Keeping ten neuron pairs separate reduced mean class error by 22–26% in two adult C. elegans routing models with recorded panel exits

We tested whether selected left-right neuron identities reduce class error in anatomical routing models. We used 162 neurons in two adult C. elegans maps from Witvliet et al. A two-step error score in one adult selected ten pairs to keep separate in the other adult. For equal class starts and steps two through six, the initial model gave mean total variation errors of 0.025908 and 0.021351 when all pairs were combined. Keeping ten pairs separate reduced these errors by 58.56 and 60.99 percent. This model normalized synapse counts within the neuron panel and held zero-output neurons fixed. A raw-target audit found that most RMEL/RMER outputs go to muscles outside the panel. Keeping RME separate alone removed 47–48 percent of the initial error. We added a model that normalized by all recorded target incidences and sent outside targets to one absorbing EXIT state. The same fixed ten-pair lists then reduced mean errors from 0.006734 to 0.004964 in map 8 and from 0.006914 to 0.005376 in map 7. These gains were 26.29 and 22.25 percent, or 0.177 and 0.154 percentage points. Keeping only RME separate gave less than one percent gain. Of 999 equal-budget subsets per map, six and four reached the selected error or lower under the exit model. Thus selected identities still improved the reduction, but the panel boundary explained much of the initial gain. The exit check followed the initial audit and is exploratory. The two test directions share animals. These unsigned count models do not establish neural signal flow, pair importance, or performance in new animals. State aggregation and memory methods are prior work. The package preserves source hashes, code, numerical inputs, all subset draws, arrays, figures and repeat steps. Raw edge counts rebuilt exactly. A clean replay matched all 16 numerical result files.

connectome · C. elegans · bilateral neurons · state aggregation · routing model · model reduction · lumpability · reproducible benchmark