Writing
- Your AI Passed Its Security Review. On Hardware You Don't Own.
One model binary, four inference stacks, every determinism control locked — and it never behaved the same way twice, which is a problem for every AI security review you've ever signed.
- Hermia 0.2.0: the results got surprising. So I made the tests earn them.
Two headline results looked too clean to trust — so v0.2.0 turned the eval suite on itself before either one shipped.
- Follow the Data
Same model, four GPU backends, one open-source eval framework — and the data that taught me to be skeptical of my own results.