are explicitly hypothetical and include zero benefit and harm. No private medical/legal material, operational scaffold, credentials, original radiographs or private reasoning traces are redistributed.