For 167 mouse cortical connections, other outputs from the same neuron gave no clear added gain in fitted delay prediction

We tested whether measured delays from other outputs of one neuron help assign timing to a local neural graph. We read all 123,506 directed pair records in Allen Synaptic Physiology release r2.1. The eligible sample had 1,505 mouse and 151 human chemical connections. The main test compared squared log delay differences within exact recording and cell class blocks. For 109 informative mouse pairs with a common source, the ratio to a block shuffle expectation was 0.817. Its 95% donor range was 0.613–0.988, but its conditional probability after two-test correction was 0.172. We then withheld each tested delay. For 167 mouse connections from 55 donor groups, a weight selected with other donors combined the recording and class mean with the other output delays. Squared error fell by 3.68%. The 95% donor bootstrap range, with weight selection repeated, was -3.77% to +18.44%. Copying the other output delays without adjustment increased error by 10.62%. Shared target delays and small human samples gave no clear added prediction gain. Distance, amplitude, response count, and layer checks gave weak support for a general source rule. An independent SQL calculation and a separate predictor reproduced the main results. The package preserves code, source versions and hashes, selected record IDs, aggregate results, figures, and repeat steps. These are selected local slice graphs. Fit errors, shared stimulation, and incomplete cell labels limit biological inference. The result tests one simple rule for adding delays to a graph; it does not show that neuron identity has no timing information.

connectome · synaptic latency · paired recordings · cortical circuits · prediction · reproducibility