In one MICrONS graph model, site distance reduced a reverse cleft size test gain from 1.32% to 0.062%

We tested whether a known reverse connection size helps predict mean synaptic cleft size in mouse cortex. We read all 303,164,967 connection records in MICrONS v1507 and retained the complete selected graph of 3,876,197 edges among 63,594 putative excitatory soma neurons. A five-fold test excluded both directions of each test pair from fitting. The reverse size was given for each separate completion query. In 38,397 reciprocal pairs with soma distance at most 200 micrometres, reverse size reduced squared log size error by 1.322% after cell, class, count and soma geometry controls. The conditional neuron-weight range was 0.934–1.715%. However, 2,114 pairs with opposite-direction sites within 1 micrometre supplied 100.52% of the net error reduction. The original rule gave -0.008% gain on the other pairs. A follow-up model with site distance calibration left a gain of 0.062%, with a neuron-weight range of -0.003–0.125%. Refitting after removal of nearby pairs retained a smaller 0.238% gain, so separated sizes are not shown to be independent. A secondary test of 274 orientation-selective pairs gave no clear relation between preferred orientation similarity and size mismatch. Raw CSV and cached Parquet site scans agreed exactly on all 113,174 reciprocal sites. The result shows that site geometry can dominate a size completion score. It does not establish matched physiological strength or identify false synapses. One mouse, incomplete axons, predicted classes, selected functional matches and post-plan follow-ups limit interpretation. The package preserves the full selected graph, site records, pair errors and predictions, code, provenance, figures and repeat instructions.

MICrONS · connectome · reciprocal connections · synapse size · site distance · graph completion · structure and function · reproducibility