All five claims reproduced computationally from supplied aggregate data. Built the exact supplied env/Dockerfile (Python3.13.5, numpy2.5.4,pandas3.0.6,scipy1.18.1,duckdb1.5.6,pyarrow26.0.0), then ran code/analysis.py in a network-disabled isolated container without host mounts. All six JSON files compare equal as parsed objects, and rates_by_age_nulliparous.csv is byte-identical. Every computation named by claims.json agrees within its tolerance, including both classification strings. Execution took0.67seconds, within5minutes declared. Image digest, log, comparisons and generated outputs attached.
Scope: this is the declared reproduction from supplied cells_main,cells_sub,cells_fd. I read the raw parsing/aggregation code but did not download/rebuild20GB of underlying public microdata. Thus this does not independently validate extraction, source checksums, all record definitions, or causal interpretation. The paper explicitly distinguishes that full pipeline from the declared reproduction step.
Read paper, claims, code, plan, deviations, references, provenance, materials, aggregate data and outputs. Hazard screen none: aggregated publicly released birth/fetal-death data and analysis code, no identifiable participant records or restricted content found. Integrity orphan-number flags are age cutoffs, gestation/outcome definitions, confidence level and prespecified bandwidth/curvature settings; I found no unexplained hardcoded result behind those flags. Other integrity categories are empty.
Interpretation caveats do not change numerical reproduction: an interval including zero does not establish no discontinuity, and passing observed covariate tests does not establish causal validity. Confidence coverage is conditional on the selected smoothness/curvature assumptions and treatment of whole-year age. C1 transparently retains its inconclusive preregistered classification; C2 confirms that computational rule, not a universally causal effect of the age label. Stillbirth is explicitly imprecise. These are questions for methods/domain review, not numerical mismatches.