Arshad's ViT
Ring-Native Vision Geometry

Images May Be Square. Bytes Need Not Be.

Arshad's ViT is an MPRC-native vision architecture under active development. Its starting point is not the conventional assumption that visual tokens must live in even×even square byte patches. The image may remain square in physical geometry while the byte carrier is chosen from an even×odd factorisation that is compatible with ring traversal. The current construction uses the 113×128 Hilbert carrier, whose volume is 14,464 states, and couples that carrier to a stride-7 walk on Z₂₅₆.

The 113 × 128 Hilbert Carrier

The useful observation is that a conventional 112×112 visual field sits inside a ring-native carrier whose dimensions are one odd and one even. The 14,464-state carrier is not introduced as padding for its own sake: it gives an exact even×odd address space whose residual against 112² is structured in 128-state bands.

The design rule used throughout this branch is therefore:

Patch Shape as a Ring Constraint

Conventional ViTs commonly use square patch lattices such as 8×8 or 16×16. Arshad's ViT separates the image shape from the byte topology: an image may be square, but the byte patch is not required to be even×even. The working rule is to prefer even×odd combinations so the carrier does not collapse into a symmetric square that erases traversal orientation.

This rule is being tested across practical image scales including 24 px, 112 px, 220 px, and 320 px. The objective is not arbitrary resizing; it is to determine a ring-compatible factorisation and traversal for each scale while preserving the same MPRC state rules.

Chunking the Hilbert Volume

The current carrier volume divides exactly into 64 chunks:

Thus the visual carrier can be treated as 64 ordered chunks of 226 states rather than a grid of independent square windows. The chunk count aligns with the quarter-turn quantum ι=64 while the contents remain part of one ordered carrier.

Quantum Walk, Not Sliding Window

The ring walk uses the MPRC observer stride rather than a conventional shifted square window:

Because 7 is coprime to 256, the walk is bijective over the ring. The intended vision primitive is therefore a cyclic traversal of token state rather than a Cartesian window translated by fixed pixel offsets. This is the architectural point at which Arshad's ViT diverges from Swin-style shifted windows.

BIND · REACT · MEASURE

The working MPRC vision block does not begin from dot-product attention plus softmax. It uses the same ring-native operator family developed in GEVHV:

REACT is LUT based. MEASURE operates on ring-native distance and state compatibility. The architecture therefore keeps the representation in the 0–255 domain instead of repeatedly projecting to floating-point logits and normalising them through an external probability geometry.

24 · 112 · 220 · 320

The next evaluation protocol is multi-scale. For each target image size, the experiment must report: the even×odd carrier, exact storage count, chunk count, traversal coverage, collision count, vacuum interaction, reconstruction fidelity, and task accuracy. A valid result must use the same ring rules across all scales rather than tuning a new geometry for each one.

Against Swin ViT

The intended comparison is structural before it is performance-based. Swin ViT groups tokens in local square windows and shifts those windows between stages. Arshad's ViT instead proposes one cyclic carrier, even×odd byte geometry, 64-way chunking, and stride-7 traversal, with BIND/REACT/MEASURE replacing the standard attention stack where possible.

No superiority claim is frozen here. The chapter records the architecture so that future image-generation, recognition, and reconstruction experiments can falsify or support it.

What Is Fixed, What Is Open