startthat.work home

CK-25 Error taxonomy and error analysis (failure mode analysis; confusion analysis)

Classify reviewed failures into defined types such as invented fact, wrong attribution, omission, wrong scope and arithmetic, and produce error counts by type.

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

What it does

Classifies reviewed failures into defined types, such as invented fact, wrong attribution, omission, wrong scope and arithmetic, and counts them. Artefact: error counts by type. Read which types dominate, and change prompts, checks or review focus.

When to use it

Use when a pilot or review has found errors and you need to know which kinds dominate so you can change prompts, checks or review focus.

When it does not fit

Use after any pilot. Misfit: overlapping types, so counts do not add.

Result format

The result is a table of error counts by defined type, such as invented fact, wrong attribution, omission, wrong scope and arithmetic, with the failed items listed under each type. It names the types that dominate and says what change to prompts, checks or review focus follows.

Requirements on the material

PropertyRequirement typeRequirement
Content typeAlways appliesrequires=text
Unit of analysisAlways appliesrequires=claim; requires=record
Appraises source credibilityApplies in some casesno

Not restricted on: Structure, Variable types, Time structure, How the sample was chosen, Level of detail available, 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 materiala set of reviewed failures, usually 30 or more
Additional sources needednone
Not for (rules)none
Not for (text)Wrong choice when error types overlap, so that counts do not add up. Types must be defined before counting.

Choosing and ordering

Kind of techniqueVerification technique, runs in step 6
Phase tagvalidate
Other phase tagsanalyse
Kind of findingpattern; risk
Question answeredwhy; what next
Strongest statement it supportsdescribe
Needs as input
Makes as output
Must exist firstNone
Loop targetreturn to analyse

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: judgement dependent.

StakesCheckPerson involvedCounts as a different method
exploratoryspot checkoptionalno
decision-informingspot checkoptionalno
decision-informinghuman review sampleneeded; left out below high stakesno
high-stakesspot checkoptionalno
high-stakeshuman review sampleneededno
high-stakesindependent reviewneededno
high-stakesassumption and limitation statementnoneno

Failure modes and limits

Model suitabilityhuman judgement required
Typical model failureAssigns tidy categories to errors it has not examined.
Error sourcesinput data; coder or rater judgement
Certainty limitsCertainty drops when types overlap (indirectness), when few failures were reviewed (imprecision) and when one reviewer classified them (risk of bias).

Evidence and tag basis

Documented. Large-scale human evaluation found substantial hallucinated content in all neural summarisers tested (2020). An eight-type leakage taxonomy from another field is a worked example. (Maynez et al. 2020, ACL; Kapoor and Narayanan 2022, arXiv 2207.07048)

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