How it compares

Heuristic confidence scores, and what replaces them

Most LLM stacks gate automation on a confidence number: a softmax probability, a verbalized score, or a single-model confidence API. Here is what changes when that number becomes a calibrated procedure.

Heuristic confidence scores compared with Conformal Logit Inference
DimensionHeuristic confidence scoresConformal Logit Inference
What you receiveA single number between 0 and 1A set, an interval, a probability interval, or a decision, each with a guarantee card
What is promisedNothing formal. “Calibrated” describes past behaviour in aggregate.Coverage, risk, FDR, cost, or human-agreement bounds that hold marginally, or per group, over exchangeable data
Evidence behind itRarely statedCalibration profile, sample size, and last audit date on every guaranteed answer; sets and intervals add an interval for realized coverage
Choosing the thresholdPicked by hand, often by reading a few examplesComputed by a calibration procedure against the risk level you set
Many decisions at onceNo control over the error rate across a batchFalse-discovery-rate control across a batch, including e-BH under arbitrary dependence
After a model updateKeeps being applied, silentlyAnytime-valid monitors raise an alarm with a controlled false-alarm rate
Which modelsTied to one model or one vendor’s single-model confidence APIAny closed API, self-hosted engine, or your own model, with the same guarantee shape
Labelling costNot applicable: nothing is calibratedPrediction-powered inference combines a small human-labelled sample with many judge labels
Where it runsWherever the model vendor serves itHosted API, your own VPC, or a fully offline statistics package

What neither approach gives you is a guarantee about one specific decision in isolation. CLI states every guarantee over the population of exchangeable decisions, and says so on the answer itself.