617 words in 1,381 bytes.

One blob, base85 wrapped around 838 bytes of bzip2. Unfold it and the page repeats the decode where you can watch: base85, then bzip2, then reverse, then the words. 1,048 characters of the blob itself. Nothing else was fetched.

Sorted forwards, these 617 words share 1,181 letters with their neighbors. Reversed, they share 1,301. That is the whole trick: -ear, -old, -ing become prefixes, and the Burrows–Wheeler transform finally has something repetitive to chew on. The blob shrinks by 30 bytes. The decoder pays five bytes for [::-1]. Net, 32.

A stored map — 22233344455566677778889999 — costs more bytes than the expression that computes it. Two boolean corrections absorb the groups that have four letters: PQRS on 7, and WXYZ on 9.

(ord(c) - 91 - (c > "r") - (c > "y")) // 3

Type a letter, or tap one. The group lit is where s lands: pqrs.

The repository’s first idea was not compression. It was Inverse Morphogenic Training: take a tiny graph, read its Laplacian spectrum, and use those eigenvalues as the singular values of a network’s initial weights. A triangle — genus 1, one independent cycle — learned this same 617-word T9 task far faster than a random start. The numbers below are the archive’s, from a pure-NumPy race in March 2026. They are reported, not re-run here.

Three nodes, one cycle, and the edge between C and F drawn thicker than the others. In the original race this crystal reached 79.6% accuracy on the 617-word task. Xavier initialization, same kind of network, managed 1.3%.

Hold the architecture fixed and the story shifts. In that ablation a genus-3 tetrahedron won, at 71.3%, and the triangle fell to 66.5%. The archive’s own revision: the spectrum is the signal. Genus predicts, but deriving depth and width from the graph was a second-order effect, and it is what made genus-3 look bad in the first race.

Genus 1 won all five. On XOR the margin gets thin—a random init is no longer lost when the task is easy. On T9 it is not close.

The neural thread left this repository and continued as Prism, a spectral-transfer method with a reported 3.33× faster GPT-2 convergence. What stayed behind is the language itself, folded until the file was five bytes from the floor.

Starting from 2,101 bytes — a trie-shaped string, zlib, base64, a docstring — an autoresearch loop tried a change, measured the file, and kept it only if every word still decoded. Twenty-eight techniques. Fifteen were thrown out.

2,101 → 1,381

Thirty-four percent gone. The behavior is the same.

Delta. Sorted words stored as a prefix length plus only the new suffix.

The blob, 1,048 characters of base85, was the smallest stdlib encoding they could find: bzip2 of the reversed delta, at every compression level, against zlib and lzma. The surrounding Python is 333 bytes. An automated scan that deleted one character at a time found five bytes that are only there because the language requires them — a space inside for, the str() around the digit. The floor is about 1,376. The file is 1,381. The gap is 0.36%.

A perfect order-1 model of the words would need about 652 bytes of payload and then hundreds more to store the probabilities and the decoder. bzip2 lands at 838 bytes of payload and borrows its model from the standard library, which is free. Fourier coefficients, a logistic map, a tiny sine-network: each one had to specify at least as many parameters as the bytes it hoped to replace. The initial condition is the data.

The formula is the compressor.

Twenty-three of them have two letters. None have five. 95 keypad codes point at more than one word—that collision is the whole game of T9, and this dictionary does not rank the answers.