@roarkanalytics/sdk 0.332.0 → 2.0.0

package/src/resources/evaluation.ts

@@ -18,79 +18,1015 @@ export class Evaluation extends APIResource {
   /**
    * Returns a specific evaluator with its blocks and configuration.
    */
-  getEvaluatorById(evaluatorId: string, options?: Core.RequestOptions): Core.APIPromise<unknown> {
+  getEvaluatorById(
+    evaluatorId: string,
+    options?: Core.RequestOptions,
+  ): Core.APIPromise<EvaluationGetEvaluatorByIDResponse> {
     return this._client.get(`/v1/evaluation/evaluators/${evaluatorId}`, options);
   }
 
-  /**
-   * Returns a list of evaluators with their blocks and configuration for the
-   * authenticated project.
-   */
-  getEvaluators(
-    query?: EvaluationGetEvaluatorsParams,
-    options?: Core.RequestOptions,
-  ): Core.APIPromise<unknown>;
-  getEvaluators(options?: Core.RequestOptions): Core.APIPromise<unknown>;
-  getEvaluators(
-    query: EvaluationGetEvaluatorsParams | Core.RequestOptions = {},
-    options?: Core.RequestOptions,
-  ): Core.APIPromise<unknown> {
-    if (isRequestOptions(query)) {
-      return this.getEvaluators({}, query);
+  /**
+   * Returns a list of evaluators with their blocks and configuration for the
+   * authenticated project.
+   */
+  getEvaluators(
+    query?: EvaluationGetEvaluatorsParams,
+    options?: Core.RequestOptions,
+  ): Core.APIPromise<EvaluationGetEvaluatorsResponse>;
+  getEvaluators(options?: Core.RequestOptions): Core.APIPromise<EvaluationGetEvaluatorsResponse>;
+  getEvaluators(
+    query: EvaluationGetEvaluatorsParams | Core.RequestOptions = {},
+    options?: Core.RequestOptions,
+  ): Core.APIPromise<EvaluationGetEvaluatorsResponse> {
+    if (isRequestOptions(query)) {
+      return this.getEvaluators({}, query);
+    }
+    return this._client.get('/v1/evaluation/evaluators', { query, ...options });
+  }
+
+  /**
+   * Retrieve details of a specific evaluation job
+   */
+  getJob(jobId: string, options?: Core.RequestOptions): Core.APIPromise<EvaluationGetJobResponse> {
+    return this._client.get(`/v1/evaluation/job/${jobId}`, options);
+  }
+
+  /**
+   * Retrieve paginated details of a specific evaluation job runs
+   */
+  getJobRuns(
+    jobId: string,
+    query?: EvaluationGetJobRunsParams,
+    options?: Core.RequestOptions,
+  ): Core.APIPromise<EvaluationGetJobRunsResponse>;
+  getJobRuns(jobId: string, options?: Core.RequestOptions): Core.APIPromise<EvaluationGetJobRunsResponse>;
+  getJobRuns(
+    jobId: string,
+    query: EvaluationGetJobRunsParams | Core.RequestOptions = {},
+    options?: Core.RequestOptions,
+  ): Core.APIPromise<EvaluationGetJobRunsResponse> {
+    if (isRequestOptions(query)) {
+      return this.getJobRuns(jobId, {}, query);
+    }
+    return this._client.get(`/v1/evaluation/job/${jobId}/runs`, { query, ...options });
+  }
+}
+
+export interface EvaluationCreateJobResponse {
+  data: EvaluationCreateJobResponse.Data;
+}
+
+export namespace EvaluationCreateJobResponse {
+  export interface Data {
+    /**
+     * ID of the evaluation job
+     */
+    jobId: string;
+
+    /**
+     * Status of the evaluation job
+     */
+    status: 'PENDING' | 'PROCESSING' | 'SUCCESS' | 'FAILURE';
+  }
+}
+
+/**
+ * Evaluator with its configured blocks
+ */
+export interface EvaluationGetEvaluatorByIDResponse {
+  /**
+   * Unique identifier for the evaluator
+   */
+  id: string;
+
+  /**
+   * Array of evaluation blocks configured for this evaluator
+   */
+  blocks: Array<
+    | EvaluationGetEvaluatorByIDResponse.UnionMember0
+    | EvaluationGetEvaluatorByIDResponse.UnionMember1
+    | EvaluationGetEvaluatorByIDResponse.UnionMember2
+    | EvaluationGetEvaluatorByIDResponse.UnionMember3
+    | EvaluationGetEvaluatorByIDResponse.UnionMember4
+    | EvaluationGetEvaluatorByIDResponse.UnionMember5
+    | EvaluationGetEvaluatorByIDResponse.UnionMember6
+    | EvaluationGetEvaluatorByIDResponse.UnionMember7
+    | EvaluationGetEvaluatorByIDResponse.UnionMember8
+  >;
+
+  /**
+   * ISO timestamp when the evaluator was created
+   */
+  createdAt: string;
+
+  /**
+   * Optional description of the evaluator
+   */
+  description: string | null;
+
+  /**
+   * Name of the evaluator
+   */
+  name: string;
+
+  /**
+   * Unique slug identifier for the evaluator
+   */
+  slug: string;
+
+  /**
+   * ISO timestamp when the evaluator was last updated
+   */
+  updatedAt: string;
+}
+
+export namespace EvaluationGetEvaluatorByIDResponse {
+  export interface UnionMember0 {
+    /**
+     * Unique identifier for the block
+     */
+    id: string;
+
+    /**
+     * Block type identifier
+     */
+    blockType: 'CUSTOM_PROMPT';
+
+    /**
+     * Optional description of what this block evaluates
+     */
+    description: string | null;
+
+    /**
+     * Name of the metric this prompt evaluates
+     */
+    metricName: string;
+
+    /**
+     * Display name of the evaluation block
+     */
+    name: string;
+
+    /**
+     * Order in which this block is executed
+     */
+    orderIndex: number;
+
+    /**
+     * The prompt to evaluate the call against
+     */
+    prompt: string;
+
+    /**
+     * Minimum score threshold to pass evaluation (0-1)
+     */
+    threshold: number;
+
+    /**
+     * Weight of this block in the overall evaluation score (0-100)
+     */
+    weight: number;
+  }
+
+  export interface UnionMember1 {
+    /**
+     * Unique identifier for the block
+     */
+    id: string;
+
+    /**
+     * Block type identifier
+     */
+    blockType: 'DATAFIELD_CHECK';
+
+    /**
+     * Optional description of what this block evaluates
+     */
+    description: string | null;
+
+    /**
+     * Criteria for evaluating the property
+     */
+    evaluationCriteria: string;
+
+    /**
+     * Whether this property must be present
+     */
+    isRequired: boolean;
+
+    /**
+     * Display name of the evaluation block
+     */
+    name: string;
+
+    /**
+     * Order in which this block is executed
+     */
+    orderIndex: number;
+
+    /**
+     * Name of the property to check
+     */
+    propertyName: string;
+
+    /**
+     * Minimum score threshold to pass evaluation (0-1)
+     */
+    threshold: number;
+
+    /**
+     * Expected type of the property value
+     */
+    valueType: string;
+
+    /**
+     * Weight of this block in the overall evaluation score (0-100)
+     */
+    weight: number;
+  }
+
+  export interface UnionMember2 {
+    /**
+     * Unique identifier for the block
+     */
+    id: string;
+
+    /**
+     * Block type identifier
+     */
+    blockType: 'EMOTION';
+
+    /**
+     * Optional description of what this block evaluates
+     */
+    description: string | null;
+
+    /**
+     * Display name of the evaluation block
+     */
+    name: string;
+
+    /**
+     * Order in which this block is executed
+     */
+    orderIndex: number;
+
+    /**
+     * The emotion to detect (e.g., "joy", "anger", "sadness")
+     */
+    selectedEmotion: string;
+
+    /**
+     * Minimum confidence threshold for emotion detection (0-1)
+     */
+    threshold: number;
+
+    /**
+     * Weight of this block in the overall evaluation score (0-100)
+     */
+    weight: number;
+  }
+
+  export interface UnionMember3 {
+    /**
+     * Unique identifier for the block
+     */
+    id: string;
+
+    /**
+     * Block type identifier
+     */
+    blockType: 'LATENCY';
+
+    /**
+     * Optional description of what this block evaluates
+     */
+    description: string | null;
+
+    /**
+     * Maximum number of silence periods allowed
+     */
+    maxAllowedSilences: number;
+
+    /**
+     * Minimum duration of silence in milliseconds to be considered
+     */
+    minSilenceDuration: number;
+
+    /**
+     * Display name of the evaluation block
+     */
+    name: string;
+
+    /**
+     * Order in which this block is executed
+     */
+    orderIndex: number;
+
+    /**
+     * Maximum allowed latency score
+     */
+    threshold: number;
+
+    /**
+     * Weight of this block in the overall evaluation score (0-100)
+     */
+    weight: number;
+  }
+
+  export interface UnionMember4 {
+    /**
+     * Unique identifier for the block
+     */
+    id: string;
+
+    /**
+     * Block type identifier
+     */
+    blockType: 'POLITENESS';
+
+    /**
+     * Optional description of what this block evaluates
+     */
+    description: string | null;
+
+    /**
+     * Display name of the evaluation block
+     */
+    name: string;
+
+    /**
+     * Order in which this block is executed
+     */
+    orderIndex: number;
+
+    /**
+     * Minimum politeness score threshold (0-1)
+     */
+    threshold: number;
+
+    /**
+     * Weight of this block in the overall evaluation score (0-100)
+     */
+    weight: number;
+  }
+
+  export interface UnionMember5 {
+    /**
+     * Unique identifier for the block
+     */
+    id: string;
+
+    /**
+     * Block type identifier
+     */
+    blockType: 'SENTIMENT';
+
+    /**
+     * Optional description of what this block evaluates
+     */
+    description: string | null;
+
+    /**
+     * Display name of the evaluation block
+     */
+    name: string;
+
+    /**
+     * Order in which this block is executed
+     */
+    orderIndex: number;
+
+    /**
+     * Minimum sentiment score threshold (0-1)
+     */
+    threshold: number;
+
+    /**
+     * Weight of this block in the overall evaluation score (0-100)
+     */
+    weight: number;
+  }
+
+  export interface UnionMember6 {
+    /**
+     * Unique identifier for the block
+     */
+    id: string;
+
+    /**
+     * Block type identifier
+     */
+    blockType: 'TOOL_CALLS';
+
+    /**
+     * Optional description of what this block evaluates
+     */
+    description: string | null;
+
+    /**
+     * Condition that must be met for tool invocation
+     */
+    invocationCondition: string | null;
+
+    /**
+     * Minimum number of times the tool should be invoked
+     */
+    minInvocationCount: number | null;
+
+    /**
+     * Display name of the evaluation block
+     */
+    name: string;
+
+    /**
+     * Order in which this block is executed
+     */
+    orderIndex: number;
+
+    /**
+     * Whether the tool should be invoked
+     */
+    shouldBeInvoked: boolean;
+
+    /**
+     * ID of the tool definition
+     */
+    toolDefinitionId: string;
+
+    /**
+     * Weight of this block in the overall evaluation score (0-100)
+     */
+    weight: number;
+  }
+
+  export interface UnionMember7 {
+    /**
+     * Unique identifier for the block
+     */
+    id: string;
+
+    /**
+     * Block type identifier
+     */
+    blockType: 'TOXICITY';
+
+    /**
+     * Optional description of what this block evaluates
+     */
+    description: string | null;
+
+    /**
+     * Display name of the evaluation block
+     */
+    name: string;
+
+    /**
+     * Order in which this block is executed
+     */
+    orderIndex: number;
+
+    /**
+     * Maximum allowed toxicity score (0-1)
+     */
+    threshold: number;
+
+    /**
+     * Weight of this block in the overall evaluation score (0-100)
+     */
+    weight: number;
+  }
+
+  export interface UnionMember8 {
+    /**
+     * Unique identifier for the block
+     */
+    id: string;
+
+    /**
+     * Block type identifier
+     */
+    blockType: 'VOCAL_CUE';
+
+    /**
+     * Optional description of what this block evaluates
+     */
+    description: string | null;
+
+    /**
+     * Display name of the evaluation block
+     */
+    name: string;
+
+    /**
+     * Order in which this block is executed
+     */
+    orderIndex: number;
+
+    /**
+     * The vocal cue to detect (e.g., "pace", "tone", "volume")
+     */
+    selectedCue: string;
+
+    /**
+     * Minimum confidence threshold for vocal cue detection (0-1)
+     */
+    threshold: number;
+
+    /**
+     * Weight of this block in the overall evaluation score (0-100)
+     */
+    weight: number;
+  }
+}
+
+/**
+ * Response containing evaluators and pagination info
+ */
+export interface EvaluationGetEvaluatorsResponse {
+  /**
+   * Array of evaluators with their blocks
+   */
+  data: Array<EvaluationGetEvaluatorsResponse.Data>;
+
+  /**
+   * Pagination information
+   */
+  pagination: EvaluationGetEvaluatorsResponse.Pagination;
+}
+
+export namespace EvaluationGetEvaluatorsResponse {
+  /**
+   * Evaluator with its configured blocks
+   */
+  export interface Data {
+    /**
+     * Unique identifier for the evaluator
+     */
+    id: string;
+
+    /**
+     * Array of evaluation blocks configured for this evaluator
+     */
+    blocks: Array<
+      | Data.UnionMember0
+      | Data.UnionMember1
+      | Data.UnionMember2
+      | Data.UnionMember3
+      | Data.UnionMember4
+      | Data.UnionMember5
+      | Data.UnionMember6
+      | Data.UnionMember7
+      | Data.UnionMember8
+    >;
+
+    /**
+     * ISO timestamp when the evaluator was created
+     */
+    createdAt: string;
+
+    /**
+     * Optional description of the evaluator
+     */
+    description: string | null;
+
+    /**
+     * Name of the evaluator
+     */
+    name: string;
+
+    /**
+     * Unique slug identifier for the evaluator
+     */
+    slug: string;
+
+    /**
+     * ISO timestamp when the evaluator was last updated
+     */
+    updatedAt: string;
+  }
+
+  export namespace Data {
+    export interface UnionMember0 {
+      /**
+       * Unique identifier for the block
+       */
+      id: string;
+
+      /**
+       * Block type identifier
+       */
+      blockType: 'CUSTOM_PROMPT';
+
+      /**
+       * Optional description of what this block evaluates
+       */
+      description: string | null;
+
+      /**
+       * Name of the metric this prompt evaluates
+       */
+      metricName: string;
+
+      /**
+       * Display name of the evaluation block
+       */
+      name: string;
+
+      /**
+       * Order in which this block is executed
+       */
+      orderIndex: number;
+
+      /**
+       * The prompt to evaluate the call against
+       */
+      prompt: string;
+
+      /**
+       * Minimum score threshold to pass evaluation (0-1)
+       */
+      threshold: number;
+
+      /**
+       * Weight of this block in the overall evaluation score (0-100)
+       */
+      weight: number;
+    }
+
+    export interface UnionMember1 {
+      /**
+       * Unique identifier for the block
+       */
+      id: string;
+
+      /**
+       * Block type identifier
+       */
+      blockType: 'DATAFIELD_CHECK';
+
+      /**
+       * Optional description of what this block evaluates
+       */
+      description: string | null;
+
+      /**
+       * Criteria for evaluating the property
+       */
+      evaluationCriteria: string;
+
+      /**
+       * Whether this property must be present
+       */
+      isRequired: boolean;
+
+      /**
+       * Display name of the evaluation block
+       */
+      name: string;
+
+      /**
+       * Order in which this block is executed
+       */
+      orderIndex: number;
+
+      /**
+       * Name of the property to check
+       */
+      propertyName: string;
+
+      /**
+       * Minimum score threshold to pass evaluation (0-1)
+       */
+      threshold: number;
+
+      /**
+       * Expected type of the property value
+       */
+      valueType: string;
+
+      /**
+       * Weight of this block in the overall evaluation score (0-100)
+       */
+      weight: number;
+    }
+
+    export interface UnionMember2 {
+      /**
+       * Unique identifier for the block
+       */
+      id: string;
+
+      /**
+       * Block type identifier
+       */
+      blockType: 'EMOTION';
+
+      /**
+       * Optional description of what this block evaluates
+       */
+      description: string | null;
+
+      /**
+       * Display name of the evaluation block
+       */
+      name: string;
+
+      /**
+       * Order in which this block is executed
+       */
+      orderIndex: number;
+
+      /**
+       * The emotion to detect (e.g., "joy", "anger", "sadness")
+       */
+      selectedEmotion: string;
+
+      /**
+       * Minimum confidence threshold for emotion detection (0-1)
+       */
+      threshold: number;
+
+      /**
+       * Weight of this block in the overall evaluation score (0-100)
+       */
+      weight: number;
+    }
+
+    export interface UnionMember3 {
+      /**
+       * Unique identifier for the block
+       */
+      id: string;
+
+      /**
+       * Block type identifier
+       */
+      blockType: 'LATENCY';
+
+      /**
+       * Optional description of what this block evaluates
+       */
+      description: string | null;
+
+      /**
+       * Maximum number of silence periods allowed
+       */
+      maxAllowedSilences: number;
+
+      /**
+       * Minimum duration of silence in milliseconds to be considered
+       */
+      minSilenceDuration: number;
+
+      /**
+       * Display name of the evaluation block
+       */
+      name: string;
+
+      /**
+       * Order in which this block is executed
+       */
+      orderIndex: number;
+
+      /**
+       * Maximum allowed latency score
+       */
+      threshold: number;
+
+      /**
+       * Weight of this block in the overall evaluation score (0-100)
+       */
+      weight: number;
     }
-    return this._client.get('/v1/evaluation/evaluators', { query, ...options });
-  }
 
-  /**
-   * Retrieve details of a specific evaluation job
-   */
-  getJob(jobId: string, options?: Core.RequestOptions): Core.APIPromise<EvaluationGetJobResponse> {
-    return this._client.get(`/v1/evaluation/job/${jobId}`, options);
-  }
+    export interface UnionMember4 {
+      /**
+       * Unique identifier for the block
+       */
+      id: string;
 
-  /**
-   * Retrieve paginated details of a specific evaluation job runs
-   */
-  getJobRuns(
-    jobId: string,
-    query?: EvaluationGetJobRunsParams,
-    options?: Core.RequestOptions,
-  ): Core.APIPromise<EvaluationGetJobRunsResponse>;
-  getJobRuns(jobId: string, options?: Core.RequestOptions): Core.APIPromise<EvaluationGetJobRunsResponse>;
-  getJobRuns(
-    jobId: string,
-    query: EvaluationGetJobRunsParams | Core.RequestOptions = {},
-    options?: Core.RequestOptions,
-  ): Core.APIPromise<EvaluationGetJobRunsResponse> {
-    if (isRequestOptions(query)) {
-      return this.getJobRuns(jobId, {}, query);
+      /**
+       * Block type identifier
+       */
+      blockType: 'POLITENESS';
+
+      /**
+       * Optional description of what this block evaluates
+       */
+      description: string | null;
+
+      /**
+       * Display name of the evaluation block
+       */
+      name: string;
+
+      /**
+       * Order in which this block is executed
+       */
+      orderIndex: number;
+
+      /**
+       * Minimum politeness score threshold (0-1)
+       */
+      threshold: number;
+
+      /**
+       * Weight of this block in the overall evaluation score (0-100)
+       */
+      weight: number;
+    }
+
+    export interface UnionMember5 {
+      /**
+       * Unique identifier for the block
+       */
+      id: string;
+
+      /**
+       * Block type identifier
+       */
+      blockType: 'SENTIMENT';
+
+      /**
+       * Optional description of what this block evaluates
+       */
+      description: string | null;
+
+      /**
+       * Display name of the evaluation block
+       */
+      name: string;
+
+      /**
+       * Order in which this block is executed
+       */
+      orderIndex: number;
+
+      /**
+       * Minimum sentiment score threshold (0-1)
+       */
+      threshold: number;
+
+      /**
+       * Weight of this block in the overall evaluation score (0-100)
+       */
+      weight: number;
+    }
+
+    export interface UnionMember6 {
+      /**
+       * Unique identifier for the block
+       */
+      id: string;
+
+      /**
+       * Block type identifier
+       */
+      blockType: 'TOOL_CALLS';
+
+      /**
+       * Optional description of what this block evaluates
+       */
+      description: string | null;
+
+      /**
+       * Condition that must be met for tool invocation
+       */
+      invocationCondition: string | null;
+
+      /**
+       * Minimum number of times the tool should be invoked
+       */
+      minInvocationCount: number | null;
+
+      /**
+       * Display name of the evaluation block
+       */
+      name: string;
+
+      /**
+       * Order in which this block is executed
+       */
+      orderIndex: number;
+
+      /**
+       * Whether the tool should be invoked
+       */
+      shouldBeInvoked: boolean;
+
+      /**
+       * ID of the tool definition
+       */
+      toolDefinitionId: string;
+
+      /**
+       * Weight of this block in the overall evaluation score (0-100)
+       */
+      weight: number;
+    }
+
+    export interface UnionMember7 {
+      /**
+       * Unique identifier for the block
+       */
+      id: string;
+
+      /**
+       * Block type identifier
+       */
+      blockType: 'TOXICITY';
+
+      /**
+       * Optional description of what this block evaluates
+       */
+      description: string | null;
+
+      /**
+       * Display name of the evaluation block
+       */
+      name: string;
+
+      /**
+       * Order in which this block is executed
+       */
+      orderIndex: number;
+
+      /**
+       * Maximum allowed toxicity score (0-1)
+       */
+      threshold: number;
+
+      /**
+       * Weight of this block in the overall evaluation score (0-100)
+       */
+      weight: number;
+    }
+
+    export interface UnionMember8 {
+      /**
+       * Unique identifier for the block
+       */
+      id: string;
+
+      /**
+       * Block type identifier
+       */
+      blockType: 'VOCAL_CUE';
+
+      /**
+       * Optional description of what this block evaluates
+       */
+      description: string | null;
+
+      /**
+       * Display name of the evaluation block
+       */
+      name: string;
+
+      /**
+       * Order in which this block is executed
+       */
+      orderIndex: number;
+
+      /**
+       * The vocal cue to detect (e.g., "pace", "tone", "volume")
+       */
+      selectedCue: string;
+
+      /**
+       * Minimum confidence threshold for vocal cue detection (0-1)
+       */
+      threshold: number;
+
+      /**
+       * Weight of this block in the overall evaluation score (0-100)
+       */
+      weight: number;
     }
-    return this._client.get(`/v1/evaluation/job/${jobId}/runs`, { query, ...options });
   }
-}
 
-export interface EvaluationCreateJobResponse {
-  data: EvaluationCreateJobResponse.Data;
-}
+  /**
+   * Pagination information
+   */
+  export interface Pagination {
+    /**
+     * Whether there are more evaluators to fetch
+     */
+    hasMore: boolean;
 
-export namespace EvaluationCreateJobResponse {
-  export interface Data {
     /**
-     * ID of the evaluation job
+     * Cursor for the next page, null if no more pages
      */
-    jobId: string;
+    nextCursor: string | null;
 
     /**
-     * Status of the evaluation job
+     * Total number of evaluators
      */
-    status: 'PENDING' | 'PROCESSING' | 'SUCCESS' | 'FAILURE';
+    total: number;
   }
 }
 
-export type EvaluationGetEvaluatorByIDResponse = unknown;
-
-export type EvaluationGetEvaluatorsResponse = unknown;
-
 export interface EvaluationGetJobResponse {
   /**
    * Evaluation job response payload
@@ -187,114 +1123,155 @@ export namespace EvaluationGetJobRunsResponse {
   export namespace Data {
     export interface Data {
       /**
-       * ID of the evaluator run
+       * All block runs for this evaluator, including skipped ones
        */
-      id: string;
+      blockRuns: Array<Data.BlockRun>;
 
-      /**
-       * When the evaluator run completed
-       */
-      completedAt: string | null;
+      evaluator: Data.Evaluator;
 
-      /**
-       * Evaluator of the evaluator run
-       */
-      evaluator: Data.Evaluator | null;
+      evidence: Array<Data.Evidence>;
+
+      metrics: Array<Data.Metric>;
 
       /**
-       * Evidence of the evaluator run
+       * Status of the evaluator run
        */
-      evidence: Array<Data.Evidence> | null;
+      status: 'PENDING' | 'IN_PROGRESS' | 'COMPLETED' | 'FAILED';
 
       /**
-       * Metrics of the evaluator run
+       * ID of the evaluator run
        */
-      metrics: Array<Data.Metric> | null;
+      id?: string;
 
       /**
-       * Score of the evaluator run
+       * When the evaluator run completed
        */
-      score: number | null;
+      completedAt?: string | null;
 
       /**
-       * Score classification of the evaluator run
+       * Score of the evaluation run (0-1)
        */
-      scoreClassification: 'SUCCESS' | 'FAILURE' | 'IRRELEVANT' | null;
+      score?: number | null;
 
       /**
-       * When the evaluator run started
+       * Score classification of the evaluator run based on score threshold (IRRELEVANT
+       * is mapped to SKIPPED)
        */
-      startedAt: string | null;
+      scoreClassification?: 'SUCCESS' | 'FAILURE' | 'SKIPPED' | null;
 
       /**
-       * Status of the evaluator run
+       * When the evaluator run started
        */
-      status: 'PENDING' | 'IN_PROGRESS' | 'COMPLETED' | 'FAILED';
+      startedAt?: string | null;
 
       /**
-       * Summary of the evaluator run
+       * Summary of the evaluation run
        */
-      summary: string | null;
+      summary?: string | null;
     }
 
     export namespace Data {
-      /**
-       * Evaluator of the evaluator run
-       */
+      export interface BlockRun {
+        /**
+         * ID of the block definition
+         */
+        blockDefinitionId: string;
+
+        /**
+         * Name of the evaluation block
+         */
+        blockName: string;
+
+        /**
+         * ID of the block run instance
+         */
+        blockRunId: string;
+
+        /**
+         * When the block run was created
+         */
+        createdAt: string;
+
+        /**
+         * Reason for the outcome (pass/fail explanation or skip reason)
+         */
+        reason: string | null;
+
+        /**
+         * Result of the block run
+         */
+        result: 'PASSED' | 'FAILED' | 'SKIPPED' | null;
+
+        /**
+         * Score of the block run (0-1)
+         */
+        score: number | null;
+
+        /**
+         * Status of the block run
+         */
+        status: 'PENDING' | 'RUNNING' | 'COMPLETED' | 'FAILED';
+      }
+
       export interface Evaluator {
         /**
          * ID of the evaluator
          */
-        id: string | null;
+        id: string;
 
         /**
          * Name of the evaluator
          */
-        name: string | null;
-      }
+        name: string;
 
-      export interface Evidence {
         /**
-         * ID of the evidence
+         * Weight of the evaluator
          */
-        id: string | null;
+        weight?: number;
+      }
 
+      export interface Evidence {
         /**
-         * Comment on the evidence
+         * Comment text of the evidence
          */
         commentText: string | null;
 
         /**
-         * Whether this is a positive example of the metric
+         * Created at of the evidence
          */
-        isPositive: boolean | null;
+        createdAt: string;
 
         /**
-         * Snippet of the evidence
+         * Is positive of the evidence
          */
-        snippetText: string | null;
-      }
+        isPositive: boolean;
 
-      export interface Metric {
         /**
-         * ID of the metric
+         * Snippet text of the evidence
          */
-        id: string | null;
+        snippetText: string;
+      }
 
+      export interface Metric {
         /**
          * Boolean value of the metric
          */
         booleanValue: boolean | null;
 
         /**
-         * Confidence of the metric
+         * Confidence level of the metric (0-1)
          */
         confidence: number | null;
 
+        /**
+         * Created at of the metric
+         */
+        createdAt: string;
+
         /**
          * Name of the metric
          */
-        name: string | null;
+        name: string;
 
         /**
          * Numeric value of the metric
@@ -302,14 +1279,14 @@ export namespace EvaluationGetJobRunsResponse {
         numericValue: number | null;
 
         /**
-         * Reasoning for the metric
+         * Reasoning of the metric
          */
         reasoning: string | null;
 
         /**
          * Role of the metric
          */
-        role: 'PRIMARY' | 'SECONDARY' | null;
+        role: string;
 
         /**
          * Text value of the metric
@@ -319,7 +1296,7 @@ export namespace EvaluationGetJobRunsResponse {
         /**
          * Value type of the metric
          */
-        valueType: 'NUMERIC' | 'BOOLEAN' | 'TEXT' | null;
+        valueType: string;
       }
     }
 
@@ -658,8 +1635,14 @@ export namespace EvaluationCreateJobParams {
 }
 
 export interface EvaluationGetEvaluatorsParams {
+  /**
+   * Cursor for pagination - evaluator ID to start after
+   */
   after?: string;
 
+  /**
+   * Maximum number of evaluators to return (default: 20, max: 50)
+   */
   limit?: string;
 }