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
feat(sdk): read the confidence an answer is gated on and the abstention report
Aashish7af12f470188
diff --git a/sdk/typescript/src/response.ts b/sdk/typescript/src/response.tsindex d6eb8468..8180c6f2 100644--- a/sdk/typescript/src/response.ts+++ b/sdk/typescript/src/response.ts@@ -47,6 +47,25 @@ export function validatePrediction(value: unknown, questions: Questions): void {if (answer.action !== undefined) {expect(isRecord(answer.action) && probability(answer.action.act_probability), `answers.${id}.action`);}+ // Checked when the server sends them, like the optional fields of `health` and of `usage`:+ // `answer_confidence` (#126) and the gate fields (#361) are recent additions to an answer that+ // predates them, so requiring a key here would fail every prediction against a deployment that+ // has not shipped it. A caller that asked for a gate (`minConfidence`) and got no `abstention`+ // back reads it as `undefined` on the typed answer, which is the truth: the gate did not run.+ if (answer.answer_confidence !== undefined) {+ expect(probability(answer.answer_confidence), `answers.${id}.answer_confidence`);+ }+ if (answer.low_confidence !== undefined) {+ expect(answer.low_confidence === true, `answers.${id}.low_confidence`);+ }+ if (answer.abstention !== undefined || answer.abstention_threshold !== undefined) {+ // The gate writes both onto the same answer or onto neither, so one without the other is a+ // half-report: an `abstention` with no echoed threshold cannot be re-split, and a threshold+ // with no state says nothing about the answer it was measured against.+ expect(answer.abstention === 'passed' || answer.abstention === 'abstained' ||+ answer.abstention === 'unevaluated', `answers.${id}.abstention`);+ expect(probability(answer.abstention_threshold), `answers.${id}.abstention_threshold`);+ }if (question.type === 'noul') {expect(probability(answer.noul), `answers.${id}.noul`);} else {diff --git a/sdk/typescript/src/types.ts b/sdk/typescript/src/types.tsindex 5a0b62a3..ac3d7403 100644--- a/sdk/typescript/src/types.ts+++ b/sdk/typescript/src/types.ts@@ -53,6 +53,20 @@ export interface PredictOptions extends RequestOptions {interface AnswerBase {/** Optional action metadata returned by Laya. */action?: { act_probability: number };+ /** The probability mass on the answer actually reported (`max(p)`). Unlike `confidence`, which+ * means normalized entropy on `choice` and `score` and `max(p_yes, p_no)` on `noul`, this is+ * the same quantity on all three types, so one threshold gates across them. */+ answer_confidence: number;+ /** Set by the abstention gate on the answers that fell below `minConfidence`, and only then:+ * `flag_low_confidence` writes `True` and never writes the key otherwise. */+ low_confidence?: true;+ /** How the gate decided this answer, reported on every answer of a gated call. Absent means no+ * gate ran (`minConfidence` unset); `unevaluated` means a gate ran and this answer carried no+ * usable confidence, which the boolean above cannot tell apart from a pass. */+ abstention?: 'passed' | 'abstained' | 'unevaluated';+ /** The threshold that produced `abstention`, echoed on the same answers: the gate consumes+ * `minConfidence`, so without this a batch run cannot be re-split by the threshold it used. */+ abstention_threshold?: number;}export interface ChoiceAnswer<Label extends string = string> extends AnswerBase {type: 'choice';