Field notes
Notes from the index
One finding per note. One argument. The numbers behind each one are in the model cards.
Relevant context made a 31B model 13 points dumber
A paired closed-book versus with-context audit, why the random-context control rules out noise, and what it does to your retrieval budget.
The August 2 problem
EU AI Act obligations bite on August 2 2026. What evidence deployers will actually be asked to produce.
Instruction-following is not a security boundary
The injection defense ladder. Prompt inoculation barely moved the hijack rate, and on some models made it worse. Ingestion sanitization is what worked.