The verified deterministic context-mixing codec.
FP-CODEC is a classical context-mixing coder — frequency-count context models, a single logistic mixer and an arithmetic coder with lpaq-style APM/SSE refinement. On enwik8 it hits ≈2.03 bits/char, below zstd-19 (≈2.29) and lzma-9e (≈2.23), byte-exact lossless — and because it is integer-only it is bit-identical across x86-64, wasm32 and ARM64 by construction.
Bit-identical across platforms
The coder is integer-only — no float, no fused-multiply-add, no transcendentals — so the same source compiled to native, wasm32 and ARM64 produces byte-identical compressed streams. Verified: 0 divergent bits over 10⁶ symbols on text, real logs and random data; SHA-256 of the stream matches on every arch.
Beats zstd-19 AND lzma-9e on text
Frequency-count context models feed a single logistic mixer and an arithmetic coder, with lpaq-style APM/SSE lookup-table refinement. On enwik8 that lands ≈2.03 bits/char — below zstd-19's ≈2.29 and lzma-9e's ≈2.23 — byte-exact lossless. It also wins materially on Apache / Linux / SSH / HDFS / BGL system logs.
Verified & never-worse
Every archive carries a SHA-256 verified-lossless trailer, and a never-worse fallback: if context-mixing does not win, FP-CODEC delegates to xz / zstd / stored. So the output is never larger than the baseline and is provably the exact original bytes.
Tiny, portable decoder
No model files, no tables to ship — the decoder is the same small integer-only kernel that made the archive. An optional tiny domain-matched prior warms the small-file and log tier without changing the determinism guarantee.
One small command surface.
Compress, decompress and verify — the decoder is the same integer-only kernel that made the archive, so a file made anywhere decodes byte-exact everywhere.
# compress a cold text / log archive to a verified deterministic stream at1 fpcodec c access.log access.log.at1 # byte-exact, integrity-sealed, never-worse at1 fpcodec c corpus.txt corpus.txt.at1 --prior logs # optional tiny domain-matched prior at1 fpcodec d access.log.at1 access.log # exact original bytes, on any platform at1 fpcodec verify access.log.at1 # -> integrity: PASS (SHA-256 matches original)
Top-tier ratio, and the only one you can trust across machines
Dedicated research coders (paq / cmix) can push raw text ratio a little further — that's not the claim. FP-CODEC is the one that pairs a top-tier text ratio with cross-platform bit-identical determinism, a verified-lossless trailer, and a never-worse fallback. paq/cmix are neither deterministic across platforms nor verified.
| Codec | enwik8 bits/char | Cross-platform bit-identical | Verified lossless | Never- worse |
|---|---|---|---|---|
zstd-19 fast general codec — great speed, but gives up ~11% on text and makes no cross-platform-determinism or verification guarantee | ≈2.29 | in practice | ||
lzma-9e (xz) strong general codec — still ~9% behind on enwik8, no verified-lossless trailer, no by-construction determinism | ≈2.23 | in practice | ||
generic context-mixing (paq / cmix) top raw text ratios, but float-based mixing is NOT bit-identical across platforms, and ships no integrity trailer or fallback | ≈1.8 | |||
FP-CODEC top-tier text ratio AND integer-only bit-identical determinism, SHA-256 verified, never worse than xz/zstd | ≈2.03 |
Cold text and log archives — enwik-class corpora, Apache / Linux / SSH / HDFS / BGL system logs — that you store forever and decode rarely, and that must decode byte-identical on any machine. Maximum ratio plus a portable determinism guarantee is exactly the trade this tier wants.
It is a context-mixing coder, so it is slower than zstd — this is a cold-archive tier, not a hot-path general codec. If you need high-throughput compression, reach for zstd; if you need queryable typed data, reach for format-aware columnar or ATLAS. We say so plainly.
Built for
Long-lived text & log archives · compliance and audit stores that must decode byte-identical years later · any cold tier where you store forever, decode rarely, and need one compressed artifact that is provably the original on every platform.
Decoding and verifying an FP-CODEC archive is always free and needs no account; encoding is metered against a connected account.