Active learning

Disagreements

Every shadow-inference call where the FT and Claude disagree meaningfully lands here. Reviewers pick a winner (or override both with their own judgment); promoted rows become high-priority training data on the next retrain.