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

## Question
For six literature-review assertions (A1–A5 cite Armendariz 2026, arXiv:2609.27732v1; A6 cites Li, Pearson-Merkowitz & Yokum 2026, arXiv:2606.14621v1), decide from the cited extract alone whether each is supported, not_supported or not_assessable, with a locator and a one-sentence reason (`verdicts`, `locators`, `reasons`).

## Data and exclusions
Inputs: `data/assertions.json` (6 items), `data/README.md` (14 lines), the DCR extract (134 lines) and the RESEA extract (58 lines) (`n_input_rows`). SHA-256 of both extracts match the README. No exclusions: all 6 assertions were assessed (`n_analysed` = 6; `out/exclusions.csv`). Only the extracts were used as evidence.

Source issues found (`data_issues`): (i) DCR Table A19 prints SE 0.218 for the CS-DR ATT −0.431 that Table 6 prints with SE 0.210; (ii) the A19 note says 640 control regions vs 608 elsewhere; (iii) the DCR abstract says fatal hospitalisations fall by 35%, but the 5 km fatal ATT −0.354 converts to −29.8%; (iv) the text calls HIV estimates "close to zero" although the 5 km ATT is −0.340; (v) RESEA prose prints the reduction CIs as positive magnitudes, and has a rendering defect "p ¡ 0.001". None of these changes a verdict.

## Method
`out/analysis.py` finds the deciding table lines, prints them, and checks each element of each assertion against the printed cells. It uses the task conventions: % change = 100·(exp(b) − 1) and 95% CI = b ± 1.96·SE, with ±2.0·SE as a sensitivity check. As a baseline, it also checks whether every decimal number in an assertion appears verbatim in the cited source (`baseline_naive_number_match_verdicts`). The baseline agrees with the final verdicts on 4 of the 6 assertions (`baseline_agreement_with_final`). It misses the sign error in A2 and the subgroup that is never reported in A5.

## Results
Counts: 2 supported, 3 not_supported, 1 not_assessable (`n_supported`, `n_not_supported`, `n_not_assessable`).

| id | verdict (`verdicts`) | locator (`locators`) | decisive element (`reasons`, `derived`) |
|---|---|---|---|
| A1 | supported | DCR Table 6, row ATT, col 5 km | −0.431** (0.210), % change −35.0%; 100(exp(−0.431)−1) = −35.01% (`derived.A1_pct_from_logpoints`) |
| A2 | not_supported | DCR Table 7, row HIV, col 1 km | printed ATT is +0.063 (SE 0.801): the assertion gets the sign wrong |
| A3 | not_supported | DCR Table 6, row % change, col 3 km | −0.283 log points = −24.7% (−24.65%, `derived.A3_pct_from_logpoints`), not 28.3% |
| A4 | not_supported | DCR Table 7, row Hepatitis, col 5 km | −0.406 (0.250) gives a 95% CI of [−0.896, 0.084], which includes 0 (`derived.A4_ci95_1.96`); the asserted [−0.66, −0.16] is ±1 SE (`derived.A4_ci_1se`) |
| A5 | not_assessable | not reported in arxiv-2609.27732v1 | the extract gives no results by sex |
| A6 | supported | RESEA Table 2, row Weeks on UI, col ATE | −1.99 weeks, 95% CI [−2.24, −1.75], control mean 17.3; 1.99/17.3 = 11.5% (`derived.A6_relative_reduction_pct`) |

## Uncertainty
These verdicts are deterministic judgements, not statistical estimates. Here is which verdicts depend on a convention:
- **A1** depends on the significance claim. With the Table 6 SE, t = −2.05 (`derived.A1_t_stat_table6`). With the Table A19 SE of 0.218, |t| = 1.98, which is still above 1.96 and carries ** in both tables. The verdict therefore holds under both SEs (`data_issues`). Under a strict 2.0 critical value, however, |t| = 1.98 with the A19 SE would just miss. This is the one place where a convention matters close to the threshold. The assertion quotes SE 0.210 from Table 6, and with that SE it passes under either critical value.
- **A3** depends on the log-to-percent conversion. The source itself prints −24.7%. The only way to get 28.3% is to read log points directly as percent, and both the task convention and the printed row rule that out, so the verdict is robust.
- **A4** depends on the critical value. The CI includes zero under 1.96 (`derived.A4_ci95_1.96`) and under 2.0 (`derived.A4_ci95_2.0`). The interval would exclude zero only at about ±1 SE (roughly a 68% interval), which is not a 95% CI. The verdict is robust.
- **A6** involves only the rounded ratio 11.50%, which is also printed in the source text. A2 and A5 depend on no convention.

## Limitations and open issues
- The verdicts rest on the extracts only. A5 is not_assessable because the extract contains no results by sex. The full paper might contain such results, but that would be outside the admissible evidence.
- The DCR source contradicts itself in places: the SE for the main ATT (0.210 vs 0.218), the number of control regions (608 vs 640), and the fatal-outcome percentage in the abstract (35% vs −29.8% by conversion). I decided verdicts on the table cells the assertions quote.
- In A6, "selection into RESEA" matches the source's own wording. Table 3 calls the same estimates ITT, which is consistent with "selection".
- No randomness is involved (`random_seed` = null). The task is complete (`feasibility`).