Selected tree scores fell by over 25% in reserved MEG and IT fMRI tests in one 15-adult study

We tested the reliability of a four-image diagnostic for additive tree models of human object responses. Development used a public 92-image study. A fixed tree rejection test then used ten reserved rows per modality from a 118-image study. No quartet passed in either modality. A later exploration found large losses from selected reference scores to test scores. We fixed a follow-up before analysis of a separate 78-image study. Five reference rows per modality selected four image-disjoint quartets. Every selected direction held in all five reference people on raw correlation distance and its chord transform. Ten other rows per modality tested the fixed directions. Mean test/reference ratios were 0.277 and 0.329 in early MEG, and -0.099 and -0.113 in IT fMRI. The fixed rule required all four means to fall below 0.75 of their reference values. The shared probability was 0.004883; Holm correction over the original and new claims gave 0.009766. Inference holds reference people and images fixed and assumes independent person scores with the stated location-model conditions. MEG and fMRI share a cohort within each study; no cross-modal person links or concatenation enter the tests. The result concerns diagnostic loss, not a shared biological cause or latent tree geometry. The package supplies the failed test, an error bound, all scores, source versions and terms, code, figures, model checks, calibration and exact source replay.

human neuroscience · MEG · fMRI · object responses · representational geometry · additive tree · four-point condition · selection · independent confirmation · sample splitting · negative results · reproducibility