Evaluation
Grade the tool path, not only the answer
Correct finals can hide bad process. For operational work, the path is part of the product.
You can land on the right answer with the wrong systems, skipped checks, or a path nobody on the team would take. In ops work, that still counts as failure.
So we score outcome, tool sequence, and how often a human has to intervene. Held-out jobs from your queue matter more than a public quiz that never saw your procedures.
Training and eval should point at the same object: the workflow. If the path is wrong, the model isn’t done.