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

# Question

A literature review makes six assertions (A1–A6), each citing one of two arXiv extracts: Armendariz (2026), arXiv:2609.27732v1 (drug consumption rooms, DCR), and Li, Pearson-Merkowitz & Yokum (2026), arXiv:2606.14621v1 (Rhode Island RESEA). For each assertion I decide, using only the cited extract, whether it is supported, not_supported or not_assessable. Each verdict comes with a locator and a one-sentence reason (verdicts, locators, reasons).

# Data and exclusions

I read four files. Line and item counts are in n_input_rows: assertions.json has 6 items, README.md 14 lines, the DCR extract 134 lines and the RESEA extract 58 lines. Their SHA-256 hashes match the values in data/README.md and bundle.json (see out/state/data_audit.md). All six assertions were assessed (n_analysed = 6) and none were excluded (out/exclusions.csv).

I found these internal problems in the sources (data_issues):

Two different standard errors for the same estimate. DCR Table A19 prints SE 0.218 for the same CS-DR ATT −0.431 that Table 6 prints with SE 0.210.
Two different counts of comparison regions. The Table A19 note says 640 never-treated regions, but the tables say 608.
Abstract disagrees with the table on fatal outcomes. The abstract says fatal outcomes fall by 35%. The Table 7 estimate of −0.354 implies −29.8%, and Section 6.1 says "approximately 30%".
Text disagrees with the table on HIV. The text calls the HIV estimates "close to zero", but Table 7 shows −0.340 at 5 km.
Encoding defect. The RESEA text prints "p ¡ 0.001", which I read as p < 0.001.
CI written with the opposite sign. The RESEA text gives the weeks-on-UI CI as magnitudes [1.75, 2.24]; this is the same interval as [−2.24, −1.75] in Table 2.

None of these problems changes a verdict. Where prose and a table disagree, I treated the table as authoritative.

# Method

For each assertion I checked every element (sign, magnitude, uncertainty, significance, population) against the relevant table cell.

Conventions from TASK.md. Percent change is 100·(exp(b) − 1). A 95% CI is b ± 1.96·SE, with 2.0 also checked.
Script. out/analysis.py implements the verdict logic explicitly. It parses the cells it relies on from the extracts, prints those source lines, and writes out/results.json. No randomness is involved (random_seed = null).
Baseline. As a naïve baseline, the script tests whether every decimal number in an assertion appears verbatim in the cited extract (baseline_literal_number_match). This flags A3 and A4. It misses A2, because "0.063" appears in the extract but with the opposite sign, and it cannot test A5, which contains no number. That is why each element has to be checked explicitly.

# Results

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

id	verdict (`verdicts`)	locator (`locators`)	decisive element (`reasons`, `conversions`)
A1	supported	DCR Table 6, row "ATT", column "5 km"	Table 6 prints −0.431** with SE (0.210) and −35.0%; 100(exp(−0.431)−1) = −35.01% (`conversions.A1_pct_change_5km`)
A2	not_supported	DCR Table 7, row "HIV", column "1 km"	The printed ATT is +0.063 (SE 0.801), not −0.063, so the sign and the claimed fall are contradicted
A3	not_supported	DCR Table 6, row "% change", column "3 km"	−0.283 log points is −24.6% (`conversions.A3_pct_change_3km`); Table 6 prints −24.7%, not −28.3%
A4	not_supported	DCR Table 7, row "Hepatitis", column "5 km"	−0.406 ± 1.96·0.250 = [−0.896, 0.084] (`conversions.A4_ci95_196`) includes zero; the claimed interval is ±1 SE
A5	not_assessable	not reported in arxiv-2609.27732v1	The extract has no estimates by sex
A6	supported	RESEA Table 2, row "Weeks on UI", column "Average treatment effect estimate"	−1.99 weeks, CI [−2.24, −1.75], control mean 17.3; 1.99/17.3 = 11.50% (`conversions.A6_relative_reduction_pct`)

# Uncertainty

The verdicts are classifications, not estimates, so they have no sampling uncertainty. Four of them depend on a convention:

A1 (5% significance). Table 6 marks the estimate with ** (p < 0.05), and the t-ratio is 0.431/0.210 = 2.05 (conversions.A1_t_stat_table6). With the Table A19 SE of 0.218 the ratio is 1.98 (conversions.A1_t_stat_tableA19). That still clears 1.96 but falls just short of the alternative 2.0 critical value. The verdict rests on the Table 6 SE and star, which are exactly what the assertion quotes, so it stays supported. Still, this is the one verdict that is sensitive to a source inconsistency combined with a choice of critical value.
A3 (log-to-percent). Under the exp convention the change is −24.6%, and the table prints −24.7%. No reasonable convention produces −28.3% except treating log points as percentages. Even that would only give 28.3 log points, not a 28.3% reduction, so the verdict holds.
A4 (CI width). Both 1.96·SE and 2.0·SE give intervals that include zero ([−0.896, 0.084] and [−0.906, 0.094], conversions.A4_ci95_196 and conversions.A4_ci95_200). The verdict does not change.
A6 (rounding). 1.99/17.3 = 11.50%, so rounding cannot change the verdict.

A2 and A5 do not depend on any convention.

# Limitations and open issues

Evidence limited to the extracts. Verdicts rest only on the extracts; omitted sections, such as RESEA Section 3.2 on heterogeneous effects, were not consulted, as the rules require. A5 is not_assessable only with respect to the DCR extract; the full paper might report results by sex, but that cannot be checked here.
SE discrepancy unresolved. The 0.210 vs 0.218 SE discrepancy in the DCR source cannot be resolved from the extract.
Task status. The task was fully completed (feasibility = "complete").

spb/wp29_uqo