✶

Binary Spatter Lab

interactive VSA playground for XOR, bundling, noise, and retrieval

public endpoint on moth.page
vector symbolic architecture

See binary spatter codes move.

Binary spatter codes represent symbols as long binary vectors. Binding is usually XOR. Bundling is usually majority vote. Similarity is measured by Hamming overlap. This lab lets you watch those operations distort, preserve, and recover structure.

You can tune dimensionality, sparsity, library size, bundle size, and noise. The page then regenerates a tiny universe of codewords, renders one of them as a radial spatter field, and estimates how well noisy retrieval still works.

512dimensions
50%1-bit density
0%noisy retrieval accuracy

What you are looking at

Each dot in the radial view is a bit position. Bright dots are 1s. Dim dots are 0s. The geometry is fixed by index, so changes in the vector show up as changing constellations rather than rearranged pixels.

Current bundle majority
0%
Selected symbol distance
0

Why BSC is strange and useful

In high dimensions, random vectors are almost orthogonal. XOR binding keeps information distributed. Majority bundling preserves coarse consensus. That means structure can survive operations that would destroy it in low dimensions.

XOR binding majority bundle Hamming similarity noise stress test

Selected symbol as spatter field

Fixed radial layout. Index drives angle and radius. Values drive glow.

symbol: S0

Density

0 of 0 bits active

Nearest neighbor

—

Bundle relation

—

Derived operators

Bound vectors scatter aggressively. Bundles retain consensus shape.

ready

XOR bind: A ⊕ B

—

Majority bundle

—

Unbind check

—

Interpretation

Generate a codebook first. Then pick two symbols to bind and a small set to bundle.

Codebook and similarities

Random high-dimensional codewords should sit near half-distance from one another.

mean pairwise distance: —
Symbol Active bits Density Nearest neighbor Distance Preview

Noisy retrieval sweep

Each point estimates how often a corrupted vector still retrieves its original symbol via nearest-neighbor Hamming distance.

waiting
Current-noise accuracy
0%
50% collapse point
—
Most robust symbol
—

Reading the curve

In a healthy high-dimensional regime, light noise barely matters. Then there is a shoulder, then a cliff. Larger dimensionality pushes that cliff outward.

Working notes

Small textual readout of what the current universe is doing.

Binary spatter codes are one branch of vector symbolic architectures. This page does not try to cover the full literature. It gives you something more tactile: an instrument for seeing how XOR, bundling, and Hamming geometry behave when the space is large enough for the weirdness to become stable.