Review and Improve Predictions
A Nyckel prediction endpoint gets more accurate as you provide feedback. Every correction you make becomes a training example that improves the model.
Every correction flows back through the same Input → Prediction → Review → Improved-endpoint cycle described in The feedback loop. This page focuses on the human side of that cycle: where to do the reviewing, and how to spend that review time well.
Reviewing predictions in the console
- Open your function in the Nyckel console.
- Go to the Review tab. It holds the predictions invoke capture selected from recent traffic — the ones most worth your review time.
- For each prediction, confirm it or correct the label if it was wrong.
- Nyckel retrains automatically after you save corrections.
Feeding back outcomes from your application
The Review tab is the recommended workflow: Nyckel manages the active learning for you, deciding which predictions deserve your attention. For most teams, that queue plus a few minutes of review per day is the whole improvement process.
If your application learns true labels on its own — a user hits “this is wrong”, an agent resolves and tags a ticket — you can also submit those outcomes as labeled samples through the API and manage sample selection yourself. That more technical path is covered in Build a feedback loop.
Using confidence scores to prioritize review
Not all predictions need review. Focus your effort where the model is least certain.
- High confidence (above ~0.90): Usually correct. Spot-check occasionally.
- Medium confidence (0.60–0.90): Worth reviewing when possible.
- Low confidence (below ~0.60): Review these first — corrections here improve the model the most.
See Confidence scores and thresholds for a detailed discussion.