Step by step, inside the software, by the people who did it. Each trajectory is a task, the actions taken, the intermediate states and the outcome, with the reasoning annotated by the person who did it.
expert trajectory · rcm · one task
reasoning annotated by the staff member who did it · 6 actions · grader: payer adjudication
Where the software keeps an audit trail, the trajectory is recovered from the record itself. Where it does not, the original staff work the same tasks live in the twin and we capture every action.
Practice management, claims and case systems log who did what and when. Those logs, de-identified, are trajectories already.
For systems without a trail, the people who ran them replay the day's work in the twin. Annotated as they go.
Supervised fine-tuning directly, as prompt, actions and outcome. Or as the reference trajectory a grader compares an agent's path against when process matters as much as the result.
Each step is a tool call or a UI action with its arguments and the state it produced, not a paraphrase.
Every trajectory ends with what the outside world said: paid, accepted, docketed, closed.
Also in for labs
One business, every tool, outcomes attached, entity-consistent across sources.
Digital twins of the industry software, seeded with real history, with verifiers.
Held-out slices scored with the same verifiers, refreshed as records arrive.
Licence chain, method document, exclusions annex, watermark, jurisdiction.
