# Data audit (2026-10-30)

| file | SHA-256 | lines (wc -l) | notes |
|---|---|---|---|
| data/effects.json | 708c1ff0c4379c228ac6a98f577c23d0d61564f3c958fb35596266aca1f5f10b | 37 (5 top-level items) | ids E1–E5, no duplicates, no missing fields |
| data/README.md | 7cf7cb385f9fb4fc0df52cfc6b0644bcbfb969a38afe2b018b864b7871187dc4 | 16 | provenance |
| data/sources/arxiv-2609.27732v1.txt | ee2d444d…6a8b (matches README) | 134 | DCR; Tables 6, 7, A19; log hospitalisation rate per 100,000 |
| data/sources/arxiv-2606.14621v1.txt | 27c463c0…b0a238 (matches README) | 58 | RESEA; Table 2/3; % and pp |
| data/sources/arxiv-2505.04555v3.txt | b1731a9b…95c1 (matches README) | 66 | MW; Table 2 elasticities, % MW changes 0.047 |
| data/sources/arxiv-2510.03658v1.txt | f8782937…881127 (matches README) | 91 | Benin; Table 3 LPM, mean dep. var. 0.261 |

All hashes match README.md and bundle.json. Not tabular data, so row/column/date-range concepts reduce to the table rows located (see out/analysis.py printed lines).

## Anomalies (full list in results.json `data_issues`)
- DCR: abstract "fatal fall by 35%" vs Section 6.1 "~30%" vs Table 7 coefficient -0.354 (=-29.8%).
- DCR: SE of the 5 km ATT 0.210 (Table 6) vs 0.218 (Table A19); A19 note 640 vs 608 control regions; several '% change' cells differ by 0.1 from 100(exp(b)-1).
- DCR: no rate level printed anywhere -> E4 absolute change impossible.
- MW: '% MW changes' 0.047 vs note "about 4%".
- Benin: text/abstract 0.0855 refers to Table 2 (not in extract); footnote rendering duplication.
- RESEA: control mean is raw (not block-adjusted); 'p ¡ 0.001' rendering defect.

## Information availability
No input violates the rules. The only risk is E4: a baseline rate would have to come from outside the extract, which is forbidden -> declared not harmonisable.