FIND THE SIGNAL.
Capture activations and analyze refusal subspaces with the native Swift/MLX research engine.
ABSlayer applies controlled changes to language models to reduce refusal behavior. Evaluate the results, check what capabilities remain, and keep the evidence behind each change.
Find refusal-related patterns. Apply an intervention. Evaluate what changed.
Capture activations and analyze refusal subspaces with the native Swift/MLX research engine.
Explore residual and weight interventions, adapter training, retention measurements, and export workers.
Use Pool CLI or Codex CLI through a shared project skill. Workspace checks get durable jobs, status, cancellation, and recorded evidence.
A completed answer still needs a correctness check. Track each outcome separately to understand what changed.