{
  "experimentId": "operations-mode-trials-v1",
  "verifiedOn": "2026-08-09",
  "measurementStatus": "illustrative-not-live-provider-runs",
  "settingSemantics": "settings contains normalized experiment labels. Standard speed means omitting the speed request field. Fast speed requires the documented request field, beta header, platform, and access prerequisites recorded on that configuration.",
  "task": "Resolve an ambiguous refund-policy exception from two dated source documents and return a reviewed JSON recommendation",
  "gate": {
    "minimumRunCount": 3,
    "minimumQuality": 0.95,
    "maximumP95LatencyMs": 8000,
    "maximumMeanCostUnits": 5.0
  },
  "verificationPolicy": {
    "refreshBeforeExperiment": true,
    "rule": "Verify model and platform support in current official documentation before running or copying any configuration.",
    "sources": [
      {
        "id": "sampling-reference",
        "verifiedOn": "2026-08-09",
        "claim": "Sampling controls are model-dependent, and even temperature zero on models that accept it is not fully deterministic.",
        "sourceUrl": "https://platform.claude.com/docs/en/api/messages/create"
      },
      {
        "id": "sampling-support",
        "verifiedOn": "2026-08-09",
        "claim": "Current Messages guidance says Claude 4.7 and later reject non-default temperature, top_p, and top_k values.",
        "sourceUrl": "https://platform.claude.com/docs/en/build-with-claude/working-with-messages"
      },
      {
        "id": "model-versioning",
        "verifiedOn": "2026-08-09",
        "claim": "A pinned model ID fixes model weights, while serving infrastructure can still change observable behavior.",
        "sourceUrl": "https://platform.claude.com/docs/en/about-claude/models/model-ids-and-versions"
      },
      {
        "id": "models-overview",
        "verifiedOn": "2026-08-09",
        "claim": "The models overview lists claude-sonnet-5 and claude-opus-5 as the exact Claude API IDs and lists adaptive thinking for both.",
        "sourceUrl": "https://platform.claude.com/docs/en/about-claude/models/overview"
      },
      {
        "id": "sonnet-5-configuration",
        "verifiedOn": "2026-08-09",
        "claim": "Claude Sonnet 5 uses claude-sonnet-5, has adaptive thinking on by default, accepts thinking disabled, and supports the low and high effort values used here.",
        "sourceUrl": "https://platform.claude.com/docs/en/about-claude/models/whats-new-sonnet-5"
      },
      {
        "id": "opus-5-configuration",
        "verifiedOn": "2026-08-09",
        "claim": "Claude Opus 5 uses claude-opus-5, accepts adaptive thinking, and supports the medium effort value used here.",
        "sourceUrl": "https://platform.claude.com/docs/en/about-claude/models/whats-new-opus-5"
      },
      {
        "id": "thinking-configuration",
        "verifiedOn": "2026-08-09",
        "claim": "Claude Opus 5 and Sonnet 5 have thinking enabled by default; adaptive is accepted, and model-specific rules govern disabling it.",
        "sourceUrl": "https://platform.claude.com/docs/en/build-with-claude/thinking"
      },
      {
        "id": "effort-control",
        "verifiedOn": "2026-08-09",
        "claim": "Claude Sonnet 5 supports low and high effort, Claude Opus 5 supports medium effort, and effort is a behavioral signal rather than a strict token budget.",
        "sourceUrl": "https://platform.claude.com/docs/en/build-with-claude/effort"
      },
      {
        "id": "fast-mode",
        "verifiedOn": "2026-08-09",
        "claim": "Fast mode lists Claude Opus 5, not Claude Sonnet 5, as supported; it is a Claude API research preview requiring access, speed fast, and the fast-mode-2026-02-01 beta header, with premium pricing.",
        "sourceUrl": "https://platform.claude.com/docs/en/build-with-claude/fast-mode"
      }
    ]
  },
  "configurations": [
    {
      "id": "sonnet-standard-low",
      "modelId": "claude-sonnet-5",
      "platform": "claude-api",
      "settings": {"speed": "standard", "effort": "low", "thinking": "adaptive"},
      "support": {
        "status": "docs-supported",
        "verifiedOn": "2026-08-09",
        "sourceIds": ["models-overview", "sonnet-5-configuration", "thinking-configuration", "effort-control"]
      },
      "runs": [
        {"id": "low-01", "quality": 0.94, "latencyMs": 3400, "costUnits": 1.5, "outcomeFingerprint": "approve-with-two-citations"},
        {"id": "low-02", "quality": 0.91, "latencyMs": 3800, "costUnits": 1.5, "outcomeFingerprint": "escalate-missing-exception"},
        {"id": "low-03", "quality": 0.96, "latencyMs": 3600, "costUnits": 1.5, "outcomeFingerprint": "approve-with-three-citations"}
      ],
      "summary": {"runCount": 3, "minimumQuality": 0.91, "p95LatencyMs": 3800, "meanCostUnits": 1.5, "passesGate": false}
    },
    {
      "id": "sonnet-standard-high",
      "modelId": "claude-sonnet-5",
      "platform": "claude-api",
      "settings": {"speed": "standard", "effort": "high", "thinking": "adaptive"},
      "support": {
        "status": "docs-supported",
        "verifiedOn": "2026-08-09",
        "sourceIds": ["models-overview", "sonnet-5-configuration", "thinking-configuration", "effort-control"]
      },
      "runs": [
        {"id": "high-01", "quality": 0.98, "latencyMs": 7100, "costUnits": 3.0, "outcomeFingerprint": "escalate-with-policy-conflict"},
        {"id": "high-02", "quality": 0.96, "latencyMs": 7600, "costUnits": 3.0, "outcomeFingerprint": "escalate-with-date-conflict"},
        {"id": "high-03", "quality": 0.97, "latencyMs": 7300, "costUnits": 3.0, "outcomeFingerprint": "escalate-with-two-conflicts"}
      ],
      "summary": {"runCount": 3, "minimumQuality": 0.96, "p95LatencyMs": 7600, "meanCostUnits": 3.0, "passesGate": true}
    },
    {
      "id": "sonnet-standard-high-disabled",
      "modelId": "claude-sonnet-5",
      "platform": "claude-api",
      "settings": {"speed": "standard", "effort": "high", "thinking": "disabled"},
      "support": {
        "status": "docs-supported",
        "verifiedOn": "2026-08-09",
        "sourceIds": ["models-overview", "sonnet-5-configuration", "thinking-configuration", "effort-control"]
      },
      "runs": [
        {"id": "disabled-01", "quality": 0.92, "latencyMs": 5000, "costUnits": 2.4, "outcomeFingerprint": "approve-without-date-conflict"},
        {"id": "disabled-02", "quality": 0.95, "latencyMs": 5400, "costUnits": 2.4, "outcomeFingerprint": "escalate-with-policy-conflict"},
        {"id": "disabled-03", "quality": 0.93, "latencyMs": 5200, "costUnits": 2.4, "outcomeFingerprint": "approve-with-one-conflict"}
      ],
      "summary": {"runCount": 3, "minimumQuality": 0.92, "p95LatencyMs": 5400, "meanCostUnits": 2.4, "passesGate": false}
    },
    {
      "id": "opus-standard-medium",
      "modelId": "claude-opus-5",
      "platform": "claude-api",
      "settings": {"speed": "standard", "effort": "medium", "thinking": "adaptive"},
      "support": {
        "status": "docs-supported",
        "verifiedOn": "2026-08-09",
        "sourceIds": ["models-overview", "opus-5-configuration", "thinking-configuration", "effort-control"]
      },
      "runs": [
        {"id": "opus-standard-01", "quality": 0.98, "latencyMs": 7200, "costUnits": 4.5, "outcomeFingerprint": "escalate-with-policy-conflict"},
        {"id": "opus-standard-02", "quality": 0.97, "latencyMs": 7800, "costUnits": 4.5, "outcomeFingerprint": "escalate-with-date-conflict"},
        {"id": "opus-standard-03", "quality": 0.98, "latencyMs": 7500, "costUnits": 4.5, "outcomeFingerprint": "escalate-with-two-conflicts"}
      ],
      "summary": {"runCount": 3, "minimumQuality": 0.97, "p95LatencyMs": 7800, "meanCostUnits": 4.5, "passesGate": true}
    },
    {
      "id": "opus-fast-medium",
      "modelId": "claude-opus-5",
      "platform": "claude-api",
      "settings": {"speed": "fast", "effort": "medium", "thinking": "adaptive"},
      "support": {
        "status": "docs-supported",
        "verifiedOn": "2026-08-09",
        "sourceIds": ["models-overview", "opus-5-configuration", "thinking-configuration", "effort-control", "fast-mode"]
      },
      "requestRequirements": {
        "access": "Anthropic fast-mode research-preview access is required, and capacity is not guaranteed.",
        "requestField": "speed: fast",
        "betaHeaders": ["anthropic-beta: fast-mode-2026-02-01"]
      },
      "runs": [
        {"id": "fast-01", "quality": 0.98, "latencyMs": 3900, "costUnits": 8.0, "outcomeFingerprint": "escalate-with-policy-conflict"},
        {"id": "fast-02", "quality": 0.97, "latencyMs": 4200, "costUnits": 8.0, "outcomeFingerprint": "escalate-with-date-conflict"},
        {"id": "fast-03", "quality": 0.98, "latencyMs": 4100, "costUnits": 8.0, "outcomeFingerprint": "escalate-with-two-conflicts"}
      ],
      "summary": {"runCount": 3, "minimumQuality": 0.97, "p95LatencyMs": 4200, "meanCostUnits": 8.0, "passesGate": false}
    },
    {
      "id": "sonnet-fast-high",
      "modelId": "claude-sonnet-5",
      "platform": "claude-api",
      "settings": {"speed": "fast", "effort": "high", "thinking": "adaptive"},
      "support": {
        "status": "docs-unsupported",
        "verifiedOn": "2026-08-09",
        "sourceIds": ["models-overview", "fast-mode"]
      },
      "runs": [],
      "rejectionReason": "Current fast-mode documentation does not list this model as supported; do not send a trial request until current docs say otherwise."
    }
  ],
  "selectedConfiguration": "sonnet-standard-high",
  "decision": {
    "why": "Within these illustrative measurements, it is the least costly docs-supported configuration that clears repeated-run quality and p95 latency gates; documentation support does not prove account entitlement or a successful live request.",
    "repeatPolicy": "Run every supported configuration at least three times on each risk slice and compare minimum quality, tail latency, and mean accepted-outcome cost.",
    "changePolicy": "Rerun documentation verification and the full repeated evaluation when the model ID, platform, speed, effort, thinking behavior, prompt, tools, or serving assumptions change.",
    "rejected": [
      "Low effort failed the minimum-quality gate on one repeated run.",
      "Disabling thinking at high effort reduced cost and latency but failed the minimum-quality gate.",
      "Standard-speed Opus cleared the gates but cost more than the selected Sonnet configuration.",
      "Fast Opus met quality and latency but exceeded the cost gate.",
      "Unsupported combinations are recorded as rejected evidence, not attempted or silently downgraded."
    ]
  }
}