Structuring expert reasoning for agent and RL training

Expert working knowledge is a different kind of operator data. There is no camera. The task is reasoning, and the ground truth is the sequence of judgments an experienced clinician makes between a patient presenting and a treatment being chosen.

For our oncology program, oncologists document diagnosis pathways step by step: presenting features, differentials considered and ruled out, staging, the treatment options weighed, and the rationale for the selection. Alongside this they record efficacy trends they have observed across de-identified cohorts.

The structure matters as much as the content. Each pathway is written as a trajectory with explicit decision points, so it can serve as a reference for agent evaluation or as the basis for a reward model in reinforcement learning. Free-form essays are easier to collect and nearly useless for this purpose.

We partner with institutions here the same way we partner with factories. We contract individual clinicians, and we also work with the organizations around them: hospital groups and cancer centres, diagnostic labs, and the quality, process and engineering teams inside industrial firms who hold the reasoning behind their own plants. An institution brings a cohort of experts, a review process, and a legal path for the data, which is usually what decides whether a program of this kind can run at all.

That reasoning exists well outside medicine. A process engineer at a chip fab, a line supervisor at a bottling plant, and a planner at a logistics hub each carry a decision tree that has never been written down, and each one is reachable through the site rather than the person.

All patient-level information is de-identified before it reaches us, and contributing clinicians work under agreements that cover consent and data handling. Specifics are available to partners on request at [email protected].