# Data audit (P1)

| file | SHA-256 | lines (wc -l) | matches README/bundle hash |
|---|---|---|---|
| data/README.md | 16ebdff389eeaee9508fe81c4edc43690f31f3463c469680c814134086ef9929 | 10 | yes (bundle.json) |
| data/sources/arxiv-2609.27732v1.txt | ee2d444d2f78f2af2ca2cf69b59a61fcd7ac5e10ea29e8ddcd67917bf07d6a8b | 134 | yes |
| data/sources/arxiv-2606.14621v1.txt | 27c463c0a144d9d5a3916dc641f15b6029f6c68b5e574ad9074e7df1aeb0a238 | 58 | yes |

Schema: plain-text extracts, UTF-8, LF line endings (0 CR characters). Tables rendered one row per line, cells separated by " | ".
- DCR (arxiv-2609.27732v1): Table 6 (3 columns 1/3/5 km; rows ATT + SE line, % change, Treated regions, Control regions, Observations), Table 7 (5 outcome rows each with SE line; 3 buffer columns), Table A19 (4 estimators, 5 km). Units: log points (ATT), region-years (Observations; unit not printed, per TASK). Sample period 1998–2022.
- RESEA (arxiv-2606.14621v1): Table 2 (6 outcomes × Treatment, Control, T-C difference, ATE; ATE cells hold "estimate; p; 95% CI: [lo, hi]"), Table 3 (3 outcomes × 3 estimators). Units: USD/yr, %, pp, weeks. Trial Feb 2022–Sep 2023, N = 23,549.
- Missing values: none in the requested cells. Table 7 "Control regions" row has empty 3/5 km cells (merged cell). Duplicates: Table 3 col 1 duplicates Table 2 ATE column (by design); Table A19 CS-DR duplicates Table 6 5 km ATT but with a different SE (see anomalies).

Anomalies (handled in results.json `data_issues`):
1. Table 6 SE (0.210) vs Table A19 CS-DR SE (0.218) for the same ATT −0.431 → use Table 6 as the key names it.
2. Control regions 608 (Tables 6/7, Fig. 6 note) vs 640 (A19 note); Observations / (treated + 640) = 25 exactly for all buffers, not with 608 → flagged; Observations reported as printed.
3. Abstract "fatal … fall by 35%" vs Table 7 5 km fatal −0.354 (≈ −29.8%); HIV called "close to zero" though −0.340 at 5 km → narrative not used.
4. RESEA text weeks CI as positive magnitudes vs Table 2 negative bounds → Table 2 used.
5. Rendering defect "p ¡ 0.001" in RESEA text → not used.

Information-availability: no input violates the rules; both extracts are the admissible evidence. No exact p-value is printed anywhere in the DCR extract (search for "p=" returns 0 lines) → dcr_hiv_att_5km_p_value = null.