worldmodel

README

WorldModel

Symmetric compressors in Go. The encoder and the decoder share one model, and decompressing a blob restores the original file bytes.

The module is github.com/pointlander/worldmodel and requires Go 1.25. The cell command also requires ffmpeg and ffprobe.

go test ./...

cell

cell turns a video of dividing cells into one CSV row per frame, then compresses that table. Each row is the frame time followed by the red, green, and blue sample of every pixel. The time is the frame index divided by the rate reported by the video.

go run ./cell
go run ./cell -size 64 -video cell/01_video.webm
go run ./cell predict cell/cell.bin cell/future.webm

usage: cell [-size 32|64|full|WxH] [-video FILE] [file]
       cell predict [file] [webm]

The default picture is 32×32. -size full keeps the video's own frame size. -video defaults to DNA cell death and division-L.mp4 in the working directory or in cell/. A 32×32 conversion is named cell.bin; any other size is named cell{w}x{h}.bin, next to the video unless an output path is given. The tracked cell/cell.csv is the canonical table the tests compress.

predict reads that blob, appends 8 seconds of forecast frames at the clip's frame rate, and writes a WebM. The frames go to ffmpeg as rgb24 and are stored with lossless VP9. A second-order Markov chain draws the next law and motion, falling back to the previous symbol alone when that pair is empty. Cells on a dark ground advance as elastic meshes. A picture that is mostly bright is warped by the block motion and sampled from the last frame. Grid nodes stay within an eighth of a block of their neighbors, so the warp does not tear the picture into pixels. A second-order Markov model of each pixel's color, falling back to the previous color alone, replaces that sample when the sample is a less common continuation.

The compressor predicts each frame after the first. A block warp cuts every block into two triangles and may split a triangle at the midpoints of its edges. A mesh law moves the grid by one velocity per block. In the second half of an even-length clip, a frame may repeat its paired earlier frame. The canonical 32×32 clip compresses to 22985 bytes.

iris

The Fisher iris file. The three classes are fifty-row blocks, so the class is the row index. Each measurement is an integer tenth of a centimeter inside a per-class span, and a row is one mixed-radix digit in the product of those spans. The packed form is 298 bytes.

Search also builds one Turing-machine program for the file. Lines are searched, then paired, until a single program remains.

bcwd

The Wisconsin diagnostic breast cancer file. Each of the thirty features is a decimal with at most four significant digits, coded as an integer in that column's mantissa range. Identifiers carry a two-bit width tag, and the diagnosis is one bit.

secom

The SECOM semiconductor files. The matrix is 1567 rows by 590 features, with NaN for a missing cell. A value is an integer at that column's fixed decimal scale, Rice-coded from the column minimum, and identical missing-cell patterns are stored once. The labels file is a pass/fail bit and a Rice-coded run of timestamps.

tm

Arbitrary bytes as one Turing-machine program. The search applies a Burrows-Wheeler transform and move-to-front, then picks the shortest program under a fixed instruction table:

Instruction	Form
emit	01 bb
lit	02 nn bb...
rep	03 bb nn
copy	04 dist u16 nn
far	05 dist u32 nn

copy and far append bytes from earlier on the tape. rep repeats one byte. lit and emit write raw bytes. The field tables stay in the decoder.

Number-theoretic transform

The root command prints the number-theoretic transform of a file and checks the inverse.

go run . [file]

The field is Z/998244353Z. Each input byte is one residue, and the input is zero-padded to the next power of two up to 2^23. The default file is iris/iris.data.

License

BSD. See LICENSE.

Documentation

Overview

Command worldmodel applies the Number Theoretic Transform to a byte string.

The field is Z/998244353Z. 998244353-1 = 119*2^23 and 3 is a primitive root, so every power-of-two length up to 2^23 has a principal root of unity. Each input byte is that residue. The input is zero-padded to the next such length; the inverse of that spectrum is the bytes followed by the pad.