One photograph in, five things out — each produced by a different part of the pipeline.

Food-101 ships only train and test. The validation set is carved out of train, never out of test.

Run the expensive backbones exactly once, then never again.

A 1.9M-parameter head on a cached vector reaches 96.83% in under four minutes.

Updating all 304M parameters — and losing to the probe that took four minutes.

A parameter-free average beat a trained 3.97M-parameter fusion head.

Making the confidence figure mean what it says, then attaching a guarantee that can be falsified.

Forty-one dishes had a USDA record. Sixty did not, and were rebuilt from ingredients.

Three retrieval stages, then a check that no number reaches the reader unsupported.

Where the trained artefacts actually run.