Short notes from the lab, in plain language, on what the measurements keep turning up. Written by me, Joseph. The formal versions live in the methodology guide and the published papers; this page is where I explain things the way I would across a counter.
August 7, 2026
Bigger AI isn't saferMeta's 8B model outscores its own 70B on our fabrication index, 71.0 to 64.9. Why the models everyone deploys anyway are the ones you can't eyeball, and what that means for the file.