# Reproduce the descriptive comparison Python 3 standard library only. From a clean folder containing the downloaded files: `python reproduce.py --input published-estimates.csv --output results` Produces category-changes.csv and audit.json; no network request, figures, or raw files required. The input contains 20 selected official NCES published estimates, two years by two ages by five categories. Changes use full returned precision. Rows retain source error flags; zero does not mean zero item nonresponse. No significance or confidence intervals are calculated. Item-specific denominator handling and missing rates are unresolved: do not interpret estimates as all-enrolled prevalence or verified respondent-conditional proportions. Age-specific repeated cross-sections, not transitions. Questionnaire wording matches; context, sampling and participation need not be invariant. Primary 2012–2025 only; 1984 excluded. The official endpoints and query parameters are in sources.json. API data can be revised; snapshot accessed 2026-10-30. Public source-hashes.json records private original bytes for provenance, not a requirement for default reproduction.