Research
We test uncertainty and risk control on public datasets, and develop the theory for AI systems that change.
Inside a validation · Chákṣu → RIGA
Local recalibration reduced omitted anatomy on a new dataset, at the cost of larger regions.
Read the studyObserved mean omission
The cost: 4.88× mean region area.
Same 182 evaluation images · Same frozen model · Six expert references
Bars share a 0–25% scale. Image-average omission, not patient-level risk.
Validation studies
Completed experiments that test the technology. Each has its own reference, risk definition and scope.
NEISS
4.93% set misses; a single label on 50.50% of 10,000 held-out records.
NTSB
6.43%, 4.63% and 5.40% observed set misses for three calibration vintages.
HAM10000
4.49% reference-label misses; 2.24 labels per lesion out of seven.
CheXpert
0 misses in 60 studies; only 8 studies received a single answer.
CBIS-DDSM
4.21% reference misses; one answer for 24.91% of participants.
LIDC-IDRI
2.75% patient-average rating exclusion; 3.79 of five ratings retained.
Chákṣu
5.37% observed omission; regions averaged 1.28× the expert cup area.
Chákṣu → RIGA
19.44% to 3.45% omission after local recalibration, with 4.88× region area.
PolypGen
33.22% versus 48.91% image area for two calibrated methods at a 5% target.
The risk quantity and limit chosen before evaluation.
Observed errors and the cost in review, set size or region area.
The statistical statement supported by calibration, under its assumptions.
Publications
MICCAI 2026 · UNSURE workshop · Research paper
Propagating uncertainty through multi-stage clinical pathways while preserving safe alternatives.
Read the paperNeurIPS 2025 · GenAI for Health workshop · Position paper
A perspective on uncertainty-aware AI, clinical trust and the European setting.
View the publicationResearch agenda
Our scientific program for deployed, adaptive AI systems.
We welcome research conversations and collaborations.
EyeTrustAI
Language & human judgment
HumAID
How much review can a calibrated policy remove?
13.88% removed from immediate review; 3.39% priority content within that subset.
GoEmotions
What does a stricter error limit cost in automation?
3.86% to 35.86% auto-tagging as the chosen limit moves from 5% to 20%.
TweetEval
One sentiment model. Different error budgets.
6.39% to 43.41% auto-tagging across the supported sentiment budgets.