runbios-sdk 0.3.0-rc.325 → 0.3.0-rc.330

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 = "0.3.0-rc.325";
+export declare const VERSION = "0.3.0-rc.330";
 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 = '0.3.0-rc.325';
+export const VERSION = '0.3.0-rc.330';
 export class RunBiOS {
     /** Search models, fetch configs, check adapter compatibility. */
     models;

package/dist/resources/loop.d.ts

@@ -382,7 +382,8 @@ export declare class Loop {
      * it made, winner or not, each the shape `attempts` has on the pipeline
      * (with its `explanation` in plain words). `limit` is 50 when absent and at
      * most 200. Narrow by `result` (`became_version`, `not_better`,
-     * `inconclusive`, `failed`, `stopped`, `running` -- which includes one
+     * `inconclusive` -- tested, could not decide --, `not_tested` -- ended
+     * before it was tested --, `failed`, `stopped`, `running` -- which includes one
      * waiting for a decision -- or `waiting_for_review`) and by `model` (`main`,
      * or `second` for the second model's attempts); `total` follows the
      * filters.

package/dist/resources/loop.js

@@ -633,7 +633,8 @@ export class Loop {
      * it made, winner or not, each the shape `attempts` has on the pipeline
      * (with its `explanation` in plain words). `limit` is 50 when absent and at
      * most 200. Narrow by `result` (`became_version`, `not_better`,
-     * `inconclusive`, `failed`, `stopped`, `running` -- which includes one
+     * `inconclusive` -- tested, could not decide --, `not_tested` -- ended
+     * before it was tested --, `failed`, `stopped`, `running` -- which includes one
      * waiting for a decision -- or `waiting_for_review`) and by `model` (`main`,
      * or `second` for the second model's attempts); `total` follows the
      * filters.

package/dist/types.d.ts

@@ -2967,9 +2967,22 @@ export interface LoopMetricPoint {
     judge_candidate_mean: number | null;
     /** The trainer's loss on its validation slice, as {@link Evaluation.trainer_eval_loss}. */
     trainer_eval_loss: number | null;
-    /** The bar this attempt had to clear, frozen with its report. */
+    /**
+     * The pipeline's margin frozen with this attempt's report: the most the
+     * mean-change bar asks. The bar it actually had to clear is
+     * {@link LoopMetricPoint.mean_change_needed}.
+     */
     promote_margin: number | null;
     promote_min_win_rate: number | null;
+    /**
+     * The average score change this attempt had to clear, frozen with its
+     * report: `promote_margin`, lowered toward a quarter of the room left above
+     * what it was compared with near the top, never below what the comparison
+     * can tell from noise ({@link EvaluationMargin.mean_change_needed}). On a
+     * report written before that bar existed, or one that compared nothing, it
+     * is `promote_margin`. Null when the attempt has no report.
+     */
+    mean_change_needed: number | null;
     checkpoint_id: string | null;
     checkpoint_step: number | null;
     attempt_no: number | null;
@@ -3316,6 +3329,15 @@ export interface TrainingRun {
      * cap when no placement price was reported).
      */
     billed_candidate_up_to: boolean;
+    /**
+     * What was taken off for comparison machines that never started (never
+     * reached running): the platform charged nothing for them, so neither does
+     * the attempt -- say "Comparison machine: not charged (it never started)",
+     * or, while a comparison machine that ran is still charged beside it, that
+     * the one that never started was not charged. 0 when none; the cost figures
+     * beside it are already net of it.
+     */
+    comparison_returned_cents: number;
     spent_eval_cents: number;
     train_rows: number | null;
     holdout_rows: number | null;
@@ -3437,6 +3459,8 @@ export interface TrainingRunSummary {
     billed_candidate_cents: number;
     /** See {@link TrainingRun.billed_candidate_up_to}. */
     billed_candidate_up_to: boolean;
+    /** See {@link TrainingRun.comparison_returned_cents}. */
+    comparison_returned_cents: number;
     spent_eval_cents: number;
     /** The fourth money column. A row that leaves it out adds up short. */
     benchmark_spent_cents: number;
@@ -3505,6 +3529,14 @@ export interface EvaluationMargin {
     promote_margin: number;
     promote_min_win_rate: number;
     min_holdout_rows: number;
+    /**
+     * The average score change this comparison actually asked of the new model:
+     * `promote_margin`, lowered toward a quarter of the room left above what it
+     * was compared with once that is smaller, never below what the
+     * comparison's size can tell from noise. Null when nothing was compared,
+     * and on a report written before, whose bar was `promote_margin`.
+     */
+    mean_change_needed: number | null;
 }
 /** The comparison report: the candidate against what serves today, same rows. */
 export interface Evaluation {
@@ -4180,8 +4212,11 @@ export interface TrainingRuleMonthlyLimitRefusal {
     /**
      * The first instant of the next UTC month, when the month's count starts
      * again from nothing. When one attempt alone can cost more than the whole
-     * limit, no month will ever fit: saving the pipeline sets the limit again
-     * from its model's prices, and nothing else changes it; `message` says which.
+     * limit, no month will ever fit, and this is not a day the pipeline trains
+     * again: saving the pipeline sets the limit again from its model's prices,
+     * or -- when its next attempt is planned to need more training time than
+     * the whole limit allows -- a smaller model does, and a save plans the
+     * attempt again; `message` says which.
      */
     resumes_at: string;
 }
@@ -4224,6 +4259,8 @@ export interface PipelineActiveRun {
      * the most it can have cost, so show it as "up to".
      */
     cost_up_to: boolean;
+    /** See {@link TrainingRun.comparison_returned_cents}. */
+    comparison_returned_cents: number;
     /** The version number it will take if it is put live. */
     will_be_version: number;
     /** Always null: nothing in flight is a version yet. */
@@ -4341,8 +4378,12 @@ export interface PipelineVersion {
  * learned from (`rebuild`).
  */
 export type PipelineAttemptTrigger = 'data' | 'schedule' | 'manual' | 'challenger' | 'rebuild';
-/** What became of an attempt. */
-export type PipelineAttemptOutcome = 'became_version' | 'not_better' | 'inconclusive' | 'stopped' | 'failed' | 'running' | 'waiting_for_review';
+/**
+ * What became of an attempt. `inconclusive` is a held-back test that ran and
+ * could not decide; `not_tested` is an attempt that ended before it was tested
+ * (its comparison machine never started, or it expired first).
+ */
+export type PipelineAttemptOutcome = 'became_version' | 'not_better' | 'inconclusive' | 'not_tested' | 'stopped' | 'failed' | 'running' | 'waiting_for_review';
 /** One attempt: every training run a pipeline makes, winner or not. */
 export interface PipelineAttempt {
     attempt: number;
@@ -4371,6 +4412,8 @@ export interface PipelineAttempt {
     cost_cents: number;
     /** See {@link PipelineVersion.cost_up_to}. */
     cost_up_to: boolean;
+    /** See {@link TrainingRun.comparison_returned_cents}. */
+    comparison_returned_cents: number;
     verdict: TrainingVerdict | null;
     win_rate: number | null;
     /**
@@ -4787,6 +4830,13 @@ export interface TrainingRunCost {
     benchmark_cents: number;
     total_cents: number;
     up_to: boolean;
+    /**
+     * What was taken off `comparison_cents` for comparison machines that never
+     * started: "Comparison machine: not charged (it never started)" (with
+     * `comparison_cents` above 0, only the one that never started). 0 when
+     * none.
+     */
+    comparison_returned_cents: number;
 }
 export interface TrainingRunActionResponse {
     run: TrainingRun;
@@ -4857,7 +4907,7 @@ export interface PipelineResponse {
     pipeline: Pipeline;
 }
 /** What became of an attempt, as `listPipelineAttempts` filters it. */
-export type PipelineAttemptResultFilter = 'became_version' | 'not_better' | 'inconclusive' | 'failed' | 'stopped' | 'running' | 'waiting_for_review';
+export type PipelineAttemptResultFilter = 'became_version' | 'not_better' | 'inconclusive' | 'not_tested' | 'failed' | 'stopped' | 'running' | 'waiting_for_review';
 /** Which model's attempts: the pipeline's own (`main`) or its second model's (`second`). */
 export type PipelineAttemptModelFilter = 'main' | 'second';
 export interface PipelineAttemptListParams {

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "runbios-sdk",
-  "version": "0.3.0-rc.325",
+  "version": "0.3.0-rc.330",
   "description": "Official TypeScript SDK for the Run BiOS training and deployment platform API",
   "type": "module",
   "main": "./dist/index.js",