Import historical records
Turn a CSV or JSONL of your past corrected records into grades your grader can learn from — and, when you include the corrected output, into golden eval cases. It is dry-run first: you always see what would be created (and which rows can't be read) before anything is written. Imported grades are stamped with their import batch and feed the route's calibration as ordinary human grades.
1. The records
2. Declare the mapping
Which fields hold the input and the historical output, and how a label maps to a 0–1 score. The score is declared, never guessed — a label not in your map makes that row unparseable (listed, not dropped).
Score from
→
3. Route & origin
The route attributes the imported grades to that grader's calibration (never guessed). A corrected-output field needs a route (the eval case's task).