The library exists.
The product didn't.
Calibration math is well-known and open source. Wiring it into a maintained, monitored, hosted workflow — with diagrams, scoring, and drift alerts — is the part nobody finishes. That's Calibrate.
| Capability | Empire Calibrate | sklearn CalibratedClassifierCV | MLflow / W&B | DIY notebook |
|---|---|---|---|---|
| Isotonic + Platt + beta, auto-selected | ✓ | isotonic/sigmoid, manual | ✗ | DIY |
| Brier / log-loss / ECE reporting | ✓ built-in | compute yourself | log a metric | DIY |
| Reliability diagrams | ✓ rendered | ✗ | if you log it | DIY |
| Live drift detection + alerts | ✓ | ✗ | custom | ✗ |
| Hosted API, no model code shared | ✓ | — | ✓ | — |
| Model-card / compliance export | ✓ | ✗ | manual | ✗ |
vs sklearn
scikit-learn gives you the calibration estimators — and that's it. You still own validation, the diagrams, the scoring, the drift monitoring, and keeping it running. Calibrate is all of that, hosted, so your team ships the decision instead of the plumbing.
vs MLflow / Weights & Biases
Experiment trackers log whatever you compute — they don't calibrate or monitor calibration for you. Calibrate complements them: it does the calibration + drift, and you log its outputs to your tracker.
vs DIY
Every team can write a one-off isotonic fit. Few maintain reliability diagrams, ECE, multi-method selection, and live drift alerts. Calibrate is the maintained version so calibration stops being a someday ticket.
Comparison reflects publicly documented capabilities as of 2026, provided for evaluation. All product names and trademarks belong to their respective owners.