# Verifier notes

**Re-run.** `python code/analysis.py` in the declared container exited 0 in 4.7 s. All 21 declared results for C1–C5 reproduce within tolerance, including the verdict strings ("inconclusive" at 35 and "confirm" at 40).

**Independent check** (evidence/rd_check.txt). This is my own code, not the bundle's. It fits a births-weighted local-linear discontinuity with a uniform kernel to the nulliparous induction rate by single year of age from data/cells_main.csv:

| Cutoff | Bandwidth 3 y | Bandwidth 4 y | Bandwidth 5 y | Bundle (triangular, h = 3) |
|---|---|---|---|---|
| Age 40 | 2.89 pp | 2.98 pp | 2.96 pp | 2.86 pp |
| Age 35 | 1.49 pp | 1.64 pp | 1.72 pp | 1.19 pp |

The age-40 jump is robust across kernel and bandwidth. The age-35 jump is positive in every specification but more sensitive to kernel and bandwidth (1.2–1.7 pp). That is consistent with the bundle's "inconclusive" verdict at 35, though the bundle reaches that verdict through its falsification check.

**Not checked.** I did not rebuild the cells from the raw NVSS natality files; the full pipeline script is provided but needs those files.

**Screening.** Hazard: none. The data are aggregated age × year × parity cells from public-use US natality and fetal-death files, with large counts and no individual records. No hidden instructions were found.