Implicit-tuning signal quantified through a tri-experiment design in a legacy spectral GCM

We quantify the climate response of the MGO atmospheric general circulation model (MGO-03 T42L25) to two alternative implementations of a structural fix in the vertical-exchange parameterization. A companion parallelization study (under review) identifies an unintended inter-latitude state leakage through SAVE-persistent working arrays in the boundary-layer scheme. Here we contrast the legacy code (REF) against two candidate repairs—zero-reinitialization at each latitude (CTRL_A) and per-latitude snapshot/restore of the prior state (CTRL_B)—in a three-member initial-condition ensemble of 30-year AMIP-style integrations (1979-2008), and evaluate the simulated climatologies against ERA5, CERES EBAF, GPCP, and ISCCP-H observational datasets. Pairwise comparisons with pooled-paired t-tests across the ensemble, an effective-sample-size correction for temporal autocorrelation, and Benjamini-Hochberg control of the false discovery rate at q = 0.05 reveal that (i) the choice between the two implementations—contrasted at identical compilation settings (CTRL_B-CTRL_A)—is itself FDR-significant in 20.34% of T42 cells in annual surface air temperature, exceeding the CTRL_B-REF leak-elimination signal in both area-weighted amplitude and FDR-significant spatial extent; (ii) both repairs enlarge the model bias in radiation and cloud cover against observations relative to REF, a pattern consistent with the hypothesis that the legacy configuration was implicitly tuned against the leaky baseline; the ensemble establishes the signal's robustness but does not isolate the mechanism among several plausible alternatives. We additionally document that the legacy implementation is not thread-reproducible across OpenMP partitions. A formal bit-reproducibility theorem, identifying three implementation conditions under which OpenMP parallelization of the model is bitwise reproducible by construction, is presented in the Supporting Information.