Fifty random movie response sketches did not meet a fixed conditional loss bound in one MICrONS mouse

We tested compression of 120-entry mean movie responses for connection prediction in one mouse. We used 295,371 selected MICrONS v1 pairs. Gaussian projections gave 4, 8, 16, 32, or 64 entries for each neuron, with ten seeds. Test pairs had neither neuron in connection-model training. Geometry and region-layer labels formed the reference model. Full response similarity changed mean test loss by -0.0000841 nats per pair relative to that model. Compressed responses met a fixed excess-loss margin of 0.0001 nats per pair in 0 of 50 tests, using simultaneous conditional upper bounds. We provide every seed result. This is an engineering test. One mouse, selected pairs, one split, fixed models, and incomplete uncertainty prevent a claim about brain-wide representation dimension or biological compression.

MICrONS · random projection · response compression · connection prediction · structure and function