startthat.work home

CK-23 Confidence calibration check (reliability diagram; expected calibration error; temperature scaling; P(True); verbalised confidence)

Check that a stated confidence of 80 per cent matches being right about 80 per cent of the time, rescale confidence if not, and produce a calibration curve and error.

Check extracted resultsEvidence: TestedVerification technique, runs in step 6Strongest statement: describe

What it does

Checks that a stated confidence of 80 per cent corresponds to being right about 80 per cent of the time, and rescales confidence if not. Artefact: a calibration curve and error. Read it before using confidence to route items to human review.

When to use it

Use when confidence scores will be used to select items for human review or to decide which outputs to trust.

When it does not fit

Needs labelled correctness data. Misfit: confidence is a prompt-generated number with no labelled data to test it. Neither study concerns free-text extraction.

Result format

The result has a calibration curve showing stated confidence against the observed rate of correct answers, and a calibration error figure. It states whether confidence can be used to route items to human review and gives the rescaling applied if it was not calibrated.

Requirements on the material

PropertyRequirement typeRequirement
Content typeAlways appliesrequires=text; requires=tabular
Unit of analysisAlways appliesrequires=claim; requires=record
Level of detail availableAlways appliesrequires=record-level
Appraises source credibilityApplies in some casesno

Not restricted on: Structure, Variable types, Time structure, How the sample was chosen, Number of sources, Whether sources can be joined, Amount of material, Known data problems, Producer interest, Source dependence, Origin, Document genre, Language, Personal information, Security classification, Licence.

Minimum materiallabelled correctness data for at least several hundred items
Additional sources needednone
Not for (rules)none
Not for (text)Wrong choice when confidence is a prompt-generated number and there is no labelled correctness data to test it. The studies cited for it do not cover free-text extraction.

Choosing and ordering

Kind of techniqueVerification technique, runs in step 6
Phase tagvalidate
Other phase tagsnone
Kind of findingrisk
Question answeredhow much
Strongest statement it supportsdescribe
Needs as input
Makes as output
Must exist firstNone
Loop targetreturn to gather

Library checks by stakes

Each stakes level includes the checks of the level below it. Checks that need a person apply at high stakes only, so the process leaves them out at lower stakes (decision 38). Repeatability: deterministic.

StakesCheckPerson involvedCounts as a different method
exploratoryinput validationnoneno
decision-informinginput validationnoneno
decision-informingcode execution for arithmeticnoneno
decision-informingspot checkoptionalno
high-stakesinput validationnoneno
high-stakescode execution for arithmeticnoneno
high-stakesspot checkoptionalno
high-stakesgold standard agreementconditionalno
high-stakesindependent reviewneededno

Failure modes and limits

Model suitabilitymodel plus code
Typical model failureVerbalised confidence contains no information, or answers are overconfident.
Error sourcesinput data; language model; statistical assumptions
Certainty limitsCertainty drops when the calibration set differs from the target data (indirectness) and when bins contain few items (imprecision).

Evidence and tag basis

Tested. Modern neural classifiers were poorly calibrated and temperature scaling worked well (2017). Larger models were well calibrated on multiple-choice and true or false questions in the right format, with weaker calibration on new tasks (2022). (Guo et al. 2017, arXiv 1706.04599; Kadavath et al. 2022, arXiv 2207.05221)

Tag basis: input profile inference; selection inference; sequencing inference; robustness inference. All tag values are inference unless a source is named.