What a reliability diagram tells you that accuracy doesn’t

Two models score 85% on the same test. The first is calibrated: when it says it's 90% sure, it's right about 90% of the time. The second is overconfident: when it says 90%, it's right 70% of the time, and it says 90% constantly. Same accuracy. One of them you can build a high-stakes system around, and the other one will hurt you — and the single accuracy number can't tell them apart.

In a system that's allowed to abstain, calibration matters more than accuracy. A confidence you can't trust is worse than no confidence at all.

The diagram

A reliability diagram is the cheapest honest picture of a model you can draw. Bin its predictions by the confidence it stated — everything it called "0.9", everything it called "0.6" — and for each bin plot the stated confidence against the accuracy it actually achieved. The diagonal line is perfect calibration: confidence equals correctness. A well-behaved model hugs that line. An overconfident one sags below it, worst at the high-confidence end, which is exactly the end you were planning to trust — you can watch that happen live in the calibration explorer. It's a picture worth being able to draw for any classifier, not just compute an accuracy for.

The single-number version of that gap is expected calibration error — the average distance between the bars and the diagonal, weighted by how many predictions land in each bin. It's a summary, and like all summaries it can hide structure, but it's a far better headline for a high-stakes model than raw accuracy.

Why this is the whole game for abstention

Everything I build is allowed to say "I don't know — send this to a human." That's the abstention behavior that makes a system safe. But abstention runs on the confidence number: you answer above a threshold, you defer below it. If the confidence is miscalibrated, the threshold is meaningless — you'll confidently answer exactly the cases you should have deferred. Calibration is the thing that turns "90% sure" from marketing into a decision you can route on.

That's why, in Littman, the approve/flag/escalate boundary isn't a softmax cutoff — it's a selective-prediction problem with a distribution-free guarantee behind it, so the wrong-approval rate on the cases it does take stays under a target I picked on purpose.

None of this is exotic. It's a diagram you can draw in an afternoon and a habit most demos skip because accuracy is the number that sells. But if you're going to let a model make decisions you'd be sad to get wrong, the first question isn't "how often is it right." It's "does it know when it isn't."