Moderation QA
Confusion Matrix
Upload a labeled CSV and see where decisions diverge. Works for QA on your moderation team, an LLM judge, a classifier, or any labeling workflow.
Two columns is all you need: the correct label and the label that was actually assigned. We’ll match them up, count where they agree and disagree, and draw the matrix. Sample data is loaded below so you can see the shape right away.
Everything runs in your browser. Your CSV is never uploaded anywhere.
Matrix
Green cells along the diagonal are correct calls. Coral cells off the diagonal are mismatches; deeper coral means more of them. Empty cells are pairs that never occurred.
Drugs | Harassment | Hate | Safe | Self-Harm | Selling | Sexual | Underage | Violence | |
|---|---|---|---|---|---|---|---|---|---|
| Drugs | |||||||||
| Harassment | 1 | 3 | |||||||
| Hate | 17 | 2 | |||||||
| Safe | 1 | 2 | 6 | 113 | 1 | 1 | 4 | 2 | 7 |
| Self-Harm | 2 | 4 | |||||||
| Selling | |||||||||
| Sexual | 1 | 2 | 24 | 1 | |||||
| Underage | |||||||||
| Violence | 2 | 3 |
This is a free mini-tool.
On the Musubi platform, you can run evals like this against your production policy and data, identify drift over time, and ship enforcement changes with confidence. This one is for a single snapshot.
See the full platform