A second fitted load pattern raised mean test error in 12 SEEG patients and 16 fMRI adults

We tested whether a second fitted response direction improves working-memory load prediction in separate human studies. Three SEEG patients and three fMRI adults supplied development. Fixed new-endpoint tests used 12 other patients and 16 other adults. The second direction held 32.56% of training load energy in SEEG and 24.58% in BOLD. Adding it raised mean test error by 1.7565% and 1.2720% of domain-mean baseline error. Pointwise 95% person intervals were 1.1024–2.4880% and 0.8903–1.6931%. The original shared positive gain test failed, with Holm p=1. A fixed middle-load curvature test also failed, with Holm p=0.21011 and mean intervals through zero in both modalities. A later audit of both directions of both endpoints gave Holm p=0.01270 for the error increase. This direction audit is exploratory. It is not independent confirmation. Full fits did not pass the baseline gate, so the result does not establish a useful scalar neural code. All fixed checks, positive sensitivities, negative results, synthetic controls, and person records are preserved. Source hash checks, a second arithmetic implementation, and an exact confirmation replay passed. The package supplies code, SEEG arrays, BOLD fetch and extraction steps, provenance, plans, and figures. Exact BOLD trial rows remain external because the derivative repository has no separate license statement. Sparse clinical sampling, distinct tasks and cohorts, source region selection, short fits, and unequal measurement scales limit scope. The result is a cross-modal estimator benchmark, not evidence for a shared neural cause or causality.

working memory · SEEG · fMRI · response patterns · cross-validation · negative results · dimensionality · reproducibility