runbios-sdk 0.2.19-dev.274 → 0.2.19-dev.277

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 = "0.2.19-dev.274";
+export declare const VERSION = "0.2.19-dev.277";
 export declare class RunBiOS {
     /** Search models, fetch configs, check adapter compatibility. */
     readonly models: Models;
@@ -75,7 +75,7 @@ export { Training } from './resources/training.js';
 export { Wallet } from './resources/wallet.js';
 export { GPU, type GPURecommendation } from './resources/gpu.js';
 export { Inference, validateChatRequest, parseSSE, buildInferenceRequest, contextSizingBasis, type ContextSizingBasis, CONTEXT_DEFAULT_CEILING, CONTEXT_EDITABLE_FLOOR, type ChatMessage, type FunctionTool, type ChatCompletionParams, type ChatCompletionResponse, type ChatCompletionChunk, type InferenceVerbParams, type CompletionParams, type EmbeddingParams, type RerankParams, type AnthropicMessageParams, type ServerlessUsageWindow, type ServerlessTimeseriesMetric, type ServerlessWorkspaceLimits, type ServerlessUsageEnvelope, type ServerlessUsageOverviewResponse, type ServerlessUsageRowsResponse, type ServerlessUsageTimeseriesResponse, type ServerlessUsageDailyResponse, type ServerlessSavingsTotals, type ServerlessSavingsPeriod, type ServerlessUsageSavingsResponse, } from './resources/inference.js';
-export type { BiOSConfig, PaginatedResponse, ApiErrorBody, AvailableGpuAlternative, CapacityMinimumRequirement, InferenceBookingAccepted, Model, ModelSearchParams, ModelSearchResponse, ModelDetailResponse, InferenceModel, InferenceModelListResponse, ModelConfig, Dataset, DatasetListParams, DatasetListResponse, DatasetUploadParams, DatasetPreview, DatasetPreviewParams, DatasetImportHFParams, DatasetRegisterHFParams, DatasetHubSearchParams, DatasetHubPreviewParams, DatasetValidation, DatasetFormatVariant, DatasetFormatSpec, DatasetFormatSpecs, DatasetStorageUsage, TrainingMethod, RLHFAlgorithm, AdapterType, TrainingJobStatus, TrainingStatusFilter, TrainingCreateParams, TrainingListParams, TrainingListResponse, TrainingJob, TrainingMetrics, TrainingResourceMetrics, TrainingDeviceMetrics, MetricPoint, MetricGraphConfig, TrainingCheckpoint, TrainingLogs, TrainingLogEntry, TrainingStopResponse, TrainingResumeResponse, CanonicalTrainingRequest, TrainingPreflightDataset, TrainingPreflightWarning, TrainingPreflightResponse, TrainingCapabilityChoice, TrainingConfigFieldCapability, TrainingCapabilities, GPUChoice, WalletBalance, Transaction, TransactionListResponse, TransactionListParams, GPUInfo, GPUPricingResponse, GPUOptionsParams, TrainingRecommendParams, TrainingAdvisorVerdict, TrainingAdvisorRecommendation, TrainingAdvisorUserShape, TrainingAdvisorJustification, TrainingAdvisorBasis, GPUOption, GPUOptionSuggestion, GPUOptionsResponse, InferenceStatus, InferenceCreateParams, InferenceUpdateParams, InferenceDeployment, InferenceDeploymentSummary, InferenceCreateResponse, InferenceListResponse, InferenceUpdateResponse, InferenceLifecycleResponse, InferenceDeleteResponse, InferenceGPUOptionMarket, InferenceGPUOption, InferenceGPUAlternative, InferenceGPUOptionsParams, InferenceGPUOptionsResponse, InferenceCanonicalRequest, InferencePreflightWarning, InferencePreflightResponse, AdapterCompatibility, AdapterCompatibilityResponse, AdapterCompatibilityParams, SupportedArchitecture, ArchitectureScope, SupportedArchitecturesResponse, ApiKey, ApiKeyScope, ApiKeyIntrospection, Organization, OrgMember, OrgInvite, Workspace, Integration, IntegrationCreateParams, StorageUsage, StorageObject, PromptTokensDetails, ChatCompletionUsage, ServingKind, TrainingRulePausedReason, TrainingTrigger, TrainingRecipe, TrainingRunState, TrainingVerdict, TrainingDecision, NotifyState, EvaluationStatus, EvaluationItemStatus, EvaluationWinner, EvaluationItemWinnerFilter, EvaluationWarningCode, TrainingToolCall, TrainingMessage, TrainingRuleServing, TrainingRunPromoteRequest, TrainingRunRejectRequest, TrainingRunRollbackRequest, TrainingRunCancelRequest, TrainingRunListParams, TrainingRun, TrainingRunSummary, TrainingRunEvent, TrainingRunLinks, TrainingRunActions, EvaluationRef, EvaluationDecoding, EvaluationJudgeDimension, EvaluationDimensionScore, EvaluationWarning, EvaluationMargin, JudgeAgreementDimension, JudgeAgreement, Evaluation, EvaluationItemSide, EvaluationItem, EvaluationItemListParams, BenchmarkStatus, BenchmarkSourceKind, BenchmarkRunStatus, BenchmarkScoring, BenchmarkDimensionScore, Benchmark, BenchmarkItem, BenchmarkRun, BenchmarkHistoryPoint, BenchmarkSource, BenchmarkCreateParams, BenchmarkListParams, BenchmarkItemListParams, BenchmarkHistoryParams, BenchmarkRetireParams, InferenceAlias, InferenceAliasRequest, TrainingRuleDeleteResponse, TrainingRunListResponse, TrainingRunResponse, TrainingRunActionResponse, EvaluationItemsResponse, InferenceAliasListResponse, InferenceAliasResponse, InferenceAliasDeleteResponse, BenchmarkListResponse, BenchmarkItemsResponse, BenchmarkHistoryResponse, TrainingRuleRunRefusalCode, TrainingRuleSaveRefusalCode, TrainingRuleMonthlyLimitRefusal, PipelineStatus, PipelineActiveRun, PipelineVersion, PipelineAttempt, PipelineAttemptTrigger, PipelineAttemptOutcome, Pipeline, PipelineListParams, PipelineListResponse, PipelineResponse, PipelineTrainWhen, PipelineSchedule, PipelineWriteRequest, PipelineMutationResponse, PipelineImportParams, PipelineFloors, LoopModel, LoopModelsResponse, LoopTracePipeline, LoopTracePipelinesResponse, LoopMetricsHistoryParams, LoopMetricPoint, LoopSignalDay, LoopMetricsHistory, PromotionPolicyState, RuleBenchmark, PipelineBenchmarksResponse, PipelineBenchmarkAttachParams, PipelineBenchmarkMoveParams, CompareBenchmark, VersionBenchmarkScore, PromotionMeasurement, PromotionOutcome, VersionScores, VersionScoreDifference, VersionCompareResponse, VersionCompareParams, } from './types.js';
+export type { BiOSConfig, PaginatedResponse, ApiErrorBody, AvailableGpuAlternative, CapacityMinimumRequirement, InferenceBookingAccepted, Model, ModelSearchParams, ModelSearchResponse, ModelDetailResponse, InferenceModel, InferenceModelListResponse, ModelConfig, Dataset, DatasetListParams, DatasetListResponse, DatasetUploadParams, DatasetPreview, DatasetPreviewParams, DatasetImportHFParams, DatasetRegisterHFParams, DatasetHubSearchParams, DatasetHubPreviewParams, DatasetValidation, DatasetFormatVariant, DatasetFormatSpec, DatasetFormatSpecs, DatasetStorageUsage, TrainingMethod, RLHFAlgorithm, AdapterType, TrainingJobStatus, TrainingStatusFilter, TrainingCreateParams, TrainingListParams, TrainingListResponse, TrainingJob, TrainingMetrics, TrainingResourceMetrics, TrainingDeviceMetrics, MetricPoint, MetricGraphConfig, TrainingCheckpoint, TrainingLogs, TrainingLogEntry, TrainingStopResponse, TrainingResumeResponse, CanonicalTrainingRequest, TrainingPreflightDataset, TrainingPreflightWarning, TrainingPreflightResponse, TrainingCapabilityChoice, TrainingConfigFieldCapability, TrainingCapabilities, GPUChoice, WalletBalance, Transaction, TransactionListResponse, TransactionListParams, GPUInfo, GPUPricingResponse, GPUOptionsParams, TrainingRecommendParams, TrainingAdvisorVerdict, TrainingAdvisorRecommendation, TrainingAdvisorUserShape, TrainingAdvisorJustification, TrainingAdvisorBasis, GPUOption, GPUOptionSuggestion, GPUOptionsResponse, InferenceStatus, InferenceCreateParams, InferenceUpdateParams, InferenceDeployment, InferenceDeploymentSummary, InferenceCreateResponse, InferenceListResponse, InferenceUpdateResponse, InferenceLifecycleResponse, InferenceDeleteResponse, InferenceGPUOptionMarket, InferenceGPUOption, InferenceGPUAlternative, InferenceGPUOptionsParams, InferenceGPUOptionsResponse, InferenceCanonicalRequest, InferencePreflightWarning, InferencePreflightResponse, AdapterCompatibility, AdapterCompatibilityResponse, AdapterCompatibilityParams, SupportedArchitecture, ArchitectureScope, SupportedArchitecturesResponse, ApiKey, ApiKeyScope, ApiKeyIntrospection, Organization, OrgMember, OrgInvite, Workspace, Integration, IntegrationCreateParams, StorageUsage, StorageObject, PromptTokensDetails, ChatCompletionUsage, ServingKind, TrainingRulePausedReason, TrainingTrigger, TrainingRecipe, TrainingRunState, TrainingVerdict, TrainingDecision, NotifyState, EvaluationStatus, EvaluationItemStatus, EvaluationWinner, EvaluationItemWinnerFilter, EvaluationWarningCode, TrainingToolCall, TrainingMessage, TrainingRuleServing, TrainingRunPromoteRequest, TrainingRunRejectRequest, TrainingRunRollbackRequest, TrainingRunCancelRequest, TrainingRunListParams, TrainingRun, TrainingRunSummary, TrainingRunEvent, TrainingRunLinks, TrainingRunActions, EvaluationRef, EvaluationDecoding, EvaluationJudgeDimension, EvaluationDimensionScore, EvaluationWarning, EvaluationMargin, Evaluation, EvaluationItemSide, EvaluationItem, EvaluationItemListParams, BenchmarkStatus, BenchmarkSourceKind, BenchmarkRunStatus, BenchmarkScoring, BenchmarkDimensionScore, Benchmark, BenchmarkItem, BenchmarkRun, BenchmarkHistoryPoint, BenchmarkSource, BenchmarkCreateParams, BenchmarkListParams, BenchmarkItemListParams, BenchmarkHistoryParams, BenchmarkRetireParams, InferenceAlias, InferenceAliasRequest, TrainingRuleDeleteResponse, TrainingRunListResponse, TrainingRunResponse, TrainingRunActionResponse, EvaluationItemsResponse, InferenceAliasListResponse, InferenceAliasResponse, InferenceAliasDeleteResponse, BenchmarkListResponse, BenchmarkItemsResponse, BenchmarkHistoryResponse, TrainingRuleRunRefusalCode, TrainingRuleSaveRefusalCode, TrainingRuleMonthlyLimitRefusal, PipelineStatus, PipelineActiveRun, PipelineVersion, PipelineAttempt, PipelineAttemptTrigger, PipelineAttemptOutcome, Pipeline, PipelineListParams, PipelineListResponse, PipelineResponse, PipelineTrainWhen, PipelineSchedule, PipelineWriteRequest, PipelineMutationResponse, PipelineImportParams, PipelineFloors, LoopModel, LoopModelsResponse, LoopTracePipeline, LoopTracePipelinesResponse, LoopMetricsHistoryParams, LoopMetricPoint, LoopSignalDay, LoopMetricsHistory, PromotionPolicyState, RuleBenchmark, PipelineBenchmarksResponse, PipelineBenchmarkAttachParams, PipelineBenchmarkMoveParams, CompareBenchmark, VersionBenchmarkScore, PromotionMeasurement, PromotionOutcome, VersionScores, VersionScoreDifference, VersionCompareResponse, VersionCompareParams, } from './types.js';
 /** The closed set of run states a training run never leaves. */
 export { TERMINAL_RUN_STATES } from './types.js';
 /** The most versions a pipeline may be set to make. */

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 = '0.2.19-dev.274';
+export const VERSION = '0.2.19-dev.277';
 export class RunBiOS {
     /** Search models, fetch configs, check adapter compatibility. */
     models;

package/dist/resources/loop.d.ts

@@ -88,10 +88,10 @@ export declare class Loop {
      * a corpus from a pile.
      *
      * Each row is read for what it actually is. A `prompt`/`chosen`/`rejected`
-     * triple becomes a preference pair; an answer plus a yes-or-no becomes a
-     * thumbs verdict; a question and an answer waits for review; a question with
-     * no answer waits for an answer; a paragraph of prose is refused, because it
-     * is not a conversation.
+     * triple becomes a correction (`chosen` is learned from); an answer plus a
+     * yes-or-no becomes a thumbs verdict; a question and an answer waits for
+     * review; a question with no answer waits for an answer; a paragraph of prose
+     * is refused, because it is not a conversation.
      *
      * A verdict that arrives *with* the file is kept -- discarding somebody's
      * judgement would be worse -- but it is recorded as having come from your
@@ -138,7 +138,14 @@ export declare class Loop {
     importRows(params: LoopImportParams): Promise<LoopImportResult>;
     /** List captured conversations, newest first. */
     listTraces(params?: LoopTraceListParams): Promise<LoopTraceListResponse>;
-    /** Read one conversation, with every verdict recorded on it. */
+    /**
+     * Read one conversation, with every verdict and tag recorded on it.
+     *
+     * A conversation kept in your own storage carries `payload_stored_externally`.
+     * When its text could not be read just now it also carries
+     * `payload_unavailable`, with `messages` null and `completion` empty: it is
+     * not an empty conversation. Read it again later rather than reviewing it.
+     */
     getTrace(id: string): Promise<LoopTrace>;
     /**
      * Delete a conversation.
@@ -158,11 +165,11 @@ export declare class Loop {
      * Append-only: posting a second verdict does not replace the first. Two
      * reviewers disagreeing about an answer is information worth keeping.
      *
-     * `source` says who judged: `human`, `verifier` (a deterministic check --
-     * tests passed, schema valid), `judge` (a model grading a model), or
-     * `behavioural` (what the user did next). When several disagree, a human
-     * outranks a verifier, a verifier outranks a judge, and a judge outranks a
-     * behavioural hint.
+     * What a pipeline trains on (every pipeline fine-tunes on examples): an
+     * `accepted` answer is an example as it stands; the `correction` of an
+     * `edited` or `gold` verdict is the example, in place of the model's answer;
+     * a `rejected` answer is left out. Nothing is learned from an answer that was
+     * rejected or replaced. `source` defaults to `human`.
      */
     signal(traceId: string, params: LoopSignalParams): Promise<LoopSignal>;
     /**
@@ -182,23 +189,18 @@ export declare class Loop {
     /**
      * The model was wrong; here is what it should have said.
      *
-     * The most valuable feedback there is, because it produces BOTH halves of a
-     * preference pair from one action: the model learns your answer and learns to
-     * avoid its own.
+     * The most valuable feedback there is: your answer is trained on in place of
+     * the model's. Nothing is learned from the answer it replaced.
      */
     correct(traceId: string, answer: string, reason?: string): Promise<LoopSignal>;
     /**
      * Record the GOLD answer for this question — the reference, regardless of
      * what the model happened to say.
      *
-     * Different from `correct()` in a way that matters: a correction asserts the
-     * model was wrong, so the original becomes the rejected half of a pair. A gold
-     * answer asserts nothing about the model, so if it MATCHES what was said, no
-     * preference pair is invented — but SFT still learns it.
-     *
-     * It is the most reusable thing you can record: it trains SFT, forms a DPO
-     * pair when it differs from the answer, and stands in as the GRPO reference
-     * when you have not supplied a separate ground truth.
+     * A pipeline trains on it exactly as it does on a correction: the gold
+     * answer is the example, in place of the model's. The difference is only
+     * what it says: a correction says the model was wrong, a gold answer says
+     * nothing about the model.
      */
     gold(traceId: string, answer: string, reason?: string): Promise<LoopSignal>;
     /** Every verdict on one conversation, oldest first. */
@@ -356,8 +358,8 @@ export declare class Loop {
      *
      * `status` says whether it is the platform working (`running`), the member
      * who has to act (`needs_review`, `needs_funds`), or nothing at all until
-     * somebody does (`paused`, with `paused_reason` or `needs_consent` saying
-     * why; `complete` at `max_versions`). `month_spent_cents` is the figure the
+     * somebody does (`paused`, with `paused_reason` saying why; `complete` at
+     * `max_versions`). `month_spent_cents` is the figure the
      * monthly limit is enforced against: what runs started this month have cost
      * at most, not an exact spend.
      */

package/dist/resources/loop.js

@@ -91,10 +91,10 @@ export class Loop {
      * a corpus from a pile.
      *
      * Each row is read for what it actually is. A `prompt`/`chosen`/`rejected`
-     * triple becomes a preference pair; an answer plus a yes-or-no becomes a
-     * thumbs verdict; a question and an answer waits for review; a question with
-     * no answer waits for an answer; a paragraph of prose is refused, because it
-     * is not a conversation.
+     * triple becomes a correction (`chosen` is learned from); an answer plus a
+     * yes-or-no becomes a thumbs verdict; a question and an answer waits for
+     * review; a question with no answer waits for an answer; a paragraph of prose
+     * is refused, because it is not a conversation.
      *
      * A verdict that arrives *with* the file is kept -- discarding somebody's
      * judgement would be worse -- but it is recorded as having come from your
@@ -182,7 +182,14 @@ export class Loop {
         const qs = q.toString();
         return this._http.fetchGet(`/api/loop/traces${qs ? `?${qs}` : ''}`);
     }
-    /** Read one conversation, with every verdict recorded on it. */
+    /**
+     * Read one conversation, with every verdict and tag recorded on it.
+     *
+     * A conversation kept in your own storage carries `payload_stored_externally`.
+     * When its text could not be read just now it also carries
+     * `payload_unavailable`, with `messages` null and `completion` empty: it is
+     * not an empty conversation. Read it again later rather than reviewing it.
+     */
     async getTrace(id) {
         const res = await this._http.fetchGet(`/api/loop/traces/${encodeURIComponent(id)}`);
         return res.trace;
@@ -205,11 +212,11 @@ export class Loop {
      * Append-only: posting a second verdict does not replace the first. Two
      * reviewers disagreeing about an answer is information worth keeping.
      *
-     * `source` says who judged: `human`, `verifier` (a deterministic check --
-     * tests passed, schema valid), `judge` (a model grading a model), or
-     * `behavioural` (what the user did next). When several disagree, a human
-     * outranks a verifier, a verifier outranks a judge, and a judge outranks a
-     * behavioural hint.
+     * What a pipeline trains on (every pipeline fine-tunes on examples): an
+     * `accepted` answer is an example as it stands; the `correction` of an
+     * `edited` or `gold` verdict is the example, in place of the model's answer;
+     * a `rejected` answer is left out. Nothing is learned from an answer that was
+     * rejected or replaced. `source` defaults to `human`.
      */
     async signal(traceId, params) {
         const res = await this._http.fetchPost(`/api/loop/traces/${encodeURIComponent(traceId)}/signals`, params);
@@ -236,9 +243,8 @@ export class Loop {
     /**
      * The model was wrong; here is what it should have said.
      *
-     * The most valuable feedback there is, because it produces BOTH halves of a
-     * preference pair from one action: the model learns your answer and learns to
-     * avoid its own.
+     * The most valuable feedback there is: your answer is trained on in place of
+     * the model's. Nothing is learned from the answer it replaced.
      */
     async correct(traceId, answer, reason) {
         return this.signal(traceId, { verdict: 'edited', correction: answer, reason });
@@ -247,14 +253,10 @@ export class Loop {
      * Record the GOLD answer for this question — the reference, regardless of
      * what the model happened to say.
      *
-     * Different from `correct()` in a way that matters: a correction asserts the
-     * model was wrong, so the original becomes the rejected half of a pair. A gold
-     * answer asserts nothing about the model, so if it MATCHES what was said, no
-     * preference pair is invented — but SFT still learns it.
-     *
-     * It is the most reusable thing you can record: it trains SFT, forms a DPO
-     * pair when it differs from the answer, and stands in as the GRPO reference
-     * when you have not supplied a separate ground truth.
+     * A pipeline trains on it exactly as it does on a correction: the gold
+     * answer is the example, in place of the model's. The difference is only
+     * what it says: a correction says the model was wrong, a gold answer says
+     * nothing about the model.
      */
     async gold(traceId, answer, reason) {
         return this.signal(traceId, { verdict: 'gold', correction: answer, reason });
@@ -491,8 +493,8 @@ export class Loop {
      *
      * `status` says whether it is the platform working (`running`), the member
      * who has to act (`needs_review`, `needs_funds`), or nothing at all until
-     * somebody does (`paused`, with `paused_reason` or `needs_consent` saying
-     * why; `complete` at `max_versions`). `month_spent_cents` is the figure the
+     * somebody does (`paused`, with `paused_reason` saying why; `complete` at
+     * `max_versions`). `month_spent_cents` is the figure the
      * monthly limit is enforced against: what runs started this month have cost
      * at most, not an exact spend.
      */

package/dist/types.d.ts

@@ -2297,18 +2297,17 @@ export interface ChatCompletionUsage {
     total_tokens: number;
     prompt_tokens_details?: PromptTokensDetails;
 }
-/** What a training set is shaped for. The three need different things. */
-export type LoopMethod = 'sft' | 'dpo' | 'grpo' | 'kto';
-/** Who judged an answer. */
+/** Who gave a verdict. Feedback you send is `human` unless you say otherwise. */
 export type LoopSource = 'human' | 'verifier' | 'judge' | 'behavioural';
-/** The judgement itself. */
 /**
  * The judgement recorded on an answer.
  *
- * `accepted` / `rejected` are the thumbs (see `upvote` / `downvote`).
- * `edited` says the model was WRONG and carries the better answer.
+ * `accepted` / `rejected` are the thumbs (see `upvote` / `downvote`): an
+ * accepted answer is trained on, a rejected one is left out.
+ * `edited` says the model was WRONG and carries the better answer, which is
+ * trained on in place of the model's.
  * `gold` records the reference answer for the question, making no claim about
- * whether the model was right — which is why it is separate from `edited`.
+ * whether the model was right; it is trained on the same way.
  */
 export type LoopVerdict = 'accepted' | 'rejected' | 'edited' | 'scored' | 'gold';
 /** A function an assistant turn asked to invoke. `arguments` is JSON-encoded. */
@@ -2495,7 +2494,7 @@ export interface LoopImportResult {
     not_saved_rows?: number[];
     /**
      * Rows whose CONVERSATION stored and whose VERDICT did not, and which rows
-     * those were. A preference pair or a thumbs label is two writes, and the
+     * those were. A chosen/rejected row or a thumbs label is two writes, and the
      * second can fail on its own: the conversation is then in the loop carrying
      * nobody's judgement, counted under `needs_review` rather than `reviewed`,
      * and not trainable. The call rejects with {@link LoopImportIncompleteError},
@@ -2532,6 +2531,8 @@ export interface LoopSignal {
     ground_truth: string | null;
     reason: string | null;
     author: string | null;
+    /** What the caller attached when recording it, or null. */
+    metadata: Record<string, unknown> | null;
     created_at: string;
 }
 export interface LoopSignalParams {
@@ -2539,9 +2540,9 @@ export interface LoopSignalParams {
     source?: LoopSource;
     /** Required when verdict is `edited`. The answer the model should have given. */
     correction?: string;
-    /** Required when verdict is `scored`. */
+    /** Required when verdict is `scored`: above zero trains the answer as it stands, otherwise it is left out. */
     score?: number;
-    /** A value or fact the answer can be checked against. GRPO needs one. */
+    /** A value or fact the answer can be checked against. Kept with the verdict; a pipeline trains on `correction`, not on this. */
     ground_truth?: string;
     reason?: string;
     author?: string;
@@ -2579,7 +2580,11 @@ export interface LoopTrace {
     deployment_id: string | null;
     model: string;
     model_version: string | null;
-    messages: LoopMessage[];
+    /**
+     * The conversation. NULL, with `completion` empty, when its text is kept in
+     * your own storage and could not be read just now: see `payload_unavailable`.
+     */
+    messages: LoopMessage[] | null;
     completion: string;
     tool_calls?: LoopToolCall[];
     tools?: unknown;
@@ -2591,9 +2596,35 @@ export interface LoopTrace {
     redacted_at: string | null;
     /** What the redaction pass removed, counted by rule. */
     redaction_report?: Record<string, number>;
+    /** What the caller attached when recording it, or null. */
+    metadata: Record<string, unknown> | null;
     created_at: string;
     expires_at: string;
+    /** Every verdict recorded on it, oldest first. On the single read only. */
     signals?: LoopSignal[];
+    /** How many verdicts it carries. */
+    signal_count: number;
+    /**
+     * The verdict that decides it -- the one training uses, not whichever
+     * arrived last. Absent when nobody has given feedback.
+     */
+    verdict?: LoopVerdict;
+    /** Who gave that verdict. */
+    verdict_source?: LoopSource;
+    /** Its tags. On the single read only. */
+    labels?: LoopLabel[];
+    /**
+     * True when the conversation's text is kept in your own storage rather than
+     * here. The text is read from there on every request.
+     */
+    payload_stored_externally?: boolean;
+    /**
+     * True when that text could not be read just now (the storage could not be
+     * reached, or the object is gone). `messages` is then null and `completion`
+     * empty: the conversation is NOT empty. Try again later; do not review it or
+     * treat it as a blank answer.
+     */
+    payload_unavailable?: boolean;
 }
 export interface LoopTraceListParams {
     deployment_id?: string;
@@ -2810,11 +2841,13 @@ export type TrainingRulePausedReason = 'consent_invalid' | 'insufficient_funds'
  */
  | 'training_not_enabled';
 /**
- * Why a run fired. `variant` is a recipe variant: the pipeline had nothing new
- * to learn, so it trained the same base model on the same conversations with
- * one training setting changed (see {@link TrainingRecipe}).
+ * Why a run fired, as the run stores it: `rows` (enough new samples),
+ * `cadence` (the schedule), `both`, `manual` (train now), `challenger` (the
+ * pipeline's challenger model trained on the same data as the attempt before
+ * it), or `variant` (a recipe variant, retired: the same base model on the
+ * same conversations with one training setting changed; old runs only).
  */
-export type TrainingTrigger = 'rows' | 'cadence' | 'both' | 'manual' | 'variant';
+export type TrainingTrigger = 'rows' | 'cadence' | 'both' | 'manual' | 'variant' | 'challenger';
 /**
  * The four training settings a recipe variant may change, as the run was
  * actually trained: the training service's own defaults are filled in where
@@ -2891,8 +2924,6 @@ export interface TrainingRun {
     workspace_id: string;
     seq: number;
     rule_revision: number;
-    authorized_by: string;
-    rule_snapshot: Record<string, unknown>;
     trigger: TrainingTrigger;
     /**
      * For a recipe variant, what it changed, in plain words ("half the learning
@@ -2921,13 +2952,9 @@ export interface TrainingRun {
      * no remedy worth printing.
      */
     error_next_step: string | null;
-    training_ceiling_cents: number;
-    candidate_ceiling_cents: number;
-    eval_ceiling_cents: number;
     billed_training_cents: number;
     billed_candidate_cents: number;
     spent_eval_cents: number;
-    loop_dataset_id: string | null;
     train_rows: number | null;
     holdout_rows: number | null;
     train_dataset_id: string | null;
@@ -2967,8 +2994,7 @@ export interface TrainingRun {
     benchmark_run_id: string | null;
     /**
      * What the replay's calls cost. Kept apart from `spent_eval_cents` so a
-     * report can say what the comparison cost and what the benchmark cost; both
-     * come out of `eval_ceiling_cents` and their sum can never exceed it, so
+     * report can say what the comparison cost and what the benchmark cost, so
      * anything totalling what a run cost has to add this one too.
      */
     benchmark_spent_cents: number;
@@ -3049,7 +3075,6 @@ export interface TrainingRunEvent {
 export interface TrainingRunLinks {
     training_job_url: string | null;
     candidate_url: string | null;
-    dataset_url: string | null;
     evaluation_url: string | null;
     /** The TREND the benchmark number belongs to, not the one replay. */
     benchmark_url: string | null;
@@ -3091,41 +3116,13 @@ export interface EvaluationWarning {
 export interface EvaluationMargin {
     promote_margin: number;
     promote_min_win_rate: number;
-    min_judge_agreement: number;
     min_holdout_rows: number;
 }
-/**
- * One rubric dimension of a judge-versus-human agreement.
- *
- * Deliberately not EvaluationDimensionScore: that type carries `incumbent` and
- * `candidate`, which are the two models being compared, and an agreement has
- * neither. `pairs` is per dimension because a reviewer who rated one dimension
- * and skipped another leaves a different denominator behind each number.
- */
-export interface JudgeAgreementDimension {
-    dimension: string;
-    pairs: number;
-    agreement: number | null;
-    mean_abs_error: number | null;
-}
-/** How often this judge agreed with the workspace's own reviewers. */
-export interface JudgeAgreement {
-    judge_id: string;
-    pairs: number;
-    agreement: number | null;
-    mean_abs_error: number | null;
-    per_dimension: JudgeAgreementDimension[];
-    window_days: number;
-    computed_at: string;
-    /** "Not enough reviewer overlap yet" is an answer; 100% of two is not. */
-    enough_pairs: boolean;
-}
 /** The comparison report: the candidate against what serves today, same rows. */
 export interface Evaluation {
     id: string;
     run_id: string;
     workspace_id: string;
-    loop_dataset_id: string | null;
     split: string;
     incumbent_ref: EvaluationRef;
     candidate_ref: EvaluationRef;
@@ -3134,7 +3131,6 @@ export interface Evaluation {
     judge_model: string | null;
     judge_instructions: string | null;
     judge_dimensions: EvaluationJudgeDimension[];
-    judge_system_prompt: string | null;
     decoding: EvaluationDecoding;
     rows_selected: number;
     rows_scored: number;
@@ -3150,7 +3146,6 @@ export interface Evaluation {
     judge_incumbent_mean: number | null;
     judge_candidate_mean: number | null;
     per_dimension: EvaluationDimensionScore[];
-    judge_agreement: JudgeAgreement | null;
     /** Informational: there is no incumbent counterpart to compare it against. */
     trainer_eval_loss: number | null;
     warnings: EvaluationWarning[];
@@ -3283,14 +3278,11 @@ export interface Benchmark {
     judge_model: string;
     judge_instructions: string;
     judge_dimensions: EvaluationJudgeDimension[];
-    judge_system_prompt: string | null;
     decoding: EvaluationDecoding;
     scoring: BenchmarkScoring;
     item_count: number;
     /** Recomputed from the rows and compared at the start of every replay. */
     items_digest: string;
-    /** A cap INSIDE the run's own amount for judge calls, never an addition to it. */
-    per_run_ceiling_cents: number;
     created_by: string;
     created_at: string;
     retired_at: string | null;
@@ -3342,7 +3334,6 @@ export interface BenchmarkRun {
     prompt_tokens: number;
     completion_tokens: number;
     spent_cents: number;
-    ceiling_cents: number;
     status: BenchmarkRunStatus;
     /** The whole explanation when there is no score, so never empty on one. */
     status_reason: string | null;
@@ -3448,6 +3439,10 @@ export interface BenchmarkRetireParams {
  *   monthly cap, so what this attempt trained could not be scored. The body
  *   carries `resumes_at`, when the cap lifts (the next UTC month, or sooner
  *   if the key's cap is raised).
+ * - `LIVE_VERSION_STOPPED` (409): the deployment the pipeline's live version
+ *   answers from is stopped, paused until funds are added, deleted or failed,
+ *   so an attempt could not be compared against it. Resume it in Deployments,
+ *   or roll the version back, then train again.
  * - `AGENT_COULD_NOT_START` (503): the key every comparison is scored with is
  *   missing or was refused, and train now could not renew it just then.
  *   Nothing was started. Press train now again in a minute.
@@ -3460,7 +3455,7 @@ export interface BenchmarkRetireParams {
  * `MONTHLY_LIMIT_REACHED` and `SCORING_CAPPED` once their date has passed;
  * retrying any other one unchanged gets the same answer.
  */
-export type TrainingRuleRunRefusalCode = 'NOT_ENOUGH_SAMPLES' | 'PIPELINE_NEEDS_SAVE' | 'RULE_PAUSED' | 'RUN_ACTIVE' | 'VERSION_LIMIT_REACHED' | 'MONTHLY_LIMIT_REACHED' | 'SCORING_CAPPED' | 'AGENT_COULD_NOT_START' | 'AGENT_OFF';
+export type TrainingRuleRunRefusalCode = 'NOT_ENOUGH_SAMPLES' | 'PIPELINE_NEEDS_SAVE' | 'RULE_PAUSED' | 'RUN_ACTIVE' | 'VERSION_LIMIT_REACHED' | 'MONTHLY_LIMIT_REACHED' | 'SCORING_CAPPED' | 'LIVE_VERSION_STOPPED' | 'AGENT_COULD_NOT_START' | 'AGENT_OFF';
 /**
  * The machine codes saving a pipeline (`createPipeline` or `updatePipeline`)
  * can be refused with: every one the two routes answer beyond a malformed
@@ -3537,7 +3532,8 @@ export interface TrainingRuleMonthlyLimitRefusal {
     /**
      * The first instant of the next UTC month, when the counted spend starts
      * again from nothing. When `run_max_cents` alone exceeds the limit, no month
-     * will ever fit and only raising the limit helps; `message` says which.
+     * will ever fit: saving the pipeline sets the limit again from its model's
+     * prices, and nothing else changes it; `message` says which.
      */
     resumes_at: string;
 }
@@ -3549,8 +3545,9 @@ export declare const PIPELINE_MAX_VERSIONS_CEILING = 100;
  * platform working on it.
  *
  * `paused` is a rule that is switched off, AND an enabled rule the platform
- * has stopped firing (`paused_reason`) or whose terms changed and were never
- * accepted (`needs_consent`). `complete` is a pipeline at its `max_versions`.
+ * has stopped firing (`paused_reason`; `consent_invalid` when its settings
+ * changed since it was last saved). `complete` is a pipeline at its
+ * `max_versions`.
  */
 export type PipelineStatus = 'running' | 'needs_review' | 'needs_funds' | 'waiting' | 'paused' | 'complete';
 /**
@@ -3588,8 +3585,6 @@ export interface PipelineActiveRun {
      * not a version, and becomes one only if a member promotes it.
      */
     competed: boolean;
-    /** What recipe it is trying, when it is a recipe variant. See TrainingRun.recipe_note. */
-    recipe_note: string | null;
 }
 /**
  * One version: an attempt that was put live -- promoted by the pipeline's
@@ -3659,18 +3654,10 @@ export interface PipelineVersion {
      */
     compared_at: string | null;
     /**
-     * Why the run that made it fired. `variant` is a version trained on the same
-     * conversations as the one before it with one setting changed; every other
-     * value is a version that learned from newly reviewed conversations (or a
-     * member's run now).
+     * Why the attempt that made it was made, in the words `attempts[]` uses for
+     * the same run; see {@link PipelineAttempt.trigger}.
      */
-    trigger: TrainingTrigger;
-    /** See TrainingRun.recipe_note. */
-    recipe_note: string | null;
-    /** See TrainingRun.recipe_of_version. */
-    recipe_of_version: number | null;
-    /** See TrainingRun.recipe. */
-    recipe: TrainingRecipe;
+    trigger: PipelineAttemptTrigger;
     finished_at: string | null;
 }
 /**
@@ -3742,8 +3729,6 @@ export interface PipelineTrainWhen {
 }
 /** What is held back beyond max(50, 5% of the data): a larger percent, or a count. */
 export interface PipelineHoldout {
-    percent: number;
-    count: number | null;
     min_rows: number;
 }
 /**
@@ -3780,10 +3765,6 @@ export interface Pipeline {
      * commit the server pinned.
      */
     challenger_model: PipelineModelRef | null;
-    /** `sft` today; `kto`, `dpo` and `grpo` cannot be chosen yet. */
-    training_type: string;
-    /** How each version starts: `fresh`, a new adapter over the pinned base trained on all accepted data. */
-    version_base: string;
     train_when: PipelineTrainWhen;
     promotion: PipelinePromotion;
     holdout: PipelineHoldout;
@@ -3793,7 +3774,6 @@ export interface Pipeline {
     /** The pinned base every version is a fresh adapter over, which is what makes their scores comparable. */
     base_model_id: string;
     base_model_revision: string;
-    train_type: string;
     /** The version serving now; null when what serves is not a version of this pipeline. */
     live_version: number | null;
     /** The same number under its older name. */
@@ -3808,25 +3788,31 @@ export interface Pipeline {
     serving_taken_over_by: string | null;
     benchmark_id: string | null;
     benchmark_name: string | null;
-    benchmark_decides: boolean;
+    /** The least a benchmark's score has to rise by to count as an improvement under a policy that reads benchmarks. */
     benchmark_min_delta: number;
     status: PipelineStatus;
     /**
      * The rule's own sentence about what it is waiting for or why it stopped,
      * verbatim. Only as fresh as the platform's last visit: when `status` is
-     * `paused`, `paused_reason` and `needs_consent` are what say why.
+     * `paused`, `paused_reason` is what says why.
      */
     next_reason: string | null;
     last_checked_at: string | null;
     next_due_at: string | null;
     auto_promote: boolean;
-    /** Why the platform stopped firing this rule, or null. Same codes as the rule's. */
+    /**
+     * Why the platform stopped firing this rule, or null. `consent_invalid` is
+     * a pipeline whose settings changed since it was last saved: saving it
+     * again starts it.
+     */
     paused_reason: TrainingRulePausedReason | null;
     /**
-     * The rule's terms changed and nobody has accepted them. It makes nothing
-     * until somebody does.
+     * When this workspace's scoring key comes off its monthly cap (the start of
+     * the next UTC month), or null when it is not capped. Until then Train now
+     * answers 409 `SCORING_CAPPED` with this time as `resumes_at`, and nothing
+     * trains on its own: an attempt would train and could not be compared.
      */
-    needs_consent: boolean;
+    scoring_capped_until: string | null;
     /** The rule's monthly limit. Null means it has none. */
     monthly_ceiling_cents: number | null;
     /**

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "runbios-sdk",
-  "version": "0.2.19-dev.274",
+  "version": "0.2.19-dev.277",
   "description": "Official TypeScript SDK for the Run BiOS training and deployment platform API",
   "type": "module",
   "main": "./dist/index.js",