Two models read the same exam, and you see both readings — including where they disagree.
The point is not a second cheerful confirmation. It is the seam: the places where two independent readings of the same images do not line up are exactly the places worth taking to a doctor.
On request, a second independent model reads the same exam from scratch, and the two readings are shown side by side. Agreement is displayed, not silently merged. Disagreement is displayed too, not quietly resolved in favour of whichever answer sounds more confident.
Each finding, in either reading, carries a confidence level — high, moderate or low — because false certainty is the real danger with medical AI, and averaging two answers into one smooth paragraph manufactures exactly that.
No model here is our own. The roster is Anthropic's Claude Opus, Google's Gemini and Google's MedGemma 1.5, compared head to head on real studies with a known diagnosis. The one that wins on a given kind of study does the reading, and the exact version is stored alongside every analysis, so a result from six months ago can still be traced to what produced it.
The most useful second opinion is often not AI versus AI but AI versus the report you already have — one reading that saw the images and never saw the radiologist's conclusion, next to the conclusion itself. That is what the paid reading gives you, and it is why the reading is kept blind to the report.