# Ridge EEG diagnostic figures

**Status:** schematic demo figures, not benchmark evidence.

**Prediction source:** schematic recomputed linear_ridge on synthetic demo data.

## Critical interpretation note

The repo's current `linear_ridge` baseline fits a ridge model after reshaping `[windows, time, channels]` into `[windows*time, channels]`. It therefore learns a regularized channel-to-channel linear map, not a full time-by-channel flattened-window model. High performance should be interpreted as evidence that near-future EEG is linearly predictable from recent channel covariance/temporal continuity under the current split, not as proof of a rich neural state model.

## Generated files

- `fig1_ridge_input_target_waveforms.png`
- `fig2_ridge_prediction_overlay.png`
- `fig3_autocorrelation_lag_structure.png`
- `fig4_ridge_coefficient_channel_map.png`
- `fig5_psd_residual_diagnostics.png`

## Metrics from plotted prediction

- `global_mse`: 0.213588
- `global_mae`: 0.37722
- `global_pearsonr`: 0.655473
- `n_train_windows`: 320
- `n_test_windows`: 80
- `window_timepoints`: 128
- `n_channels`: 8
- `sfreq`: 128

## Provenance

- `source`: `synthetic schematic generated by plotting script`

## Optional outputs skipped

- skipped topomap: schematic channels do not have valid sensor positions