Decisions, Jev-style

A System One decision is the fast kind: read the situation, pick one of the allowed options, say how sure you are. Jev made the shape popular — send a state and typed questions, get back a typed answer with a calibrated probability for every option. No text is generated, so there is nothing to parse and no way to answer outside the options you gave.

toolnexus-web does this in the tab, on WebGPU or CPU.

Jev plays Snake

Every move below is one decision, made on your GPU in this tab: the board goes in as one sentence, the moves that do not crash go in as the allowed options, and the model scores them. Load the model once (348 MB, cached after that), then press Jev plays — or play it yourself. The code panel under the board is the exact call it just made.

A model trained to decide — NexusDecider

import { NexusDecider } from 'toolnexus-web';

// open-jev: DeBERTa-v3, trained only on typed decisions. 348 MB (q4f16) on WebGPU.
const decider = await NexusDecider.load();

const { answers, ms } = await decider.systemOne(ticket, {
  team:   { type: 'choice', instructions: 'Which team handles this?', options: ['billing', 'support', 'sales'] },
  urgent: { type: 'noul',   instructions: 'The customer needs help right now.' },
  mood:   { type: 'score',  instructions: 'How upset is the customer?', options: ['calm', 'annoyed', 'furious'] },
});

answers.team.choice;          // 'billing'
answers.team.probabilities;   // { billing: 0.94, support: 0.03, sales: 0.03 }
answers.urgent.noul;          // 0.12  — P(true)
answers.mood.score;           // 1.4   — expected level, 0 = first option

Every question is answered in the same forward pass. The state can be a string or an object (serialized as JSON) and is cut to 256 tokens; the whole input is capped at 512.

NexusDecider needs Transformers.js 4.x and loads it itself. The version the chat side uses (3.8.1) ships an ONNX Runtime that rejects this model’s graph with a bare exception and no message — NexusDecider names that failure instead of passing the number through.

The model’s authors measure 0.854 accuracy in-domain and 0.690 on unseen instructions and option sets, with calibrated confidence (ECE 0.022). Out of distribution, measure on your own data before trusting it.

Any chat model — chat.decide()

The lighter path: no second model, just the chat model you already loaded. Options are labelled A, B, C… and scored by the logit of each letter at the start of the answer.

const d = await chat.decide(
  'Email: Payroll asks for your password on a non-company sign-in page.',
  ['Legitimate', 'Spam', 'Phishing'],
);
d.choice;    // always one of the three
d.options;   // [{ choice, label, probability, logprob, logit }, …]

It reproduces the published readout from “Jev in 25 lines of Python” exactly, and runs in about 46 ms on WebGPU for a 50-token prompt. The conversation history is not touched.

Which one — measured, not assumed

Same 60 random Snake boards, WebGPU, one decision each: which way gets the snake closer to the food?

Engine	Toward the food	Per decision
chat.decide() on Qwen3-0.6B	45%	~66 ms
a random move that does not crash	49%	—
NexusDecider (open-jev)	100%	36 ms

A small chat model is fast at this and weak at it — SemIf measured Qwen3-0.6B at 40.7% agreement with Jev on its public subset. Use chat.decide() when a second download is not worth it and the choice is easy; use NexusDecider when the decision matters.

Wording matters too: for the model above, “the food is 3 to the right and 2 up” scored 100%, and “the food is up and right” scored 47%. Measure your phrasing the same way.