Other outputs of the same neuron gave no clear strength prediction gain in shrinkage tests of 934 human layer 2–3 connections

We tested whether measured connection strengths can help assign missing strengths to a human cortical circuit. The primary sample had 1,040 positive EPSP amplitudes from 23 donors in released human temporal cortex layer 2–3 recordings. In 618 edge pairs from 22 donors, connections with a common source had squared log strength differences 21.3% below a shuffle within each recording. The ratio was 0.787, with a 95% donor range of 0.649–0.952. Removal of the weakest tenth of responses gave a ratio of 0.974, with range 0.856–1.095. Raw amplitude and rank tests gave little evidence for extra source similarity. We then hid each tested strength and selected shrinkage weights with other donors only. In 934 connections with a source or target peer, other outputs of the same neuron reduced mean squared log test error by 1.83%. A donor bootstrap that repeated weight selection gave a 95% gain range of -3.05% to +4.15%. Direct use of the source mean increased error by 21.52%. A small tumor cohort also gave no clear gain. These results show a metric-sensitive source association, but no reliable gain for the tested completion rules. Selected slice graphs, missing responses, coarse cell labels, and shared measurement errors limit inference. The package contains exact source tables, provenance, code, predictions, resample draws, figures, independent checks, and repeat instructions.

human cortex · synaptic strength · connectome · structure and function · EPSP · graph completion · shared source · donor validation