Interactive CPU lab · original synthetic spectra

Hide a spectrum. Inspect the learning signal.

A tiny ridge model really fits on 72 training curves. Separate validation and inspection queries reveal its errors. The second panel explains contrastive arithmetic without training an encoder. This is not a trained foundation model.

1. Predict the hidden wavelengths

● Visible input · × Ridge prediction · solid line: target for inspection · shaded bands: hidden

Hidden-band MSE: training 1.514e-4 · validation 1.367e-4 · inspection query 5.666e-5. Validation mean-only baseline: 6.941e-3. Lower is better for this reconstruction task only.

Errors use only hidden channels and have squared signal units. Each mask setting fits a separate model; a lower query score after exploration is not final-test evidence.

2. What should count as a positive view?

Solid line: clean query · dashed line: transformed view

Reversing values while keeping wavelength metadata fixed changes spectral meaning. This is an intentionally invalid positive view, not a recommended augmentation.

One-anchor contrastive loss: 1.339962. Positive share: 26.19%. These are loss terms, not task accuracy.

Bars are denominator softmax shares, not class probabilities or confidence. Raw unit-normalized spectra act as demo embeddings. Three negatives are fixed training curves. Positive gain leaves cosine similarity unchanged; no encoder learned this invariance. Jitter is a sine perturbation, not a calibrated sensor-noise model.

Download runnable CPU source · Download all synthetic spectra

Static defaults shown. JavaScript enables controls; the Python code runs independently.