Non-autoregressive System 1 decision engine. Typed choice, score and yes/no decisions over any text in a single forward pass, in 100+ languages, with a router that picks the right checkpoint per request.
Python calibrationclassificationdecision-modelhuggingfacejevmodernbertmultilingualnlppythonpytorchroutingtyped-decisionszero-shot
import { describe, expect, it } from "vitest";
import { checkQuestion } from "../src/agent.js";
import { q, best, topK, isConfident } from "../src/dx.js";
import { predictShortlist } from "../src/shortlist.js";
describe("q builders", () => {
it("produce valid defs", () => {
const c: any = q.choice("What?", { a: "x", b: "y" });
const s: any = q.score("Rate?", ["bad", "ok"]);
const n: any = q.noul("Holds?");
expect(() => checkQuestion("c", c)).not.toThrow();
expect(() => checkQuestion("s", s)).not.toThrow();
expect(() => checkQuestion("n", n)).not.toThrow();
expect(c.criteria).toEqual({ a: "x", b: "y" });
});
it("choice accepts label list", () => {
const c: any = q.choice("Pick?", ["a", "b"]);
expect(() => checkQuestion("c", c)).not.toThrow();
});
});
describe("result helpers", () => {
it("best/topK/isConfident", () => {
expect(best({ type: "choice", choice: "b", probabilities: { a: 0.2, b: 0.8 } } as any)).toBe("b");
expect(best({ type: "score", probabilities: { "0": 0.1, "1": 0.7, "2": 0.2 } } as any)).toBe("1");
expect(topK({ a: 0.2, b: 0.8, c: 0.5 }, 2)).toEqual(["b", "c"]);
expect(isConfident({ answer_confidence: 0.9 } as any, 0.8)).toBe(true);
expect(isConfident({ answer_confidence: 0.5 } as any, 0.8)).toBe(false);
});
it("gates on answer_confidence alone, ignoring entropy confidence", () => {
expect(isConfident({ confidence: 0.5, answer_confidence: 0.9 } as any, 0.8)).toBe(true);
expect(isConfident({ confidence: 0.9, answer_confidence: 0.5 } as any, 0.8)).toBe(false);
expect(isConfident({ confidence: 0.9 } as any, 0.8)).toBe(false);
});
});
describe("predictShortlist auto-embed", () => {
it("uses agent encoder when embedFn omitted", async () => {
const agent: any = {
tok: { encode: (t: string) => [t.length % 100], padId: 0 },
provider: {
async runEncoder(batch: any) {
return { lastHidden: batch.inputIds.map((row: number[]) => row.map(() => [1, 0])) };
},
},
async predict(_s: unknown, reduced: any) {
return { model: "m", answers: {}, usage: { input_tokens: 0, output_tokens: 0 }, _reduced: reduced };
},
};
const questions: any = {
d: { type: "choice", instructions: "pick", criteria: { aaa: "x", bbb: "y", ccc: "z" } },
};
const out: any = await predictShortlist(agent, "state", questions, undefined as any, 1);
expect(out.shortlist.d.labels).toHaveLength(1);
});
});