With two direction groups assumed and one label per graph, contact area classified 54 of 56 other hair cells in seven zebrafish neuromasts

We tested how much direction information can be read from zebrafish neuromast contact graphs with one known label per graph. A fixed spectral method assumed two direction groups. It used contact area to split all cells before reading any direction value. One cell, selected by its absolute structural score, then set the direction of the split. That cell was excluded from accuracy. In seven normal neuromasts, the method classified 54 of 56 other recorded bundle directions correctly. Equal-animal accuracy was 95.92%. Six graphs had no errors. Both errors came from WT5. Binary contacts gave 76.20% accuracy. The paired sign-flip reference gave p = 0.0625, so the small panel gives limited evidence for an area advantage across animals. Age-conditioned label shuffles gave mean accuracy 51.82%; none of 9,999 draws reached the measured score. Ribbon weights gave 53 correct tests and three unknown classes. Contact area classified all three zero-ribbon test cells correctly. Testing every possible single known label reduced equal-animal mean contact accuracy to 93.88%. Terminal omissions showed further failures. These results concern bundle direction as an indicator of mechanical sensitivity. They do not measure nerve firing or behavior. The source already established direction-specific wiring, and the spectral method is standard. The contribution is a low-label benchmark, its limits, and a reproducible package with source data, provenance, code, forecasts, controls and figures.

connectome · zebrafish · neuromast · contact area · bundle direction · spectral partition · single label · structure and function