Cubrim is my lossless-compression research project and experimental archiver. I am trying to make a file smaller without losing a single byte when it is decompressed. In Arcanada, agents help me explore possible approaches, while exact reconstruction and the measurements tell me whether an idea worked.
One idea sounded convincing right up to the moment I ran the test. I have seen the same pattern many times: an idea looks good on paper, but the archive comes out no smaller, or the reconstructed file differs from the original. Either result closes that line of research for me.
Agents can bring me several candidates while I am still studying one. I pick a candidate, ask them to attack its weak points, and run the full compression and decompression cycle. Eloquence has no score in that test.
The Cubrim page currently puts ten general-purpose archivers side by side across three data sets: Silesia, enwik8, and Canterbury. On that table, Cubrim ranks first with an overall ratio of 0.1890. A lower ratio means better compression.
That number sits on top of many failed attempts. I kept this hypothesis because the file came back exactly and the size went down. I keep the failures too; they tell me which paths are not worth reopening.
After reaching 0.1890, I moved on to the next experiment. Cubrim still has plenty of open questions. That test gave me a version worth testing further.