1,381 bytes contains the entire T9 predictive text language — 617 English words, full keyboard mapping, and a working CLI. No dependencies. No training. No data files. Just one Python script smaller than this paragraph.

python main.py 843              # the tie
python main.py 4663             # gone good home
python main.py                  # prints all 617 words

The 1,048-character blob is 838 bytes of bz2-compressed data encoded in base85. Inside: 617 English words, reversed and delta-encoded. bz2's Burrows-Wheeler Transform compresses reversed English words 30 bytes smaller than forward-sorted. See archive/docs/WHITEPAPER.md for the full compression research (28 techniques tested).

The neural network research in this repo grew into Prism: Prismic Pretraining Acceleration — a spectral transfer method that achieves 3.33x faster GPT-2 convergence. The Prism code and findings now live in their own repo.

Apache 2.0