Degree control left a borderline reciprocal count and joint feature control failed a mixing check in one MICrONS panel

We tested seven reciprocal neuron pairs after exact control of observed input and output counts in one MICrONS mouse. The panel had 1,722 pairs with opportunity in both directions, 145 cells, and 148 connections. A conditional graph model removes additive cell effects without fitting their coefficients. The degree-only sampler gave mean 3.59 reciprocal pairs and an upper tail of 0.047. Its batch-based Monte Carlo standard error was 0.0036, so simulation uncertainty spans 0.05. A second null also fixed edge counts in 27 joint anatomy and movie-response groups. It failed a mixing check after four chains of ten million proposals: chain tail estimates ranged from 0.013 to 0.154 and count R-hat was 1.106. Its pooled values cannot support a stationary-null claim. The biological excess remains unresolved. This extends an earlier fitted-probability check of the same pairs, not an independent replication. One mouse, seven events, selected cells, model assumptions, detection error, and sampling limits prevent generalization. Code, selected data, source hashes, sampling records, and a conditional-model proof support inspection.

MICrONS · reciprocity · conditional graph model · degree control · null result · Allen Institute