How one step works follows a recorded step (Team room calendar, step 1) from the raw screen through the DOM's list of clickable elements, the numbered boxes and the exact request, to image tokens, one forward pass per question, probabilities and the decision; boxes come from the page's DOM or a canvas app's hit regions, never a detector model. Every step of a run now shows what blink got: the box source, earlier actions and the questions word for word, with the request JSON. Data: cua/pipeline.json (build_pipeline.py, checked pixel for pixel) and refreshed step records in the dataset (5da6693d).
One visual system for all tabs: clean cards and controls, compact navigation, readable typography, aligned probability displays, sharp document previews and proper disclosures. Removes nested gray wrappers and closed-panel scrollbars. Responsive figures preserve methodology disclosures. Inference, APIs, weights and model readout are unchanged.
Keep PDF preparation independent of the model runtime
PDF classification and packet splitting with local English OCR, original-page inspection, category rules, review flags, JSON and PDF exports, and a same-text 4B comparison. Five primary navigation entries preserve existing deep links. Bundled fictional examples use real recorded model outputs. Published model weights are unchanged.
GitHub and Docs beside the name on every tab; Wire format docs under the API tab's lede; GitHub and Docs beside the API tab's steps and at the end of How it works' Running it yourself; a Code/Docs line on the Space card. Each opens in a new tab.
API tab: the Space reply shows a live answer instead of placeholders
