Assured Prediction
Release: Beta — core workflows available; some capabilities still maturing
Entropy-based selective prediction for classifiers
Assured Prediction is a model-agnostic post-hoc filter for neural network classifiers. Given any trained model's softmax output, it computes per-prediction entropy and rejects low-confidence predictions — trading coverage for accuracy on the samples it keeps.
A 70% accurate model can deliver substantially higher accuracy on the fraction of
samples it keeps — if you let it abstain on low-confidence predictions. No retraining required.
Validated on defense SAR radar (MSTAR), medical ECG, and multi-domain benchmarks.
The AssuredFilter API ships in the production wheel.
What it does
A classifier with moderate overall accuracy can deliver much higher accuracy on the samples it is most confident about — if you let it say "I don't know" on the rest. Assured Prediction is a post-hoc filter on softmax outputs: compute entropy per prediction, keep the low-entropy (confident) fraction, route the rest to a human or fallback.
No retraining. Works on any classifier with softmax outputs. Coverage calibration is precise: ask for 50% coverage, get ~50%. Most valuable when aggregate accuracy looks acceptable but individual errors are unacceptable — defense imagery triage, clinical screening, fraud ops.
When to use it
High-stakes classification where a wrong answer is worse than no answer.
See Examples for runnable code.
Who it is for
- Defense intelligence analysts triaging SAR/EO imagery volume
- Radiologists and pathologists routing uncertain cases to specialist review
- Fraud operations teams improving alert precision via coverage targeting
What ships
- Python wheel with licensing (
assured-prediction) - C library with
AssuredFilter— entropy thresholding for 10–90% coverage targets - Guided journeys 01→06 under
python/examples/journeys/