Zero-shot prompt routing
Categories are plain prose you write here — nothing is trained or configured. The encoder reads the category list and the text together in a single bidirectional pass, then scores the text against every category at once.
- Results appear as you type.
Zero-shot policy linting
Write your rules as free text. Every token in the draft is scored against every rule in one pass, so adding a rule costs a handful of tokens rather than another model.
Scoring is one pass over the whole draft against every rule, which is seconds of work under WASM — so it runs when you press Lint rather than on every keystroke. The threshold slider re-thresholds the last result without re-scoring it.
Press Lint to score the draft.
Draft or policy has changed — press Lint to re-score.
Fill‑mask
The base encoder with its masked-LM head. Bidirectional, so each blank is predicted from the whole sentence rather than only what comes before it.
This panel loads the base encoder, a separate 424 MB download from the routing and linting heads.
Results appear as you type.
Masked diffusion
The same encoder, fine-tuned to answer questions by denoising. Instead of emitting one token after another, it starts from a canvas of blanks and fills them in over a handful of passes — keeping only the words it is confident about and re-guessing the rest, now able to see what landed on both sides of them.
This panel loads the diffusion checkpoint, a separate 424 MB download. Every pass is a full forward over the canvas, so expect a second or two per pass under WASM.
Press Generate to watch the answer condense.