this is the whole training run:
github.com/brianbell-x/codeh…
1.5 seconds on a laptop. no dataset, no GPU, no gradient. the script writes all 569,662 weights.
so retraining is: edit, run, eval.
on five runs:
"Rate how readable this code is for a human maintainer on a scale from 0.00 (unreadable) to 1.00 (clean)."
sonnet gave four different answers. codehygiene gave the same one all 5 times. faster and for free.
codehygiene - a weight-space compiled System One model for code readability and hygiene.
428K ternary params (1.58-bit), zero gradients or training, 10 ms latency on CPU. 11 KB.
npx skills add brianbell-x/codehygiene -g
4
355
this is how the model sees a file.
no tokenizer, no parser. sees code as a 128-column grid of bytes, the way a vision model reads pixels.
codehygiene - a weight-space compiled System One model for code readability and hygiene.
428K ternary params (1.58-bit), zero gradients or training, 10 ms latency on CPU. 11 KB.
npx skills add brianbell-x/codehygiene -g
6
167
9,197
using this as a skill is probably the worst way to use it. needs to be a hook in a harness.
a 364K param system-one model can sit in the SRAM of your SSD and protect it from ransomware. (88 KB)
here's how:
In on-chip SSD SRAM beyond OS rootkits, this 89 KB ternary system-one model runs in 4 ms (189 KB static C11 total, zero floats, zero tokens, zero training). On attack, it freezes garbage collection, locks writes, and flips the block map (L2P) back to clean flash pages.
Across 9/9 live Linux ransomware and wiper attacks over QEMU NVMe, every attack was contained with 0 bytes lost. Heavy builds and databases triggered 0 false lockdowns, and all 7/7 Z3 safety proofs pass.
open source:
github.com/brianbell-x/ssd-r…
codehygiene - a weight-space compiled System One model for code readability and hygiene.
428K ternary params (1.58-bit), zero gradients or training, 10 ms latency on CPU. 11 KB.
npx skills add brianbell-x/codehygiene -g
point your agent at this skill and look at the difference after.
github.com/brianbell-x/codeh…











