We tried per‑point conformal intervals and rejected them, and the rejection is in the product

A constant confidence band is the crude version of calibrated uncertainty: every prediction gets the same interval regardless of whether the input is easy or hard. Per-point conformal promises better — an interval that widens where the model is likely wrong. We tried it on three targets.

Where it worked

On the two capacitance targets it did exactly what it promises: intervals 22% and 10% narrower at the same coverage, and width correlated with actual error at +0.45 and +0.49. The interval was carrying real information about difficulty.

Where it failed

On resonator frequency the same method produced intervals that were wider, covered less — 0.893 against a 0.95 guarantee, so the guarantee was violated, not merely tight — and whose width correlated with error at −0.03. No signal at all, and worse than the thing it replaced on every axis that matters.

So per-point conformal ships for capacitance and does not ship for frequency. The constant band stays.

Why this is on our website

Because the honest version of "we use conformal prediction" is that we tried the better method, measured it, and it lost on one of three targets. A vendor page that says "calibrated uncertainty" without saying which method on which target is not telling you anything you can act on.

The rejection is recorded in the product's own validation card under per_point_rejected, with a test, so it cannot quietly come back without someone re-measuring it. A method you tried and rejected is part of the specification, not an embarrassment to leave in a notebook — and if you are deploying per-point conformal on a target where width does not correlate with error, you have added complexity and lost coverage.

What we would do differently

Check the width–error correlation before adopting the method, not after. It is one number, it is cheap, and it would have told us in an afternoon what the full comparison told us in a week.