Define the target
Choose one electronic response, units and primary score. Match each structural measurement to the same specimen.
Methods →Let the prediction
take the risk.
Compare a candidate with a uniform baseline and a conventional model on test specimens. Better stories need better predictions.
This scores supplied predictions; it does not fit a model or verify that predictions were frozen in advance. Use the same outcome units and genuinely independent specimen IDs.
specimen_id,split,observed,uniform,conventional,candidate A,test,10,8,9.8,9.2
Exact header required. Split must be train or test. Test outcomes alone are scored. A specimen appearing in both splits is rejected. Repeated measurements within a specimen are averaged before specimens receive equal weight.
—Independent test specimens
—Test observations
—Lowest RMSE in this input
| Model | Equal-specimen RMSE | Candidate improvement vs model | 95% bootstrap interval |
|---|
Positive improvement favors the candidate; negative favors the reference. The synthetic candidate beats the uniform baseline and loses to the conventional model. That is an intentional illustration of why both comparisons matter.
Make the test credible before seeing its answer.
Choose strong rivals
Specify a uniform baseline and a material-specific conventional model. Calibrate all models using training data only.
Methods →Freeze predictions
Record the code, fitted parameters, data split and predictions before test outcomes are opened. This browser cannot establish that history.
Methods →Report both directions
Publish favorable and unfavorable results. Do not let repeated pixels or repeated measurements masquerade as independent specimens.
Methods →