Every field the legacy IDP got wrong and the pipeline got right, with what the legacy OCR read on the line its rule used. Click a row to run that document.
The pipeline, then every module of the model in the order it runs: each card is a layer with weights (dense layers, norms, the convolution, the embedding table), dashed cards are steps written as code (attention, residual adds, 2 × 2 pooling). Each shows its real output on this page and a picture of its weights. Click a card to step there; ▸ opens a part, ▾ folds it. Keys: ← → step, space play.