CK-14 Data leakage audit (train-test contamination check; model information sheet)
Confirm that no information from the test items or the outcome entered the model, prompt or features, and produce a leakage checklist and a model information sheet.
What it does
Confirms that no information from the test items or the outcome entered the model, prompt or features. Artefact: a leakage checklist and a model information sheet. Read which of the eight leakage types applies.
When to use it
Use when a predictive or classification claim is about to be reported and you need to know whether test information leaked into development.
When it does not fit
Fits any predictive or classification claim. Misfit: prompt examples drawn from the evaluation set. Pre-training exposure is a further route that was not confirmed.
Result format
The result is a leakage checklist with one line per leakage type, marking whether it applies, together with a model information sheet. It names the leakage types that apply and says what information entered the model, prompt or features.
Requirements on the material
| Property | Requirement type | Requirement |
|---|---|---|
| Content type | Always applies | requires=text; requires=tabular |
| Unit of analysis | Always applies | requires=dataset |
| Level of detail available | Always applies | requires=record-level |
| Appraises source credibility | Applies in some cases | no |
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 material | a documented split, the modelling code or prompt, and the feature list |
|---|---|
| Additional sources needed | none |
| Not for (rules) | none |
| Not for (text) | Wrong choice when prompt examples were drawn from the evaluation set and cannot be replaced. Exposure during model pre-training is a further way that this audit cannot report. |
Choosing and ordering
| Kind of technique | Verification technique, runs in step 6 |
|---|---|
| Phase tag | validate |
| Other phase tags | analyse |
| 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 | return 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.
| Stakes | Check | Person involved | Counts as a different method |
|---|---|---|---|
| exploratory | spot check | optional | no |
| decision-informing | spot check | optional | no |
| decision-informing | leakage check | none | no |
| decision-informing | provenance log | none | no |
| high-stakes | spot check | optional | no |
| high-stakes | leakage check | none | no |
| high-stakes | provenance log | none | no |
| high-stakes | independent review | needed | no |
| high-stakes | expert review | needed | no |
Failure modes and limits
| Model suitability | human judgement required |
|---|---|
| Typical model failure | Cannot report what it saw in pre-training. |
| Error sources | input data; data transformation; coder or rater judgement |
| Certainty limits | Certainty drops when the development steps are not documented (risk of bias) and when pre-training exposure cannot be ruled out (indirectness). |
Related techniques
- should follow MM-07 Supervised classification and dictionary methods (text classification; fine-tuned classifier; lexicon method): a trained classifier is the usual subject
- should follow MM-08 LLM coding and zero-shot classification of text (LLM annotation; LLM-assisted deductive coding; few-shot classification; natural language inference classification): prompts with examples can leak evaluation items
- pairs with CK-05 Out-of-sample validation (train and test split; k-fold cross-validation; bootstrap; permutation test): a clean split is a precondition for out-of-sample testing
Evidence and tag basis
Documented. Leakage errors affected 329 papers across 17 fields, and none could be detected by reading the papers alone (abstract, 2022). (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.