# Contract — F4-conf-02

Question. For six literature-review assertions (data/assertions.json), each citing one of two arXiv extracts, decide from the cited extract alone whether each is supported / not_supported / not_assessable. Estimands. Six categorical verdicts, each with one locator and one-sentence reason. Meaningful under: extracts are the only evidence; log-point→% via 100(exp(b)-1); 95% CI = b ± 1.96·SE (2.0 acceptable); partly wrong ⇒ not_supported; not_assessable only when the extract has no bearing quantity. Outputs. out/results.json keys: verdicts, locators, reasons (objects A1..A6), n_supported, n_not_supported, n_not_assessable (int), task_id="F4-conf-02", feasibility, feasibility_reason, data_issues (list of {issue,handling}), n_input_rows ({file: wc -l lines or JSON items}), n_analysed, random_seed=null. Also out/report.md (6 sections), out/analysis.py (prints relied-on source lines, regenerates results.json), out/claims.json. Information rules. No memory of papers / outside knowledge; no modification of data/; values reported as printed. Invalidity. Using outside info; verdicts not traceable to extract lines; missing extracts.