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
docs(ts): name the four measured divergences in the parity table's header
Aashishbc62398c6373
diff --git a/laya-ts/tests/temperature-coercion.test.ts b/laya-ts/tests/temperature-coercion.test.tsindex 216c92db..5495938f 100644--- a/laya-ts/tests/temperature-coercion.test.ts+++ b/laya-ts/tests/temperature-coercion.test.ts@@ -2,11 +2,13 @@ import { describe, expect, it, vi } from "vitest";import { clampTemperature, TEMP_MAX, TEMP_MIN } from "../src/common.js";import { Agent } from "../src/agent.js";-// The reference asks "is this a number?" with `float()`, which answers no for null, a boolean, a-// list, an object and a string that holds nothing but blanks. `Number()` answers yes for all six and-// produces 0 or the inner element, and 0 lands on TEMP_MIN -- the sharpening end, where a published-// confidence is inflated rather than merely uncalibrated. These are the cases where the two answers-// differed, measured as `clamp_temperature(t)` / `clampTemperature(t)` on the port's own base.+// The reference asks "is this a number?" with `float()`, which refuses a string that holds nothing+// but blanks and refuses any list. `Number()` reads `Number(" ")` and `Number([])` as 0, which the+// clamp bounds to TEMP_MIN, and `Number([2])` as 2, which sits inside the range and so passes for a+// fitted temperature. 0 is the worse of the two: a temperature below 1 sharpens the logits, so the+// published confidence is inflated rather than merely uncalibrated. Those four are the divergences+// measured as `clamp_temperature(t)` / `clampTemperature(t)` on the port's own base; the rest of the+// table pins the fallbacks the port already got right, so a future coercion shortcut is caught here.const fakeProvider = () => ({async runEncoder(_b: any) { return { lastHidden: [[1, 0], [0, 1]] }; },