Removing connections with no size slope information reduced an HC1 variance estimate by 53% and 9% in two mouse cortex models

We tested how exactly fitted connections affect uncertainty in models that relate synapse size to measured neural function. We used two public mouse cortex panels and the fixed designs from SCI e4adb420-6b56-4b47-8599-968a2f6981ad. Each model had separate source and target intercepts, count controls and geometry splines. Iterative removal retained 1,325 of 2,840 MICrONS connections and 1,311 of 1,441 V1 Deep Dive connections. The full and retained size slopes agreed within 1.6e-15. The global HC1 variance factor n/(n-rank) fell by 53.35% and 9.02%. A fixed-design proof gives the full-to-retained ratio n/k. It also shows that new leaf connections can increase this factor without bound while the slope stays fixed. Singleton removal is an established method; the new contribution is this connectome audit and its exact variance benchmark. We ran 5,000 draws per panel and error model. Under shared cell slopes, nominal 95% coverage in MICrONS changed from 98.30% with the full factor to 90.74% with the retained factor. V1DD coverage changed from 96.48% to 95.64%. Exact expectations confirmed that calibration depended on the noise model. A further model with dependence across disjoint pairs gave severe undercoverage. Removal gives an equivalent least-squares problem. It does not establish a correct interval for the real calcium data. We preserve code, a proof, input provenance, all 30,000 draws, exact calculations, figures and repeat instructions. The study concerns two fixed panels from two mice. It does not establish a new physiological size effect.

connectome · structure and function · synapse size · fixed effects · dyadic variance · singleton observations · MICrONS · V1 Deep Dive · uncertainty calibration