# Verification checks (P4)
| check | outcome |
|---|---|
| Clean re-run of out/analysis.py regenerates identical results.json (SHA-256 75af39dc…25c before and after deleting it) | PASS |
| Source SHA-256 values match data/README.md (asserted in analysis.py) | PASS |
| arXiv id in each entry equals id in its extract header (asserted) | PASS |
| unverifiable_references == ids failing any of the 3 checks | PASS |
| n_verified + len(unverifiable) = 5 = n_analysed = n_raw − 0 excluded (exclusions.csv) | PASS |
| n_input_rows equal `wc -l` counts (17/134/58/107/66/91) and 5 JSON items | PASS |
| Every locator has form "<source id> \| <table/section>" | PASS |
| Arithmetic: exp(−0.431)−1 = −35.0% (R1); −0.0962/0.261 = −0.3686 vs printed −0.3685 (R5, rounding) | PASS (rounding noted) |
| Year rule: R4 entry 2025 = v1 year (v3 is 2026) → consistent; R3 entry 2024 ∉ {2026, 2026} | PASS |
| Baseline (title/year only) flags R3; full method also flags only R3 — consistent in magnitude | PASS |
| Internal inconsistencies in sources (R1 SE 0.210 vs 0.218 etc.) | Reported in data_issues, not silently ignored |