{
  "created_at": "2026-09-24T07:12:47.758109Z",
  "failure": null,
  "partial": false,
  "provenance": {
    "endpoint": {
      "base_url": "http://127.0.0.1:39283/v1",
      "model": "thinkingcap-awq-5090"
    },
    "finished_at": "2026-09-24T07:12:47.757978Z",
    "started_at": "2026-09-24T07:12:37.965276Z",
    "submitted_at": "2026-09-24T04:58:24.725917Z",
    "worker": {
      "id": "operator"
    }
  },
  "results": {
    "evidence": {
      "completeness": "completed",
      "created_at": "2026-09-24T07:12:47.744379Z",
      "evidence_kind": "measured",
      "failure": null,
      "identities": {
        "concurrency": {
          "requests": 1,
          "workers": 1
        },
        "context": {
          "configured_tokens": 131072,
          "output_headroom_tokens": "not-applicable",
          "profile_buckets": "not-applicable",
          "selected_buckets": "not-applicable"
        },
        "dataset": {
          "name": "deterministic-native",
          "profile_sha256": "090745451099214af7c222f736769737911453fe9e9c5d14ffc04f5a4a3e4586"
        },
        "hardware": {
          "id": "rtx5090"
        },
        "harnesses": {
          "native": {
            "schema": "anvil-serving.agentic/v1"
          }
        },
        "image": {
          "status": "not-applicable"
        },
        "model": {
          "id": "thinkingcap-awq-5090",
          "revision": "f8fe157f207a13f977bc3d620ce10a3e9ba5ab11"
        },
        "runtime": {
          "engine": "vllm",
          "image": "sha256:082ca6f035279109041ffd3fe0695cb568b29bc580b35c4f297a66a08b216c1b",
          "revision": "98dff2a81d747d1dba01a47f939f48c3526d4206"
        },
        "served_model": {
          "configured_context": 131072,
          "id": "thinkingcap-awq-5090"
        },
        "topology": {
          "kind": "co-resident-benchmark-client",
          "model_host_id": "operator",
          "worker_id": "operator"
        }
      },
      "promotion": {
        "authorized": false,
        "message": "Benchmark evidence does not authorize model promotion; promotion is a separate human decision."
      },
      "run": {
        "ownership_id": "thinkingcap-5090",
        "profile": "smoke",
        "run_id": "thinkingcap-5090-agentic-bf16-128k-mtp-uva3-v1",
        "spec_sha256": "32b33ac5b465091dd39c0efefcfe07f53114ee5571224de2a86f8ab8711b105b",
        "suite": "agentic"
      },
      "schema": "anvil-serving.benchmark-evidence/v1",
      "stages": [
        {
          "evidence": [
            {
              "bytes": 567,
              "path": "evidence/0-assets.json",
              "sha256": "f3b81c4404e73151b371512eede16c8d74c9f6e3ac69049e8175fcc7c12c8679"
            }
          ],
          "name": "asset_preparation",
          "sequence": 0,
          "status": "completed"
        },
        {
          "evidence": [
            {
              "bytes": 1991,
              "path": "evidence/0-preflight.json",
              "sha256": "225cd99cdc7f55387e9994f0a517c6c97047d1bac42fb6c0e7c6e1727f5e7952"
            }
          ],
          "name": "preflight",
          "sequence": 1,
          "status": "completed"
        },
        {
          "evidence": [
            {
              "bytes": 4431,
              "path": "evidence/0-agentic.json",
              "sha256": "486e312e2ba4e71a2d0628972bb5c57692230f80c8621b77f68b2ca7f1939d86"
            }
          ],
          "name": "agentic",
          "sequence": 2,
          "status": "completed"
        }
      ],
      "summary": {
        "attempted": 2,
        "pass_rate": 1.0,
        "passed": 2,
        "required_pass_rate": 0.8
      }
    }
  },
  "run": {
    "ownership_id": "thinkingcap-5090",
    "profile": "smoke",
    "run_id": "thinkingcap-5090-agentic-bf16-128k-mtp-uva3-v1",
    "spec_sha256": "32b33ac5b465091dd39c0efefcfe07f53114ee5571224de2a86f8ab8711b105b",
    "suite": "agentic"
  },
  "schema": "anvil-serving.benchmark-result/v1",
  "status": "completed"
}