How should a lab choose a useful first correction pilot?

Begin with the decision you want the model to learn or demonstrate. A useful pilot might ask whether an agent can determine when a cost reclassification is ready for posting. That objective is easier to inspect than a broad request for realistic finance data.

Define the permitted correction method, participating roles and evidence needed to reach the decision. Then state the accounting assumptions: whether the period is open, which currencies are supported and whether tax treatment is outside scope. These boundaries prevent unspoken requirements from dominating the result.

Ask for a sample that includes a normal path and cases that stop at the chosen decision boundary. The purpose is to inspect whether the proposed material presents the intended problem and whether the expected outcomes are adequately justified.

MirrorFirm could scope generated datasets, tasks or an environment around that brief. Expansion can follow the findings: add another correction method if method selection is weak, or add more complex evidence if interpretation is weak. A broader domain mix is useful only when it answers an identified training or evaluation question.