EvalSampleObject

Example Usage

import { EvalSampleObject } from "@meetkai/mka1/models/components";

let value: EvalSampleObject = {
  id: "<id>",
  object: "eval.sample",
  runId: "<id>",
  taskId: "<id>",
  model: "Taurus",
  sampleIndex: 823042,
  status: "running",
  datasetRow: {},
  prompt: "<value>",
  target: "<value>",
  responseId: null,
  outputText: "<value>",
  reasoning: null,
  extractedOutput: null,
  scores: {},
  judge: {
    "key": "<value>",
    "key1": "<value>",
  },
  error: {
    "key": "<value>",
  },
  costUsd: 6761.6,
  inputTokens: 809932,
  outputTokens: null,
  cachedTokens: 573425,
  phaseTimings: {},
  createdAt: 663201,
  startedAt: 954309,
  completedAt: 947537,
};

Fields

Field	Type	Required	Description
id	string	✔️	N/A
object	"eval.sample"	✔️	N/A
runId	string	✔️	N/A
taskId	string	✔️	N/A
model	string	✔️	N/A
sampleIndex	number	✔️	N/A
status	components.EvalSampleStatus	✔️	N/A
datasetRow	Record<string, any>	✔️	N/A
prompt	string	✔️	N/A
target	string	✔️	N/A
responseId	string	✔️	N/A
outputText	string	✔️	N/A
reasoning	string	✔️	Model reasoning ('thinking') captured from the response's reasoning items, kept separate from output_text so graders and extraction never see it. Truncated at 64 KB of UTF-8 with a marker appended. Null when the model returned none, when the provider only sent encrypted reasoning, and for transcription tasks.
extractedOutput	string	✔️	N/A
scores	Record<string, any>	✔️	N/A
judge	Record<string, any>	✔️	N/A
error	Record<string, any>	✔️	The exception that ended this sample, if any. For a harness trial this also carries an agent timeout — which is still a scored trial, so do not read a non-null error here as an unscored one; check status and scores.
costUsd	number	✔️	Spend for this one trial in USD. Null outside harness runs, and when the harness reported no cost.
inputTokens	number	✔️	Prompt tokens for the whole trial session. Includes cached_tokens.
outputTokens	number	✔️	Completion tokens for the whole trial session.
cachedTokens	number	✔️	The share of input_tokens served from the provider's prompt cache. Reported separately because it is what makes two agents at the same token count cost an order of magnitude apart.
phaseTimings	Record<string, any>	✔️	Per-phase wall clock for a harness trial: environment_setup, agent_setup, agent_execution, verifier, each { started_at, finished_at } as ISO 8601 timestamps exactly as the harness recorded them. A phase the harness did not time is absent rather than null. Drives the trial timing bar. Null outside harness runs.
createdAt	number	✔️	N/A
startedAt	number	✔️	N/A
completedAt	number	✔️	N/A