Decision dataset. Trains models on one judgment: “is a candidate answer supported by the given state?” Each record holds a state, a query, and two candidate decisions, one supported and one not; the model answers with a choice or a probability, never text. Truth is relative to the state, not the world.
Verified agentic coding traces. Each trace records a coding agent solving a real problem. Traces contain reasoning, tool calls, and results, from a controlled runtime. One demonstration per task.
Verified agentic coding traces. Each trace records a coding agent solving a real problem. Traces contain reasoning, tool calls, and results, from a controlled runtime. A chosen and a rejected trace of the same task.