CK-24 Review sampling design (random audit sample; stratified sample; risk-based sampling; acceptance sampling; rule of three)
Choose which outputs to inspect and how many so that the observed error rate has a stated bound, and produce an error-rate estimate with an upper bound.
What it does
Chooses which outputs to inspect and how many, so that the observed error rate has a stated bound. Artefact: an error-rate estimate with an upper bound. With zero errors in n independent random items, the 95 per cent upper bound is about 3 divided by n.
When to use it
Use when you need to audit model output by hand and must decide how many items to review and which.
When it does not fit
Fits rare errors and independent items. Misfit: clustered items, or errors concentrated in a subgroup, which need stratification. No source sets a sample size for LLM output.
Result format
The result states how the items to inspect were chosen, the sample size and the number of errors found. It gives an error-rate estimate with an upper bound, and notes the strata used if any.
Requirements on the material
| Property | Requirement type | Requirement |
|---|---|---|
| Unit of analysis | Always applies | requires=claim; requires=record; requires=document |
| Level of detail available | Always applies | requires=record-level |
| Appraises source credibility | Applies in some cases | no |
Not restricted on: Structure, Content type, 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 material | independent items and a stated error rate that would be unacceptable, with at least 59 error-free items for a 5 per cent upper bound |
|---|---|
| Additional sources needed | none |
| Not for (rules) | none |
| Not for (text) | Wrong choice when items are clustered or errors concentrate in a subgroup, unless sampling is stratified. No source sets a sample size for language model output. |
Choosing and ordering
| Kind of technique | Verification technique, runs in step 6 |
|---|---|
| Phase tag | validate |
| Other phase tags | none |
| Kind of finding | risk |
| Question answered | how much |
| Strongest statement it supports | describe |
| Needs as input |
|
| Makes as output |
|
| Must exist first | None |
| Loop target | proceed with limit stated |
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: repeatable with fixed settings.
| Stakes | Check | Person involved | Counts as a different method |
|---|---|---|---|
| exploratory | input validation | none | no |
| decision-informing | input validation | none | no |
| decision-informing | code execution for arithmetic | none | no |
| high-stakes | input validation | none | no |
| high-stakes | code execution for arithmetic | none | no |
| high-stakes | independent review | needed | no |
| high-stakes | assumption and limitation statement | none | no |
Failure modes and limits
| Model suitability | model plus code |
|---|---|
| Typical model failure | Picks "representative" examples to check, which gives a biased sample. |
| Error sources | input data; source selection; statistical assumptions |
| Certainty limits | Certainty drops when sampled items are not independent (risk of bias) and when errors are rare so the bound is wide (imprecision). |
Related techniques
- pairs with CK-26 Independent human review of model output (expert review; adjudication; human oversight): reviewers inspect the drawn sample
- pairs with CK-25 Error taxonomy and error analysis (failure mode analysis; confusion analysis): errors found in the sample are classified
- pairs with CK-03 Bootstrap confidence intervals for agreement and accuracy (bootstrap resampling of items): bootstrap intervals give an alternative bound
Evidence and tag basis
Documented. The rule of three is read from an encyclopaedia page and is valid roughly for n above 30. The note author's arithmetic gives an upper bound near 5 per cent for 59 error-free items and near 1 per cent for 299. (Wikipedia: Rule of three (statistics); Grimmer and Stewart 2013, Political Analysis; Artstein and Poesio 2008, Inter-coder agreement for computational linguistics)
Tag basis: input profile inference; selection inference; sequencing inference; robustness inference. All tag values are inference unless a source is named.