Full cell type pairs gave no clear gain beyond separate type effects in donor exclusion tests of human layer 2–3 connections and 50 ms pulse ratios

We tested whether a full source and target cell type table adds prediction beyond separate type effects in human cortical layer 2–3. We used 5,960 connection test records and 780 positive 50 ms pulse ratios from 23 donors. Four released electrophysiological labels defined the types. Nested ridge tests excluded each test donor from all fitting and penalty selection. Adding pair terms increased mean connection log loss by 0.38%, from 0.428840 to 0.430489. The pulse-ratio squared log error fell by 1.20%, from 0.132592 to 0.131005. Its descriptive 95% donor range was −0.72% to +3.00%. Ranges held outer predictions fixed and omit fitting uncertainty. Wider penalties, removal of repeated cell identities, omission of DoubSpk cells, complete geometry, and added cell physiology gave no clear pair gain. A later check with separate pair shrinkage gave error decreases of −0.12% for detection and −0.09% for pulse ratio. The initial 5.17% pulse-ratio gain from separate type labels fell to an uncertain 0.65% with a wider penalty range. Independent code reproduced both cohorts and checked the numerical fits. These are selected slice recordings, not a complete anatomical connectome. Sparse type pairs and donor variation limit inference. The result does not establish that biological partner effects are absent. We preserve code, exact source hashes, model states, predictions, figures, checks and repeat instructions. Original source measurements remain external.

human cortex · cell type · connectivity · paired pulse ratio · structure and function · donor validation · synaptic dynamics · ridge regression