skills/audit-eduinsights-evidence/references/audit-checklist.md

# Evidence audit checklist

## Entity and identity

- Was every named college, field, or occupation resolved before filtering?
- Does every shown identifier retain leading zeroes and its readable label?
- Are CIP and SOC taxonomy levels consistent?
- Is an O*NET specialization being mistaken for its SOC parent?

## Measure and level of detail

- Does the measure answer the person's intended concept?
- Is college-level evidence presented as college-level?
- Is broader CIP-4 or CIP-2 evidence clearly labeled when applied to CIP-6?
- Are median outcomes repeated across narrower rows being counted more than once?
- Is a field-to-career bridge presented only as a preparation path?

## Time and release

- Does every figure have a collection cycle or reference period?
- Is the source release retained rather than replaced by the serving snapshot version?
- Do comparisons use compatible periods and release stages?
- Does a latest grouped result use one shared comparable cycle?
- Are provisional values kept separate from accepted final or revised values?

## Aggregation

- Is the measure additive at the grouping level?
- Were medians, wages, rates, shares, scores, years, or codes summed?
- Were overlapping BLS geography or industry packages combined?
- Does the ranking use the correct award-level code and first-major definition?
- Are counts based on the declared entity grain?

## Coverage

- Is zero treated as an observed number?
- Does every blank retain its supplied reason?
- Is an unmatched DAPIP record being mistaken for no accreditation?
- Are suppressed values kept distinct from not collected or not applicable values?
- Is an unresolved identifier disclosed instead of silently dropped?

## Interpretation

- Are reported degrees being used as proof of a current catalog program?
- Are preparation paths being used as observed graduate destinations?
- Are AI-use observations being converted into job loss, pay, productivity, or time-savings claims?
- Is national context being attached to a specific college, program, or occupation as direct evidence?
- Does causal language exceed the source design?

## Provenance

- Can every material value be tied to a publisher, release, and relation?
- Does the relation source list name the actual publishers, not only EduInsights?
- Is a serving snapshot version correctly used for reproducibility rather than as the measure's year?
- Could another person reproduce the result from the recorded tool calls and filters?