How a fine-tuned local model turns molecular evidence into a cancer-origin verdict: waterfall, cloud vs local probabilities, calibration. Demo case, no patient data.
Each bar picks up where the previous ends — base 0 pp to +29.0 pp behind the 42% NSCLC call. Hover for the reason; click to pin it open.
supports the call opposes cumulative total
Jev cloud zero-shot vs GUT local fine-tuned, same 0–45% scale. Hover a bar for what the class means.
Key: NSCLC non-small cell lung · COADREAD colorectal · STAD stomach · PRAD prostate · insufficient not enough evidence for a call · other outside the 22-class taxonomy. Top pick per engine in red.
Where the model's confidence sits on the trust scale — and whether to take it at face value.
In plain English: how honest the model's confidence is. Trustworthy models score 0.10 or lower; this one scores 0.539 — stated confidence runs ahead of actual accuracy. Until it trains on more labeled cases, treat the percentages above as a rough ranking, not reliable probabilities.