# Independent reproduction checks

The analysis program and registered program are byte-identical. All 33 externally referenced NHANES files were fetched from the CDC public-use distribution; each byte count and SHA-256 digest was independently checked after the harness staged it. input-verification.json records those checks.

independent_points.py reconstructs the analytic cohort and all five measures directly from the transport files without importing the submitted program. It asserts one-to-one participant joins. It estimates partial Spearman correlations by inverting correlation matrices of ranked variables rather than by separate residual regressions. The 79 resulting sample counts, missingness counts, primary point estimates, differences and secondary item correlations agree with the declared values at their four-decimal precision. The eligible cohort has 14,722 adults; H1 has 14,483 complete cases, and H3 has 13,643. Each analysis-specific sample size is retained in the output. This independent check does not itself recompute bootstrap intervals.

The initial unchanged launch was stopped after 82 seconds because numerical-library thread overhead consumed about ten CPU-seconds per elapsed second on the small linear algebra solves. It produced no result file and was not used to judge the claims. The fresh full rerun sets OPENBLAS_NUM_THREADS, OMP_NUM_THREADS and MKL_NUM_THREADS to 1; code, data, seed, bootstrap count, package versions and model specification remain unchanged. The initial interrupted environment record is preserved for transparency. The final harness report records the completed run and its command.

The node reported no integrity flags, and the harness found no hidden content in all 13 text files. The paper, claims, both copies of the analysis code, plan, external-data manifest, requirements, results, provenance, references and materials were inspected. These contain a statistical analysis, public-source pointers, and aggregate results, with no identified hazard. The [CDC overview](https://wwwn.cdc.gov/nchs/nhanes/continuousnhanes/overview.aspx?BeginYear=2005) explains that confidential and administrative data are excluded and variables are recoded to protect participants. The work points to these holder-released public-use files rather than attempting to de-identify private records. The [physical-function codebook](https://wwwn.cdc.gov/Nchs/Data/Nhanes/Public/2005/DataFiles/PFQ_D.htm) and [depression codebook](https://wwwn.cdc.gov/Nchs/Data/Nhanes/Public/2005/DataFiles/DPQ_D.htm) document the relevant variables.

Reproduction establishes that the submitted computations yield the declared results. It does not validate causal or system-level interpretations, treat a nonsignificant association difference as equivalence, or establish nationally representative inference. In particular, the submitted analysis uses unweighted observations and individual-level bootstrap draws, whereas NHANES uses a complex sampling design. Those choices are retained, not silently replaced, in this numerical verification.

The full run finished successfully in 5 minutes 15 seconds. Its entire R1 JSON object, including all bootstrap intervals and p-values, is equal to the declared object. All 22 result bindings across C1-C3 reproduce, and the independent point-estimate checks also agree with the freshly generated results.