runbios-sdk 0.2.1-dev.134 → 0.2.1-dev.136

package/dist/index.d.ts

@@ -36,7 +36,7 @@ import { Loop } from './resources/loop.js';
  * SDK version. Sent as part of the User-Agent header.
  * Must match package.json "version" -- enforced by a contract test.
  */
-export declare const VERSION = "0.2.1-dev.134";
+export declare const VERSION = "0.2.1-dev.136";
 export declare class RunBiOS {
     /** Search models, fetch configs, check adapter compatibility. */
     readonly models: Models;

package/dist/index.js

@@ -36,7 +36,7 @@ import { Loop } from './resources/loop.js';
  * SDK version. Sent as part of the User-Agent header.
  * Must match package.json "version" -- enforced by a contract test.
  */
-export const VERSION = '0.2.1-dev.134';
+export const VERSION = '0.2.1-dev.136';
 export class RunBiOS {
     /** Search models, fetch configs, check adapter compatibility. */
     models;

package/dist/resources/loop.d.ts

@@ -312,6 +312,13 @@ export declare class Loop {
      * twice the weight, and the combined score is the weighted mean over the
      * rules that actually applied.
      *
+     * `matches_gold` needs no `expected`: it compares each answer to the gold
+     * answer recorded on that same conversation, or the human correction when
+     * there is no gold, and scores sampled alternatives against the same gold.
+     * A conversation with neither is skipped, not failed, so one rule checks
+     * every labelled question without punishing the unlabelled ones. An
+     * optional `tolerance` widens the match when both sides are bare numbers.
+     *
      * A caution worth knowing before you write a set: a rule made only of
      * `forbidden` phrases is satisfied VACUOUSLY by an answer that says nothing.
      * Pair it with a `required` phrase, or you are rewarding silence.
@@ -408,8 +415,9 @@ export declare class Loop {
      */
     agentStatus(): Promise<LoopAgentStatus>;
     /**
-     * Turn the agent on: mints the workspace's managed serverless key. After
-     * this, automatic judges and sample runs make model calls billed to the
+     * Turn the agent on: mints the workspace's managed serverless key and
+     * resumes the automatic judges a previous turn-off paused. After this,
+     * automatic judges and sample runs make model calls billed to the
      * workspace. Idempotent.
      *
      * `monthly_spend_cap_cents` caps what that key may spend on model calls in
@@ -423,9 +431,11 @@ export declare class Loop {
         monthly_spend_cap_cents?: number;
     }): Promise<LoopAgentCredential>;
     /**
-     * Turn the agent off: revokes its key and sets every automatic judge back to
-     * manual. Open runs stop where they are and continue if it is turned back
-     * on. Nothing already scored or written is removed.
+     * Turn the agent off: revokes its key and pauses every automatic judge in
+     * the workspace (`judges_paused`; each judge shows `auto_paused`). Turning
+     * it back on resumes exactly those judges. Open runs stop where they are
+     * and continue if it is turned back on. Nothing already scored or written
+     * is removed.
      */
     disableAgent(): Promise<{
         revoked: boolean;

package/dist/resources/loop.js

@@ -417,6 +417,13 @@ export class Loop {
      * twice the weight, and the combined score is the weighted mean over the
      * rules that actually applied.
      *
+     * `matches_gold` needs no `expected`: it compares each answer to the gold
+     * answer recorded on that same conversation, or the human correction when
+     * there is no gold, and scores sampled alternatives against the same gold.
+     * A conversation with neither is skipped, not failed, so one rule checks
+     * every labelled question without punishing the unlabelled ones. An
+     * optional `tolerance` widens the match when both sides are bare numbers.
+     *
      * A caution worth knowing before you write a set: a rule made only of
      * `forbidden` phrases is satisfied VACUOUSLY by an answer that says nothing.
      * Pair it with a `required` phrase, or you are rewarding silence.
@@ -579,8 +586,9 @@ export class Loop {
         return this._http.fetchGet('/api/loop/agent');
     }
     /**
-     * Turn the agent on: mints the workspace's managed serverless key. After
-     * this, automatic judges and sample runs make model calls billed to the
+     * Turn the agent on: mints the workspace's managed serverless key and
+     * resumes the automatic judges a previous turn-off paused. After this,
+     * automatic judges and sample runs make model calls billed to the
      * workspace. Idempotent.
      *
      * `monthly_spend_cap_cents` caps what that key may spend on model calls in
@@ -598,9 +606,11 @@ export class Loop {
         return res.credential;
     }
     /**
-     * Turn the agent off: revokes its key and sets every automatic judge back to
-     * manual. Open runs stop where they are and continue if it is turned back
-     * on. Nothing already scored or written is removed.
+     * Turn the agent off: revokes its key and pauses every automatic judge in
+     * the workspace (`judges_paused`; each judge shows `auto_paused`). Turning
+     * it back on resumes exactly those judges. Open runs stop where they are
+     * and continue if it is turned back on. Nothing already scored or written
+     * is removed.
      */
     async disableAgent() {
         return this._http.fetchDelete('/api/loop/agent');

package/dist/types.d.ts

@@ -2672,8 +2672,16 @@ export interface LoopCandidate {
     metadata: Record<string, unknown>;
     created_at: string;
 }
-/** The deterministic checks a grader can perform. */
-export type LoopGraderKind = 'exact_match' | 'contains' | 'regex' | 'json_valid' | 'numeric' | 'tool_called';
+/**
+ * The deterministic checks a grader can perform.
+ *
+ * `matches_gold` is the one kind with no expected value of its own: it
+ * compares each answer to the gold answer recorded on that same conversation
+ * (or the human correction when there is no gold), scores sampled
+ * alternatives against the same gold, and is NOT APPLIED to a conversation
+ * that carries neither, so it never marks down an unlabelled answer.
+ */
+export type LoopGraderKind = 'matches_gold' | 'exact_match' | 'contains' | 'regex' | 'json_valid' | 'numeric' | 'tool_called';
 export interface LoopGraderConfig {
     expected?: string;
     /** Every phrase that must appear. */
@@ -2683,7 +2691,10 @@ export interface LoopGraderConfig {
     pattern?: string;
     /** Keys the answer must carry, for `json_valid`. */
     keys?: string[];
-    /** How far from `expected` still counts, for `numeric`. */
+    /**
+     * How far from `expected` still counts, for `numeric`. For `matches_gold`,
+     * how far from the gold answer still counts when both are bare numbers.
+     */
     tolerance?: number;
     /** The function that must have been called, for `tool_called`. */
     function?: string;

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "runbios-sdk",
-  "version": "0.2.1-dev.134",
+  "version": "0.2.1-dev.136",
   "description": "Official TypeScript SDK for the Run BiOS training and deployment platform API",
   "type": "module",
   "main": "./dist/index.js",