01
Provenance
Data without a lineage does not enter the training set.
Method
A dataset is an apparatus,not a file dump.
Every sample is treated like a collision event: origin, calibration, error, reproducibility. Without those four, there is no training science.
Four delivery laws
02
Calibration
Sensors, label protocols and model versions are frozen.
03
Uncertainty
Quality is not binary. We report agreement and expert disagreement.
04
Signature
Final review is a named expert. Anonymous crowds cannot underwrite medicine or safety.
The QC pyramid
L1
Machine gates
Format, sync, missingness, distribution shift.
L2
Dual review
Independent annotators, measured agreement.
L3
Expert sign-off
Gold sets and disputes released under a name.
L4
Living ratings
The pass threshold only rises.
Methods we are opening
2026 China Medical AI Training-Data Quality White Paper
Standards, evaluation, trajectory.
Forthcoming
MedLabel-CT open lung CT samples
Expert labels and QC metadata.
Open preview
Evidence loops in high-order legal annotation
From entity to causality.
Research