Research notes on AI systems, the law, and the places they meet. Each one is written to be read in plain English, with result bundles available to download from the paper page.
Hidden credibility judgments in legal AI: internal inspection on an open 7B model, plus the same behavioural battery run on Claude Sonnet 5 and GPT-5.5. Hardened placebos, behavioural cross-check, closed-model replication.