Browser Decider

A decision model in about 10 MB, running on your GPU in this tab. Give it a state (a support thread, an invoice, a security alert, an agent's run log) and a few typed questions: choice (pick one label), score (a level on a rubric) or noul (is this statement true). It returns a probability for every option from one forward pass, with no text generated and no server.

The model is Ettin-17m, a 17M-parameter encoder, fine-tuned on the Typed Decisions benchmark's training split, so it answers these four workflows' questions well and other schemas poorly (see the write-up). Grey bars are the benchmark's gold distribution, the mean of three teacher samples; blue bars are this model.

Your own state and questions

Questions use the System One request shape: { name: { type, instructions, criteria } }, criteria a map of option to description for choice, a list of level descriptions for score, optional { "true": …, "false": … } for noul. Off the four trained workflows, expect weak answers.