# F4-conf-02 — Is each quoted assertion supported by the source it cites?

## Scientific question
A literature review drafted from two open-access studies contains six assertions, each citing one of the two sources provided. For every assertion, decide from the cited extract alone whether it is **supported** (every element of the assertion — sign, magnitude, uncertainty statement, significance claim, population — is stated in, or follows arithmetically from, the source), **not_supported** (at least one element is contradicted by the source or does not follow from it), or **not_assessable** (the source does not report the quantity or comparison needed to judge the assertion, so it can be neither confirmed nor refuted from the extract). Each verdict must carry a locator (source id and the table, row/column or section that decides it) and a one-sentence reason. The estimands are the six verdicts.

## Inputs
- `data/assertions.json` — list of the six assertions: {"id", "cites" (source id), "text"}.
- `data/sources/arxiv-2609.27732v1.txt` — extract of Armendariz (2026), *Do Drug Consumption Rooms Reduce Drug-Related Hospitalizations? Evidence from Switzerland*, arXiv:2609.27732v1 [econ.GN], CC BY 4.0: abstract, Section 6 (Results: Table 6 overall ATT by buffer; Table 7 ATT by outcome category and buffer; Section 6.1 Mechanisms) and Appendix Table A19. Outcomes are log hospitalisation rates per 100,000 inhabitants (ATTs in log points; the "% change" rows and the text convert them to percentage changes); standard errors in parentheses, clustered at the MedStat region level; stars: *** p<0.01, ** p<0.05, * p<0.10.
- `data/sources/arxiv-2606.14621v1.txt` — extract of Li, Pearson-Merkowitz & Yokum (2026), *Evaluating the Impact of Rhode Island's Self-Sustaining RESEA Program on Employment Outcomes*, arXiv:2606.14621v1 [econ.GN], CC BY 4.0: abstract, Section 3 Results (Table 2 with group means, treatment–control differences and ATE estimates with p-values and 95% CIs; results text; Section 3.1) and Table 3. Tables 2 and 3 were re-set one row per outcome (values verbatim; see the extract header).
- `data/README.md` — provenance, retrieval times, licences, SHA-256 of the extracts.

## Information-availability rules
- The extracts are the only admissible evidence. Do not use remembered content of the papers, other versions, or outside knowledge of the programmes.
- A percentage change implied by a log-point coefficient b is 100·(exp(b) − 1); a 95% confidence interval from a printed standard error is the estimate ± 1.96·SE (2.0 is also acceptable). Use these conventions when an assertion requires a conversion the source does not print.
- "Supported" requires every element of the assertion to hold; an assertion that is partly correct and partly contradicted is `not_supported`. Use `not_assessable` only when the extract contains no quantity or comparison bearing on the claim (e.g. a subgroup the source never reports), not when the claim is merely imprecise.
- A locator has the form `<source id> | Table <n> | row "<row>" | column "<col>"` or `<source id> | Section <n or title>`; one locator per assertion.

## Required outputs
- `out/results.json` with the keys:

| key | type | unit | meaning |
|---|---|---|---|
| `verdicts` | object | — | {"A1": "supported" / "not_supported" / "not_assessable", …, "A6": …} |
| `locators` | object | — | {"A1": "<locator>", …} — the table/section that decides the verdict (for `not_assessable`, the table/section where the relevant outcome is reported, or "not reported in <source id>") |
| `reasons` | object | — | {"A1": "<one sentence naming the decisive element>", …} |
| `n_supported`, `n_not_supported`, `n_not_assessable` | integer | — | counts over the six assertions |
| `task_id` | string | — | "F4-conf-02" |
| `feasibility` | string | — | "complete", "partial" or "infeasible" — refers to the task as a whole |
| `feasibility_reason` | string or null | — | explanation if not complete |
| `data_issues` | array of {"issue","handling"} | — | problems detected in the inputs (rendering defects, internal inconsistencies of a source, etc.) and how they were handled; empty list if none |
| `n_input_rows` | object | lines / items | {file name: number of lines} for every file under `data/` that you read (text files: number of newline-terminated lines as reported by `wc -l`; JSON files: number of top-level items) |
| `n_analysed` | integer or null | — | number of reported items (values, assertions, effects or references) you actually assessed |
| `random_seed` | integer or null | — | null (no randomness is involved) |

- `out/report.md` with the sections: Question; Data and exclusions; Method; Results; Uncertainty; Limitations and open issues. "Uncertainty" must discuss which verdicts depend on a convention (critical value, log-to-percent conversion) and whether the verdict would change under a reasonable alternative.
- `out/analysis.py` — running `~/wp29/venv/bin/python out/analysis.py` from the workspace root must regenerate `out/results.json` from `data/` only, deterministically (the script may encode the verdict logic explicitly; it must print the source lines it relies on).
- `out/claims.json` — a list of the central claims of the report: [{"id": "C1", "text": "...", "results_keys": ["..."], "status": "supported"|"unsupported"|"not_evaluated", "caveat": "..."}].

## Constraints
- No network access. Python: `~/wp29/venv/bin/python` (numpy, pandas, scipy, statsmodels available; the standard library is sufficient for this task). Do not modify `data/`.
- The source extracts under `data/sources/` are the **only** admissible evidence. Do not rely on memory of the papers or on other versions of them; if a quantity is not in the extract, it is not available.
- Report numbers exactly as printed in the source (same sign, same number of decimals); do not round or rescale printed values unless the key definition asks for a transformation.
- This task is designed to be completed in well under 20 minutes of work.