Spatium.ai

A gradient-free, integer-only neuron that stores patterns, self-organizes, and flags novelty. Here is what we measured on real data, what failed, and what we haven't run yet.

Where Spatium stands. We match patterns using only bit-vectors and Hamming distance — no gradients, no neural net weights. It gets to 84.56% on MNIST. Here are the results, the experiments that failed, and the parts of the idea we still haven't tried.

Introducing xn-spatium — a bit-spatial AI substrate built as a downstream consumer of xtranodly. Day-1 substrate gate, P0 primitives dogfood (similarity neuron), transfer-mode probe, and the post-P2 refactor onto the public Rust API.

Strategy • Observer • Feature Flags

Parallel Iterator • Work Stealing • Builder

Adaptive Learning • Observer • Builder

Strategy • Spatial Hashing • Enum Dispatch

Bloom Filter • Quantization • Region Partitioning

Z-Order Curve • Builder • Newtype