We probe what open-weight models do when they leave the rails: manipulation transfer, euphemization, dishonesty under social pressure. We ship the benchmarks that catch each one.
Manipulation transfer, persuasion realism, honesty under pressure. Evals for behaviors that don't show up on a leaderboard.
Read researchWhat open-weight pretrains are taught to flinch from, measured at the token level across labs and corpus generations.
Read researchAbliteration, Heretic, Obliteratus: what these tools actually change about a model's behavior, beyond removing the refusal text.
Read research
April 2026 · Pretrain Forensics
March 2026 · Benchmark Design
March 2026 · Uncensoring Methods
Bespoke inference needs, related research, or feedback? We'd like to compare notes.
Contact us