# Reference capstone · measured results

This is a report of the executed reference configuration, not a required score for every learner. Values are from the bundled output JSON.

## Protocol

Predict the fictional nonlinear sensor response from two features. Use independently generated regression rows and fixed random 360/120/120 partitions. Fit mean/population standard deviation on training only. Compare a linear model and a two-hidden-layer width-16 MLP, each with 100 epochs. Select checkpoint and model by validation MSE. The visible bundled test is a teaching fixture, not a hidden assessment.

## Selection and final assessment

- linear: best epoch 57; validation MSE 1.26921833, RMSE 1.126596

- mlp: best epoch 99; validation MSE 0.02235237, RMSE 0.149507

Selected mlp. On 120 test examples, MSE 0.02597234, RMSE 0.161159, MAE 0.125866. The constant train-mean baseline predicts 1.180640 and has test RMSE 2.285362. Only the already-selected learned model is evaluated on test; the naive baseline is fixed from train.

## Interpretation and limitations

The nonlinear candidate better represents this generator’s curved relationship. The comparison is consistent with improved fit, not proof of a causal feature effect or real sensor utility. The fixed synthetic process has low noise and no deployment drift. A real application would require availability-aware features, realistic group/time holdouts, error-cost decisions and external validation. Repeated reference runs verify execution, not additional independent test evidence.

## Classification extensions

rings: selected mlp by validation loss. Test accuracy 1.0000 on 120 examples; confusion counts {'tn': 57, 'fp': 0, 'fn': 0, 'tp': 63}. Fixed train-majority baseline test accuracy 0.4750.

bars: selected cnn by validation loss. Test accuracy 1.0000 on 120 examples; confusion counts {'tn': 57, 'fp': 0, 'fn': 0, 'tp': 63}. Fixed train-majority baseline test accuracy 0.4750.

Both classification fixtures are deliberately easy, balanced class-conditional samples. The bars linear baseline also achieved perfect validation accuracy; the CNN was selected by lower validation loss, not a demonstrated accuracy advantage. Zero test errors on these fixtures do not establish real-world performance, robustness or calibrated confidence.

## Inspect the underlying files

See outputs/reference_results.json for full candidate histories, split IDs, settings and metrics. See outputs/reference_capstone/selection.json and final.json for the two-stage capstone record. These are reference outputs, not replacements for learner implementation or explanation.