Mechanistic-interpretability experiments, research records, and ML visualization tools, including SELF-GROUND and construct-mismatch studies.
