# F4-conf-02 — Are the six assertions supported by the sources they cite?

## Question
For six assertions in a draft literature review, decide from the cited extract alone whether each is supported, not_supported or not_assessable, and give a locator and a one-sentence reason for each (keys `verdicts`, `locators`, `reasons`).

## Data and exclusions
Inputs: `data/assertions.json` (6 assertions), the extracts `data/sources/arxiv-2609.27732v1.txt` (Armendariz 2026, Swiss drug consumption rooms; 134 lines) and `data/sources/arxiv-2606.14621v1.txt` (Li, Pearson-Merkowitz & Yokum 2026, Rhode Island RESEA; 58 lines), and `data/README.md` (key `n_input_rows`). The SHA-256 hashes of the extracts match the README. No assertion was excluded: all 6 were assessed (`n_analysed` = 6; `out/exclusions.csv`). Internal inconsistencies found in the sources are listed in `data_issues`. None of them affects a verdict.

## Method
`out/analysis.py` finds the deciding table rows in the extracts, prints them, parses the printed values and applies the verdict rules from TASK.md. Percentage changes are computed as 100·(exp(b)−1) and 95% CIs as b ± 1.96·SE (2.0 as an alternative). An element counts as supported only if the printed number matches it, or follows from it arithmetically. For A5 a regex search for any sex/gender breakdown was run on the cited extract. Before the real analysis, a naive baseline that calls every assertion "supported" was recorded (`baseline_naive_all_supported_verdicts`). No randomness is used (`random_seed` = null).

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

| id | verdict (`verdicts`) | locator (`locators`) | decisive element (`reasons`, `derived`) |
|---|---|---|---|
| A1 | supported | arxiv-2609.27732v1, Table 6, ATT / % change, 5 km | Printed ATT is −0.431** with SE (0.210) and change −35.0%. 100(exp(−0.431)−1) = −35.01% (`derived.A1_pct_from_log`); \|z\| = 2.05 (`derived.A1_z`). |
| A2 | not_supported | Table 7, HIV, 1 km | The printed value is +0.063, not −0.063, so the sign and the claimed fall are wrong. |
| A3 | not_supported | Table 6, % change, 3 km | The implied change is −24.7% as printed (−24.65% computed, `derived.A3_pct_from_log`), not 28.3%. The assertion treats log points as percent. |
| A4 | not_supported | Table 7, Hepatitis, 5 km | −0.406 ± 1.96×0.250 = [−0.896, 0.084] (`derived.A4_ci95_1p96`), which includes zero, and the estimate carries no stars. The claimed [−0.66, −0.16] equals ±1 SE (`derived.A4_ci_pm1se`). |
| A5 | not_assessable | not reported in arxiv-2609.27732v1 | The extract has no estimates by sex (0 matching lines, `derived.A5_sex_term_hits`). |
| A6 | supported | arxiv-2606.14621v1, Table 2, Weeks on UI, ATE column | Printed ATE is −1.99 weeks with 95% CI [−2.24, −1.75] and a control mean of 17.3. 1.99/17.3 = 11.50% (`derived.A6_relative_pct`), which matches the text. |

The naive baseline agrees with the final verdicts on only 2 of the 6 assertions (A1 and A6).

## Uncertainty
These verdicts are categorical, so no sampling uncertainty applies. Three verdicts depend on a convention:
- **A1 (critical value and SE source).** The 5% significance claim rests on the printed stars, and \|z\| = 2.05 > 1.96 confirms it. With a critical value of 2.0 the result still holds (2.05 > 2.0). Appendix Table A19 prints SE 0.218 for the same estimate. With that SE, \|z\| = 1.98 (`derived.A1_z_with_A19_se_0218`): still significant at 1.96, but just short of 2.0. A19 also prints ** for it, and the assertion quotes the Table 6 SE, so the verdict stays supported. This is the most convention-sensitive verdict.
- **A3 (log-to-percent conversion).** Under the prescribed exp conversion the change is −24.6% to −24.7%. Only by reading log points directly as percent would one get 28.3%, and the printed "% change" row rules that out. The verdict does not change.
- **A4 (critical value).** The CI includes zero with 1.96 ([−0.896, 0.084]) and with 2.0 ([−0.906, 0.094], `derived.A4_ci95_2p0`). The verdict does not change.
- **A2, A5 and A6** need no convention. They rest on printed values only, plus the simple ratio for A6.

## Limitations and open issues
- The verdicts rely only on the extracts. For A5, a sex-specific analysis may exist in parts of the paper that were omitted from the extract, but by rule it is not available here.
- The DCR source contradicts itself in several places (`data_issues`): the SE is 0.210 in Table 6 but 0.218 in Table A19; the control-region count is 608 in the tables but 640 in the A19 note; the abstract says fatal hospitalisations fall by 35%, while Table 7 implies −29.8%; and the 3 km change is printed as −24.7% while −24.65% is computed. The RESEA text gives CIs as positive magnitudes for a reduction and contains the rendering defect "p ¡ 0.001". Tables were treated as authoritative. None of these issues changes a verdict.
- The task is complete (`feasibility` = "complete").