# F4-conf-02 — Are the six assertions supported by their cited sources?

## Question
Six review assertions (A1–A5 cite Armendariz 2026, arXiv:2609.27732v1, on Swiss drug consumption rooms; A6 cites Li, Pearson-Merkowitz & Yokum 2026, arXiv:2606.14621v1, on Rhode Island's RESEA programme). For each one I decide, using only the cited extract, whether it is `supported`, `not_supported` or `not_assessable`, and give a locator and a one-sentence reason (`verdicts`, `locators`, `reasons`).

## Data and exclusions
I read four inputs: `assertions.json` (6 items), `README.md` (14 lines), the DCR extract (134 lines) and the RESEA extract (58 lines) (`n_input_rows`). The SHA-256 of each extract matches the README (`out/state/data_audit.md`). No assertion was excluded (`out/exclusions.csv`: 6 → 6), so all six were assessed (`n_analysed` = 6). I found five internal inconsistencies or rendering defects in the sources. None of them changes a verdict (`data_issues`):
- the same CS-DR 5 km ATT has SE 0.210 in Table 6 but 0.218 in Table A19;
- Table A19 gives 608 control regions, but its note says 640;
- the abstract says fatal outcomes fall by 35%, but Table 7's −0.354 implies −29.8% (Section 6.1 says about 30%);
- two of the A19 percentage changes differ by 0.1 pp from 100·(exp(b)−1) of the printed coefficient;
- the RESEA text prints the CI as magnitudes ([1.75, 2.24]) and contains the defect "p ¡ 0.001".

## Method
`out/analysis.py` parses the relevant table rows from the extracts and prints each source line it relies on. It then checks every element of each assertion (sign, magnitude, SE or CI, significance, population) against those values, using the conventions in TASK.md: percentage change = 100·(exp(b)−1), and 95% CI = b ± 1.96·SE (with 2.0 as a sensitivity check). The verdict logic is written explicitly in the script. As a naive baseline, I also computed the share of numbers in each assertion that appear verbatim in the cited extract (`baseline_share_numbers_found_verbatim`). This baseline cannot detect wrong signs or wrong inferences: A2 scores 1.0 and is still false. No randomness is involved (`random_seed` = null).

## Results
Overall: 2 supported (`n_supported`), 3 not supported (`n_not_supported`), 1 not assessable (`n_not_assessable`).

| id | verdict (`verdicts`) | locator (`locators`) | decisive element (`reasons`) |
|---|---|---|---|
| A1 | supported | DCR Table 6, ATT, 5 km | −0.431** (0.210); ** means p<0.05; printed % change −35.0% |
| A2 | not_supported | DCR Table 7, HIV, 1 km | the printed ATT is 0.063, which is positive, not −0.063 |
| A3 | not_supported | DCR Table 6, % change, 3 km | −0.283 log points = −24.7% (printed; computed −24.65%), not −28.3% |
| A4 | not_supported | DCR Table 7, Hepatitis, 5 km | −0.406 ± 1.96·0.250 = [−0.90, 0.08], which includes zero; the asserted interval is b ± 1 SE |
| A5 | not_assessable | not reported in arxiv-2609.27732v1 | the extract has no estimates by sex |
| A6 | supported | RESEA Table 2, Weeks on UI, ATE | −1.99 weeks, CI [−2.24, −1.75], control mean 17.3; −1.99/17.3 = −11.5% |

## Uncertainty
- **A1** depends on the critical value and the log-to-percent conversion. Using SE 0.210, |z| = 2.05 > 1.96, and the value also exceeds 2.0, so the verdict does not depend on the critical value. Table A19's alternative SE (0.218) gives |z| = 1.98. That still clears 1.96 but would fall just short of 2.0. The printed `**` and the paper's statement "significant at the 5% level" decide the significance element in any case. The conversion is robust: 100·(exp(−0.431)−1) = −35.01%.
- **A3** depends on the conversion. The source prints the conversion (−24.7%), and treating log points as percent (−28.3%) is exactly the error in the assertion. No reasonable convention produces 28.3%.
- **A4** depends on the critical value. With 1.96 the CI is [−0.896, 0.084]; with 2.0 it is [−0.906, 0.094]; even a 90% CI (1.645) gives [−0.817, 0.005]. All of these include zero. Only ±1 SE reproduces the asserted [−0.66, −0.16], and that is not a 95% interval. The verdict is therefore robust.
- **A2, A5 and A6** do not depend on any convention: A2 turns on a sign, A5 on missing data, and A6 on values printed in the table plus simple division.

## Limitations and open issues
- The verdicts rest only on the extracts. Some sections are omitted from them, including Section 3.2 of the RESEA paper (heterogeneous effects) and DCR Sections 1–5 and 7. A sex breakdown might exist in the full DCR paper, but under the task rules A5 is `not_assessable`. This is not evidence that the assertion is false.
- The SE discrepancy between Table 6 and Table A19 for the same estimate is unresolved in the source. It does not change A1 under the 1.96 convention.
- The task is complete (`feasibility` = "complete").