Train a cheaper model on reviewed examples.

The problem

A stronger model can produce useful examples. You want to know whether a less expensive model could learn the task well enough to run it economically.

Teacher output is treated as truth, or apparent savings are calculated before anyone checks the student on unfamiliar cases. The cost of preparing and reviewing training data is easily missed.

How Nibuno would handle it

The planned expansion connects teacher proposals, human qualification, approved training examples, the resulting student configuration, and an independent evaluation. Training orchestration remains a future capability.

What you’d compare

Compare quality and critical failures on unseen examples. Include data preparation, human review, training, inference, and fallback costs. The comparison needs to show whether the student meets the required quality at a lower total cost.