# Independent reproduction checks

The two externally referenced datasets were fetched by the reference harness and independently checked for exact byte counts and SHA-256 digests. Their contents are published aggregate environmental time series. The archive members are the supplied adjustment programs, a data table, and a README; the original archive code was not executed in this verification.

The three registered analysis modules and parameters match their final counterparts byte for byte. The added summarizer reads newly computed R1/R2 results. Code inspection found that the simulations regenerate the adjusted series, including lag search and backfitting, rather than reading declared simulated results. The initial results directory was empty.

independent_gls.py checks the central statistic separately. It uses direct AR(1) whitening, least-squares solves, and bounded scalar likelihood optimization rather than the supplied covariance-polynomial and cofactor implementation. It shares the adjustment port, but also compares that port against the published adjusted series after the documented constant offset. Across five series, the maximum difference between corresponding fitted changepoint statistics is less than 0.000001. The published adjusted-series deviation is below 0.0001 degrees after offset correction. See independent_gls.json. This verifies numerical agreement of the specified estimator, not equality to a full-Hessian standard error or validation of the physical adjustment assumptions.

check_simulation_counts.py independently reaggregates all saved simulation groups from the fresh output CSV using threshold counts and the complementary error function. It permits at most one draw of discrepancy because the saved CSV and observed-statistic JSON use six significant digits. Its output records actual discrepancies, so the allowance is not a claim of bitwise agreement.

The node's integrity checks found no flags, and its scan found no hidden content in 27 text files. The code, paper, parameters, provenance, result schemas and input data inspected concern statistical simulations of aggregate climate series. No private records, credentials, mass-harm capability, or identifiable unlicensed copied work was found. External public data are referenced with declared licenses rather than redistributed in the bundle.

These checks concern the submitted computations. They do not establish that a fitted noise process is physically complete, that the five observational records are independent confirmations, or that conditional simulation p-values integrate uncertainty over every modeling choice.

All 30,000 saved full-pipeline draws were independently reaggregated across 15 groups of 2,000. The largest absolute difference from the saved R2 summary was 4.6876561718423027e-07, consistent with display rounding; no one-draw allowance was needed. All 37 numerical bindings across C1-C4 reproduced.