Notes from an independent AI safety practice.

Misalignment write-ups, eval-design pieces, and occasional method notes. Most carry a Hugging Face artefact or a public repo; everything is reproducible.