AI research
The AI program builds the parts around the model that make it useful and inspectable: calibrated decision heads, persistent agent runtimes, and controlled experiments on what actually changes a model's output.
Decision models
Open-weight System One models that return calibrated choices, scores and yes/no decisions in one prefill pass.
LibertAI Labs
The paper behind Deem, our open-weight System One decision model family. We document the letter-slot architecture, two failure modes of decision SFT that accuracy gates cannot see, a teacher study whose pilot result did not transfer at scale, and how we arrived at the released deem-9b-v1 and deem-0.8-v1 checkpoints.
Model behaviour
A controlled abliteration study that falsifies the textbook projection method on a hybrid architecture and isolates when steering breaks generation quality.
LibertAI Labs
A formal writeup of our agent-run abliteration study on GLM-5.3-Flash. We falsify projection-based abliteration on a 4-stream mHC hybrid, isolate decode-time constant injection as the cause of long-generation quality collapse, and combine community abliterated weights with a light prefill-only activation top-up — reaching 0% true refusal at 10/10 capability.
