Cortical Columns, Capsules, and Transformers: A Typed Benchmark for Part-Whole Representation, Reference Frames, Routing, and Receiver-Relative Dynamics

Abstract

# Cortical Columns, Capsules, and Transformers Recognizing a moving or partly hidden object requires a system to preserve relations among parts, wholes, pose, context, time, and action while its immediate inputs change. This paper makes a direct operational comparison among cortical-column circuits, Hawkins-style reference-frame models, capsule networks, transformer attention, and a Self Aware Networks receiver-state model. The study introduces a typed architecture contract and a shared eight-task benchmark. Seven small trainable implementations receive the same declared inputs, task interface, data partitions, optimizer, training schedule, and frozen seeds. The models expose their internal states, undergo targeted interventions, report their resource use, preserve complete adverse and null evidence, and replay their scientific outputs deterministically. No target operation passed the complete frozen success rule. Two comparisons were favorable but incomplete, while several others were clearly adverse or moved in the wrong direction under intervention. A bounded Lean model also checks finite interface, provenance, refusal, and claim-class rules without admitted proofs. These results apply to the declared small implementations and finite model only. They are not a ranking of complete theories, a biological validation result, or a final-test result. This is a preprint and has not undergone peer review.