compress something. real numbers.

Paste text. Press compress. The same engine that lives in github.com/dot-protocol/pipernet runs your bytes server-side, gives you the real ratio, and verifies the round-trip is byte-exact.

What this is. Track B v0.3 — order-3 Markov + four match-window predictors mixed multiplicatively, single adaptive arithmetic-coded stream, byte-exact round-trip. Pure Python; no LZ, no Huffman, no codebook. On 250 KB of enwik8 it produces +43.71% lift over the order-3 Markov baseline (3.01% behind gzip). On your text it produces whatever the text actually compresses to. Honest both ways.

What this isn't. Not SOTA. cmix v21 holds the Hutter Prize at 14.6 MB on full enwik8; we are far behind in absolute terms. The architectural ground we're on (corpus-wide retrieval- augmented context mixing) is empty space in the 20-year sediment record of Hutter submissions — that's where the work is.

Source. Same code in your browser is compression/track-b/mixer_multi.py. Reproduce locally with python3 compression/track-b/bench.py 250000.