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 type { QuestionDef, SystemAnswer } from "./agent.js";
/** One-line question factories; output always passes checkQuestion. */
export const q = {
choice(instructions: string, criteria: Record<string, unknown> | string[]): QuestionDef {
return { type: "choice", instructions, criteria } as QuestionDef;
},
score(instructions: string, criteria: unknown[]): QuestionDef {
return { type: "score", instructions, criteria } as QuestionDef;
},
noul(instructions: string, criteria?: Record<string, unknown>, labels?: { false: string; true: string }): QuestionDef {
const out: Record<string, unknown> = { type: "noul", instructions };
if (criteria !== undefined) out["criteria"] = criteria;
if (labels !== undefined) out["labels"] = labels;
return out as QuestionDef;
},
};
/** Argmax label: choice string, score key, noul boolean. */
export function best(answer: SystemAnswer): string | boolean {
if (answer.type === "choice") return answer.choice;
if (answer.type === "score") {
let bk = "";
let bv = -Infinity;
for (const [k, v] of Object.entries(answer.probabilities)) {
if (v > bv) {
bv = v;
bk = k;
}
}
return bk;
}
return answer.noul >= 0.5;
}
/** Top-k probability keys, descending. */
export function topK(probs: Record<string, number>, k: number): string[] {
return Object.entries(probs)
.sort((a, b) => b[1] - a[1])
.slice(0, Math.max(0, k))
.map(([key]) => key);
}
/** Confidence gate (default 0.8) on answer_confidence (max(p)) alone, so the
* gate uses one scale across question types instead of mixing in entropy. */
export function isConfident(
answer: Pick<SystemAnswer, "answer_confidence">,
thresh = 0.8,
): boolean {
return (answer.answer_confidence ?? 0) >= thresh;
}