Compression research has produced ratios nobody can deploy, because reading the data back costs more than the bytes are worth. Our work is organised around a different question: what is this asset worth, and what is the best result reachable at that budget.
The method is unpublished. We would rather show you a number you can reproduce than a diagram you have to trust, so this page describes the shape of the work and the benchmarks page carries the evidence. If you are evaluating us seriously, the useful conversation is about your data and your baseline, not our internals.
These are not three products. They are three points on a curve trading encode compute against bytes saved, and the interesting engineering is knowing which point a given asset deserves.
A single model trained across a wide range of data distributions, so one engine handles mixed real-world data instead of a different specialised tool per file type. The cost of training is amortised across every domain it covers and every customer who uses it, and each training run raises the floor for all of them at once.
Broad coverage at production cost. This is the mode most data should be run through.
Compression has a floor, and that floor is specific to the individual file. When an asset is valuable enough to justify the compute, we specialise against that asset and get closer to its particular floor than any general-purpose tool will reach, because a general tool has to be right about everything at once and this does not.
The best result available for a given file, at a compute budget you choose. This is the mode behind our benchmark result.
For the most valuable data we run a large parallel search for a shorter exact representation of the bytes, evaluating many candidates at once. Every candidate is verified to reconstruct the input bit for bit. Anything that does not verify is discarded rather than shipped, and nothing unverified is ever counted in a published number.
Results that conventional coders cannot reach, on data where the bytes are worth the search.
Research compressors have removed far more than any deployable tool for years. They stayed in the lab because reading the data back was astronomically expensive, and a ratio you cannot afford to use is a paper rather than a product. Keeping decode cheap while the ratio moves is the work that turns a benchmark result into infrastructure, and it is the reason we start with encode-once, decode-many data.
Every number we publish counts the decompressor and all of our side information, on a named public corpus, against a tuned baseline rather than a default one. A result exists for us only after the decoded output matches the input byte for byte. The public board is the authority for the listed row; later models stay off this site until they are listed.
