Project context and scope

A reference codec for storing predictor hits, miss residuals, manifests, and integrity metadata around a shared next-unit model.

AIXC explores a simple compression question: if encoder and decoder share the same deterministic predictor, how much of the original text needs to be stored? The current reference implementation stores a seed, packs hit/miss decisions from a predictor, and records residual literals only when prediction fails. The repository includes a binary container, deterministic byte backends, optional Hugging Face model-token experiments, rank residual modes, entropy-coded sections, CRC/SHA integrity checks, a CLI, tests, benchmarks, and a local workbench. This paper summarizes the codec contract, the economics of the reference format, the binary container, and the engineering constraints that matter for reproducible decoding.

Source repository →