Data audit (P1)
| file | SHA-256 | lines (wc -l) | items | matches README/bundle hash |
|---|---|---|---|---|
| data/assertions.json | 2965ab3acc6f326653169b83a5345698b51fac88a556de65faf89210aea0cad4 | 32 | 6 assertions (A1–A6, ids unique, keys id/cites/text, no missing) | yes (bundle) |
| data/README.md | 5348bd49912099735108dbbe19ae09fa72bd600cb3d12b878658e6b9806198a7 | 14 | — | yes (bundle) |
| data/sources/arxiv-2609.27732v1.txt | ee2d444d2f78f2af2ca2cf69b59a61fcd7ac5e10ea29e8ddcd67917bf07d6a8b | 134 | Tables 6, 7, A19 + text | yes (README) |
| data/sources/arxiv-2606.14621v1.txt | 27c463c0a144d9d5a3916dc641f15b6029f6c68b5e574ad9074e7df1aeb0a238 | 58 | Tables 2, 3 + text | yes (README) |
Units: DCR outcomes are log hospitalisation rates per 100,000 (ATT in log points, SE in parentheses, clustered at MedStat region); RESEA outcomes in $, pp, weeks with p-values and 95% CIs. Date ranges: DCR openings 1998–2022; RESEA RCT Feb 2022–Sep 2023. Assertions A1–A5 cite arxiv-2609.27732v1; A6 cites arxiv-2606.14621v1. Both cited sources present.
Anomalies / internal inconsistencies (none blocks the verdicts)
- DCR Table A19 CS-DR column: same ATT −0.431 as Table 6 5 km, but SE (0.218) vs (0.210).
- DCR Table A19 note: "640 regions" never-treated vs 608 control regions in Table 6/7, A19 body, Figure 6 note.
- DCR abstract: fatal hospitalisations "fall by 35%"; Table 7 5 km fatal ATT −0.354 → 100(exp(−0.354)−1) = −29.8%, matching Section 6.1 "approximately 30%", not 35%.
- DCR: text says "five diagnostic sub-categories"; Figure 7 note says "four"; Table 7 note lists ICD codes for four only (fatal not coded).
- DCR results text says HIV estimates "close to zero" although the 5 km point estimate is −0.340 (SE 0.294) — interpretive, not numeric.
- RESEA line 39: rendering defect "p ¡ 0.001" (should read p < 0.001).
- RESEA results text prints the weeks-on-UI CI as magnitudes of reduction "[1.75 weeks, 2.24 weeks]" while Table 2/3 print signed [−2.24, −1.75]; equivalent.
Availability rules
No input violates the rules: both extracts are the admissible evidence; assertions.json is the object of assessment, not evidence. No external information used.
Exclusions
None: all 6 assertions assessed (see out/exclusions.csv).