Method
Taught, then checked against the job.
Training has an operational shape. Specialists teach the model, a second reading settles disagreement, and a held-out set judges the result. This is not a claim that every public corpus has been remade, or that a training run is a guarantee.
Sequence
Name the job
The task, the audience, and the standard are written before the model is taught. Every example has to mean something for that job.
Staff the field
Reviewers are drawn from the specialist network for the domain: physicians, academics, or engineers, depending on the work.
Read it twice when it matters
High-stakes items get an independent second read. Disagreement goes back to the rubric. It does not get averaged into a false consensus.
Keep a set aside
Evaluation examples stay out of the training pile. A model is still checked against the job after it has been trained.
Hand over the set
Delivery is the labeled set, the rubric, and the notes on what was rejected. A downloadable public corpus is not what this page offers.
Stop where the claim stops
We do not treat a labeled set as a guarantee of model behavior, a clinical device, or a filing with a regulator.