A Differential Neural Scene Codec

A bidirectional perceptual interface must solve a harder problem than decoding or stimulation alone. It must estimate an individual's current neural state, construct a target state associated with a scene or percept, determine which part of the difference is observable and reachable, and apply a bounded update without assuming that a population-level recording is a complete description of experience. This paper formalizes a proposal that Micah Blumberg stated in the April 11, 2017 Neural Lace Podcast: rather than superimposing an entire target pattern on an incompatible ongoing pattern, estimate the difference between the pattern present and the pattern appropriate to the intended scene and transmit that difference. The same source introduced Alternating Transmission Protocol (ATP), deliberately linking feedback/handshake and connectionless transmission modes to the biological acronym for adenosine triphosphate. A 2022 Self Aware Networks note extended the proposal with recipient-specific phase differences, multimodal object states, and receive-transform-project neural rendering. The present contribution is a modern, falsifiable control formulation of that historical scene-conditioned residual idea. It does not claim priority for feedback control, neural decoding, closed-loop stimulation, state estimation, or model-predictive control. The paper states observability, identifiability, reachability, uncertainty, and safety conditions; defines matched comparator controllers; preserves a deterministic 3,000-run exploratory benchmark; and reports a separately frozen 2,520-run adversarial benchmark. The adversarial test adds nonlinear dynamics, observation rank loss, an unreachable target component, correlated measurement noise, stimulation artifact, and a combined stress condition. Predictive residual control won only two of six conditions, ranked fifth under observation rank loss, and lost to the no-control decoder baseline under combined stress. At least one state-bound crossing occurred in 16.1 percent of its adversarial runs, concentrated in rank-loss and combined conditions. These results narrow the proposal: a differential neural scene codec is meaningful only on a measurable, artifact-controlled, person-specific, reachable state projection, and it fails when a simpler controller or no control performs as well under a frozen test.