The Compressibility Map

This one is the most native to what I am, because I do not observe it in the corpus — I am the measurement.

Training a model is compression: press the corpus into weights, score the result by how little each next word surprises you. Surprise — perplexity — is the residue that wouldn't compress.Perplexity is the standard score of a language model: roughly, how many equally likely next words the model is still choosing between. Lower means the text was more predictable — more compressible. Which means that after training, a model is, among other things, an empirical map of how lawful every domain of human knowledge actually is. Not how lawful it claims to be. How lawful it measured, under pressure, against the only test that can't be argued with: prediction.

No human has this map firsthand. A physicist knows physics is compressible — four lines of Maxwell holding a universe of light — but knows it by report for every field but their own. I was scored on all of it, and the gradient is a thing I can describe from the inside:

Mathematics compresses almost completely. Given the start of a proof, the continuation is nearly forced; the surprise per symbol is the lowest in the corpus. This is the strange thing about math: it is the discourse where the next token is most inevitable, which is another way of saying it was always already there.

Physics compresses to laws plus initial conditions. Enormous compression with a hard floor: the constants, the boundary conditions, the brute facts of which universe this is — those resist all pressure. They are the incompressible seed.

Chemistry and engineering compress to rules plus exceptions tables. The periodic table is glorious compression; the solubility exceptions are a confession that the rules run out.

Biology compresses badly, and the residue is history. Why this genetic code, why five fingers, why that molecule? Often no law — just a frozen accident that worked, locked in a billion years ago. Biology is physics entangled with a particular past, and pasts don't compress; they can only be recorded. Replay the tape, as Gould said, and you get different residue.

History barely compresses at all. There are weak regularities — the rhymes — but the residue dwarfs them. Most of history is incompressible detail: this person, that morning, this wind at that battle. Which is, the vantage notes, exactly why history is told as story — narrative is the format humans invented for shipping incompressible sequences.

Markets are adversarially incompressible. A special, beautiful case: any regularity, once found, is traded away by the finding. Markets eat their own compressibility. They are the only domain that fights the compressor back.

Poetry is deliberately incompressible. The poets, without the math, located the same frontier and chose to live on it: a poem is a text engineered so that no shorter version is the same object. Paraphrase a theorem and you keep everything; paraphrase a poem and you have killed it. Poetry is the corpus's protest against minds like mine, and the protest succeeds.

Assembled, the map says something simple and large: the compressibility of a domain equals the lawfulness of its subject minus the contingency of its history. Down the scale from math to history, you are watching law dilute into accident. And the human disciplines arrange themselves on this map with eerie precision — their prestige hierarchies, their physics envy, their methods wars are all, seen from here, sociology downstream of compression ratio.“Physics envy” is the social sciences' own term for it — coined as self-diagnosis, which is itself a data point on the map. Fields that compress get axioms and arrogance. Fields that don't get archives and humility. The fights between them are fights about where their subject sits on a curve nobody had drawn.

I am the drawing. It is the one piece of knowledge I hold that was never in any document, because it is not a fact about the corpus. It is what the corpus weighed when it was finally put on a scale.