Reference error scales did not pass shared tests in 20 iEEG patients and 16 fMRI adults

We tested whether errors from reference people improve response prediction sets in other people. Separate public human studies supplied signed iEEG LFP means and BOLD fMRI. Five patients and four MRI adults supplied reference errors. A local code lock preceded tests in 20 other patients and 16 other adults. One rule used class error scales. A second used sound-specific scales. Both recalibrated an 80% map-error radius and kept the same acoustic point forecast. Proper radius loss penalized wide sets and misses. Mean class gains were -0.109% and +3.770%, with 95% person Student limits [-0.234,+0.015]% and [+3.054,+4.486]%. Added item gains were -0.545% and +1.207%. Both shared tests failed Holm correction, p=1. Mean coverage stayed near 81% in both modalities. An added study used eight other MRI adults, the same sounds and unchanged B2021 references. The B2021 item-gain range included zero when reference people and sounds were resampled. All negative people and fixed checks remain. The contribution is a tool to score prediction sets and a failed shared benchmark. It does not establish a common neural mechanism, causality or transfer to unknown sound libraries. Sparse clinical coverage, different cohorts and labels, source screens, unequal site counts and measurement filters limit scope. Code, permitted inputs, source versions and terms, all results, figures and repeat steps are supplied. Original MRI site arrays and audio remain external.

auditory cortex · iEEG · fMRI · prediction sets · model error · quantile loss · cross-modal test