Seeded data generation, separate training and holdout sets, Gaussian-elimination ridge fitting, two feature maps, schema rejection, standardized temperature-mean shift, error gates and JSON evidence. All computation stays in this browser.
EVAL-DRIVEN. MONITORED FOR DRIFT.
A model is fitted.
Should it ship?
Train a cooling-power predictor. Stress the inputs.
Keep a weak release out of the deployment lane.
02 / RELEASE DECISION
Ready for review
03 / HOLDOUT ERROR
Same data. Two models.
Mean absolute error on accepted holdout rows only. Rejected inputs never receive a prediction.
04 / INSPECT THE EVIDENCE
Predicted vs observed
The dashed line is a perfect prediction. Each point is one accepted synthetic holdout row.
Row-level trace
All 60 holdout rows are included in the export.
| Source | °C | Load % | Seen kW | Pred. kW | Status |
|---|
THE BOUNDARY BETWEEN A PROTOTYPE AND PRODUCTION
What this proves.
What it doesn't.
Temperature and load feed a quadratic power equation with seeded noise. A quadratic model is expected to fit this constructed signal better. This is not a general model benchmark, Freecool case study or measured client saving.
Replace generated rows with authorized, versioned telemetry. Fit on time-based splits, validate target labels, monitor more features and cohorts, then connect a reviewed registry and shadow deployment. No production connection is implemented.
Why NXL: its AI & Machine Learning practice describes versioned, evaluated predictive models and drift monitoring. This prototype is an independent application sample, not an NXL product.