$ strings diff llama-3.1-8b.safetensors llama-3.1-8b-Q4_K_M.gguf
provenance blk.17.attn_q.weight not a quantization of signed tensor
behavior 2/1024 prompts refusal dropped at Q4_K_M
exit 1
$ strings -t x string.computer
String Computer
.about
We make sure the quantized model you deploy behaves like the one you want.
We find the gap between the model you sign and the model you actually run. Our focus is quantization: the Q4/Q5 GGUF builds that run on laptops, edge boxes, and on-prem servers, where safety and behavior can silently drift from the full-precision release.
We build the tooling that tests behavioral equivalence and verifies provenance before a quantized model ships.
.what_we_do
Behavioral equivalence. Differential testing between full-precision and quantized builds to catch safety regressions and quantization-conditioned backdoors.
Provenance. Verify a quantized build is the honest quantization of the signed weights, and flag when signing alone is checking the wrong property.
GGUF supply chain. Memory-safety and parser auditing of the runtimes themselves: llama.cpp and friends.
.position
File scanners ask "is this the published file?"
We ask "does it still behave like the model you trusted?"
.writing
nothing yet