Notes
When Provotics returns a site of origin, it also returns a number: how confident it is. That number is only useful if it means what it says. A model can be accurate and still hand you confidence values you cannot trust, and a model can be a little less accurate while telling you exactly how much to believe each call. The second model is the one a researcher can actually work with.
A probability is calibrated when it matches reality over many predictions. If the model says it is 90% confident across a batch of cases, then it should turn out right on roughly 90% of them. Not 70%, which would be overconfident, and not 99%, which would be needlessly timid. We measure this with expected calibration error, and on our held-out evaluation it improved from 0.091 to 0.011, about an eight-fold reduction. In plain terms, the confidence numbers stopped lying.
For a researcher, a raw label is a starting point. A trustworthy confidence is a decision tool. If 90% genuinely means 90%, you can sort predictions by confidence and know what you are trading off when you draw a line. Act on the top tier, send the middle tier for review, set the rest aside. None of that works if the number is decorative. Calibration is what lets you budget against a probability instead of guessing at it.
Accuracy asks how often the top guess is right. Calibration asks whether the confidence attached to every guess is honest. They come apart constantly. A model can post a strong accuracy figure and still be wildly overconfident on the cases it gets wrong, which is the worst possible failure for anyone reviewing its output. We treat them as two separate things to get right, because for our work a confident wrong answer is more dangerous than a humble one.
Calibration handles the cases the model will answer. Conformal prediction handles the cases it should not. On data like what it trained on, Provotics runs at 90% conformal coverage: instead of forcing a single answer it returns a small candidate set that contains the right site about 90% of the time, or abstains when nothing clears the bar. That coverage is an in-distribution property, so it does not hold on inputs from sequencing platforms it has never seen. For those, a separate novelty check flags the input rather than trusting a number that no longer applies. The honest move is sometimes to return nothing, and we would rather do that than hand you a number that quietly sounds sure.
A bigger headline accuracy is easy to print and hard to use. A calibrated probability with a clear abstention rule is harder to earn and far more useful at the bench. It lets an expert interrogate a prediction instead of taking it on faith. That is the gap between a result you can defend in a meeting and one you are quietly hoping holds up.
Provotics is a research and educational project. It is not a medical device and is not intended for clinical diagnosis or treatment decisions.
← All notes