cpath-triage: diagnostic analysis of verbalized confidence in a medical vision-language model

Evaluation and analysis code, plus the prediction artifacts behind every table and figure, for a diagnostic study of verbalized confidence in MedGemma-27b-it on colorectal H&E histopathology patches. Covers zero-shot classification, temperature scaling and calibration metrics, selective-prediction routing with explicit tie handling, a design-effect sensitivity analysis in place of clustered intervals, and K=5 prompt-variant consistency routing. Version 4.0 added a cohort-weighted re-analysis of the K=5 subsets, a wording-only prompt family, image-removal, image-swap and pixel-shuffle controls, and a PLIP image-text baseline. Version 5.0 adds the wording-only family on the external cohort, each prompt variant's own stated confidence, Platt and isotonic calibration maps with a Brier decomposition, agreement against the number of prompts, operating points with measured query cost, per-class results under each pilot prompt, a QuiltNet image-text baseline, and the per-call logs of the runs that kept them.