Determinism gives you a stable function; it does not give you compression. The question is whether a small cause can be recovered from real audio at all, and the bookkeeping has to be brutal to mean anything: the genome, the manifest, and both SHA-256 hashes all count toward container size, and a model only wins by reproducing the complete file.
Integer math end to end, so decode is byte-identical on every machine. Eight decoder models compete on every exact encode, from oscillator genomes to reversible mid/side to integer wavelets. The lossy ladder is masked wavelets with integer rate-distortion optimization, screened by ViSQOL and a built-in ABX listener against size-matched Opus controls. Forty experiments in the log, and the rejections are written up as carefully as the wins.
A working lab: drop audio in, get the winning model, exact byte accounting, matching hashes, and an ABX session on your own ears. On the four-track SQAM diagnostic the exact codec beat maximum-effort FLAC on bass drum, piano, and soprano, and lost on speech. It is research, not a shipped format, and the experiment log says exactly which is which.