How it works
Observable signals in. Honest error bars out.
The pipeline is built so the estimator can never cheat. It reads only what a BMS actually emits, its features never see the latent truth, and its uncertainty is calibrated on held-out packs rather than asserted.
Model card soh-quantile-1.0.1The error bars are published. Read the methodology.The scoring pipeline, stage by stage
- 01
Observable BMS signals only
Equivalent full cycles, temperature exposure, an internal resistance proxy, coulombic efficiency, depth of discharge, fast-charge fraction. Signals every compliant pack already emits.
- cycles
- temperature
- resistance proxy
- coulombic efficiency
- depth of discharge
- fast-charge fraction
- 02
A feature layer that never sees the truth
Features are computed from the observables alone. The latent state of health stays hidden from the estimator, in training and in production.
- 03
Gradient-boosted quantile regression
The model predicts a P10, P50, and P90 for every pack: a central estimate inside an explicit band, never a bare point.
- P10
- P50
- P90
- 04
Conformal calibration
The band is widened on held-out calibration packs (conformalized quantile regression), so quoted coverage is honest: a stated 80 percent band covers close to 80 percent.
- 05
The decision lenses
The calibrated estimate becomes the six lenses: grades, values, and risk figures, one per decision that touches the pack.
- 06
Reason codes and drift monitoring
Every score ships with its attribution, and the fleet is watched for drift that calls for recalibration.
Run the pipeline on your fleet.
A pilot scores a sample of your packs end to end, with the same calibrated bands and reason codes described here.