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

## Question
A literature review makes six assertions (A1–A6). Each one cites either Armendariz (2026, arXiv:2609.27732v1, drug consumption rooms, DCR) or Li, Pearson-Merkowitz & Yokum (2026, arXiv:2606.14621v1, RESEA). Using only the cited extract, I classify each assertion as supported, not_supported or not_assessable. Each verdict comes with one locator and a one-sentence reason (`verdicts`, `locators`, `reasons`).

## Data and exclusions
I read four inputs; their line and item counts are in `n_input_rows`. They are `assertions.json` (6 items), `README.md` (14 lines) and the two extracts (134 and 58 lines). The SHA-256 hashes of both extracts match the values in the README (`out/state/data_audit.md`). No assertion was excluded (`out/exclusions.csv`: 6 − 0 = 6), and all six were assessed (`n_analysed` = 6). The task was completed in full (`feasibility` = "complete").

I found these problems in the sources (`data_issues`):
- In the DCR paper, Table A19 prints the CS-DR 5 km ATT with SE 0.218, while Table 6 gives SE 0.210 for the same specification.
- The Table A19 note says there are 640 control regions; everywhere else the paper says 608.
- The abstract says fatal hospitalisations fall by 35%, but Table 7's coefficient of −0.354 implies −29.8%, which matches the "≈30%" in Section 6.1.
- The Figure 7 note mentions four sub-categories, but Table 7 has five.
- The RESEA results text prints confidence intervals for reductions as unsigned magnitudes, and has a rendering defect ("p ¡ 0.001").

## Method
`out/analysis.py` locates the deciding table rows in the extracts by matching strings and prints those lines. It then applies the conventions set by the task:
- percentage change = 100·(exp(b) − 1);
- 95% CI = b ± 1.96·SE, with ±2.0·SE as a check;
- for significance, the printed stars.

The verdict rules are written explicitly in the script and involve no randomness (`random_seed` = null). As a simple baseline, I also checked whether every decimal number in an assertion appears anywhere in the cited extract (`baseline_verdicts_number_match`). This baseline agrees with the final verdicts on 4 of 6 assertions (`baseline_agreement_with_final`). It gets A2 wrong because it ignores the sign, and A5 wrong because that assertion contains no numbers.

## Results
| id | verdict | locator | decisive element |
|---|---|---|---|
| A1 | supported | DCR Table 6, row "ATT", col "5 km" | −0.431** (0.210); ** means p<0.05; the printed % change is −35.0%, and 100(exp(−0.431)−1) = −35.01% (`derived_quantities.A1_pct_change_5km`) |
| A2 | not_supported | DCR Table 7, row "HIV", col "1 km" | the printed ATT is **+0.063** (SE 0.801), not −0.063, so the claimed fall is contradicted |
| A3 | not_supported | DCR Table 6, row "% change", col "3 km" | −0.283 log points is −24.7% (printed), or −24.65% when computed (`derived_quantities.A3_pct_change_3km`), 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 zero (`derived_quantities.A4_ci95_1.96`); the asserted interval equals ±1 SE (`derived_quantities.A4_ci_pm_1se`) |
| A5 | not_assessable | not reported in arxiv-2609.27732v1 | the extract has no results by sex (no line mentions men, women, sex or gender) |
| A6 | supported | RESEA Table 2, row "Weeks on UI", col "Average treatment effect estimate" | −1.99 weeks, CI [−2.24, −1.75], control mean 17.3; −1.99/17.3 = −11.50% (`derived_quantities.A6_pct_of_control_mean`) |

Counts: `n_supported` = 2, `n_not_supported` = 3, `n_not_assessable` = 1.

## Uncertainty
The verdicts are deterministic judgments, not statistical estimates. Here is which ones depend on a convention, and whether a reasonable alternative would change them:
- **A1** depends on the significance threshold. With Table 6's SE of 0.210, |t| = 2.05, which passes both the 1.96 and the 2.0 critical values; ** is printed as well, so the verdict is robust. If a reader used Table A19's SE of 0.218 for the same estimate, |t| = 1.98 would pass 1.96 but not 2.0. The paper still prints ** there, and the assertion quotes SE 0.210, which matches Table 6. The % change does not depend on any convention, because −35.0% is printed.
- **A3** depends on the log-to-percent conversion. Under the exp convention and in the source's own printed row, the change is −24.7%. The figure 28.3% would only come from treating log points as percent, which is an approximation the task rules out and which is also inaccurate at this size. The verdict is robust.
- **A4** depends on the critical value. With 1.96 the interval is [−0.896, 0.084]; with 2.0 it is [−0.906, 0.094] (`derived_quantities.A4_ci95_2.0`). Both include zero, and no significance star is printed, so the verdict is robust. The asserted interval matches only a ±1 SE (≈68%) band.
- **A2, A5 and A6** depend on no convention. A6's 11.5% is printed in the source and also follows arithmetically.

## Limitations and open issues
- The verdicts rely only on the extracts. For example, A5 might be assessable from omitted sections of the full paper, but those are inadmissible here.
- The DCR paper is internally inconsistent in places (the SE in A19, the control-region count, the abstract's fatal-outcome percentage). None of these changes a verdict, but they are reported in `data_issues`.
- For A6, "selection into RESEA" is taken to match the paper's ITT/ATE of RESEA selection.