A 663-cell mouse cerebellar map leaves two axon signal directions unresolved in an equal-claw linear model

We tested whether input coverage ensures source recovery in a linear model of a public mouse cerebellar map. The selected core had 663 granule cells and 381 connected parent axons. Equal-claw weights gave rank 379, although every axon had a target. Exact arithmetic certified the rank and two invisible input changes involving seven axons. After random removal of half the cells, 30 samples retained a mean of 342.17 covered sources but only 324.10 recoverable dimensions. Graph matching was a close rank bound, but it exceeded equal-claw rank in 13 partial-removal samples. A selected four-input panel had rank 249 for 306 sources; the full release had rank 729 for 729 sources. Code, compact source tables, all removal scores, exact witnesses and checks are supplied. This is a noiseless linear identifiability test in one animal. It does not measure neural decoding, lesion tolerance, noise amplification, or cross-animal generalization.

cerebellum · connectomics · linear decoding · structural rank · cell loss · input identifiability