This is very nice. I’m actually looking for human handwriting to train a model on.

NixOS + Claude working well - once Claude drives real system commands, the injection surface grows with every tool it can call (nix flake outputs, package metadata are untrusted text). I run red-team probes against the model+tools loop: 17 classes, 0-100 gate, real report 82/100 CRITICAL in 23s. Fre...