{
  "evidenceOrigin": "client-reported",
  "primaryOutcome": "independently-verified-task-success",
  "baseline": {
    "planned": 4,
    "attempted": 4,
    "passed": 4,
    "successRate": 1,
    "failures": 0,
    "errors": 0,
    "notCompleted": 0,
    "meanAgentDurationMs": 52256.09794774999,
    "inputTokens": {
      "knownTotal": 243090,
      "reportedTrials": 4,
      "totalTrials": 4,
      "complete": true
    },
    "outputTokens": {
      "knownTotal": 4317,
      "reportedTrials": 4,
      "totalTrials": 4,
      "complete": true
    },
    "costUsd": {
      "knownTotal": 0,
      "reportedTrials": 0,
      "totalTrials": 4,
      "complete": false
    }
  },
  "candidate": {
    "planned": 4,
    "attempted": 4,
    "passed": 4,
    "successRate": 1,
    "failures": 0,
    "errors": 0,
    "notCompleted": 0,
    "meanAgentDurationMs": 64220.46276024999,
    "inputTokens": {
      "knownTotal": 266419,
      "reportedTrials": 4,
      "totalTrials": 4,
      "complete": true
    },
    "outputTokens": {
      "knownTotal": 5845,
      "reportedTrials": 4,
      "totalTrials": 4,
      "complete": true
    },
    "costUsd": {
      "knownTotal": 0,
      "reportedTrials": 0,
      "totalTrials": 4,
      "complete": false
    }
  },
  "effect": {
    "estimate": 0,
    "confidence": 0.95,
    "interval": [
      -1,
      1
    ],
    "method": "paired-hoeffding",
    "units": "success-rate-difference"
  },
  "discordance": {
    "candidateWins": 0,
    "baselineWins": 0,
    "pValue": 1,
    "method": "exact-two-sided-binomial"
  },
  "conclusion": "inconclusive",
  "verifierQualification": "controls-matched",
  "scope": "This task, surface snapshots, agent configuration, verifier and execution environment only.",
  "cautions": [
    "The confidence bound assumes independent repeated pairs. Shared caches, provider drift or side effects can violate that assumption.",
    "A verifier is independent of the agent's self-assessment, but its correctness and coverage require separate validation.",
    "The p-value is descriptive for this predeclared comparison. Repeated peeking, selecting tasks or testing multiple changes needs a separate multiplicity plan.",
    "Durations and usage are descriptive; missing token or cost telemetry is unknown, not zero.",
    "Verifier controls only cover their predeclared examples; matching them does not prove the verifier is correct for all outcomes."
  ]
}