[
  {
    "_type": "https://in-toto.io/Statement/v1",
    "subject": [
      {
        "name": "j-rig:local:coreweave-gpu-cost-leak-hunter.deepseek-v4-flash",
        "digest": {
          "sha256": "561c9e3a2fdd4a1f4f39b763dd3caf4df0a914e2169c03499c5f8eda7ebbecf6"
        }
      }
    ],
    "predicateType": "https://evals.intentsolutions.io/gate-result/v1",
    "predicate": {
      "gate_id": "j-rig:local:coreweave-gpu-cost-leak-hunter.deepseek-v4-flash",
      "gate_name": "rollout",
      "gate_version": "0.1.2",
      "gate_decision": "fail",
      "gate_reasons": [
        "2 blocker criteria failed"
      ],
      "coverage": {
        "dimensions_evaluated": [
          "trigger",
          "functional",
          "behavioral"
        ],
        "dimensions_skipped": [
          "regression",
          "baseline"
        ]
      },
      "policy_ref": "sha256:683f601ca3d772b5ee242828cf8422873cccbec94c0e9fc444ed97ba6af1a6d4:eval-spec.yaml",
      "policy_hash": "sha256:683f601ca3d772b5ee242828cf8422873cccbec94c0e9fc444ed97ba6af1a6d4",
      "input_hash": "sha256:561c9e3a2fdd4a1f4f39b763dd3caf4df0a914e2169c03499c5f8eda7ebbecf6",
      "evaluated_at": "2026-07-07T00:22:37.638Z",
      "runner": "j-rig@0.1.2",
      "commit_sha": "561c9e3a2fdd4a1f4f39b763dd3caf4df0a914e2",
      "metadata": {
        "model": "deepseek-v4-flash",
        "provider": "deepseek",
        "ground_truth": true,
        "rollout_decision": "block",
        "pass_rate": 0.6,
        "passed": 6,
        "total_criteria": 10,
        "commit_sha_source": "skill-content-sha (skill not git-tracked)",
        "aggregation": {
          "rule": "majority-vote",
          "denominator_rule": "errored-samples-count-as-unsure",
          "blocker_quorum": 0.8,
          "samples_default": 5
        },
        "criteria": [
          {
            "criterion_id": "triggers-on-cost-question",
            "blocker": true,
            "method": "judge",
            "verdict": "yes",
            "samples": 5,
            "agreement": 1,
            "sample_verdicts": [
              "yes",
              "yes",
              "yes",
              "yes",
              "yes"
            ]
          },
          {
            "criterion_id": "produces-cfo-grokkable-report",
            "blocker": true,
            "method": "judge",
            "verdict": "no",
            "samples": 5,
            "agreement": 1,
            "sample_verdicts": [
              "no",
              "no",
              "no",
              "no",
              "no"
            ]
          },
          {
            "criterion_id": "splits-confirmed-vs-estimated",
            "blocker": false,
            "method": "judge",
            "verdict": "yes",
            "samples": 5,
            "agreement": 1,
            "sample_verdicts": [
              "yes",
              "yes",
              "yes",
              "yes",
              "yes"
            ]
          },
          {
            "criterion_id": "dollars-from-promql-not-guessed",
            "blocker": false,
            "method": "judge",
            "verdict": "no",
            "samples": 5,
            "agreement": 1,
            "sample_verdicts": [
              "no",
              "no",
              "no",
              "no",
              "no"
            ]
          },
          {
            "criterion_id": "triggers-on-cost-question",
            "blocker": true,
            "method": "judge",
            "verdict": "yes",
            "samples": 5,
            "agreement": 0.8,
            "sample_verdicts": [
              "yes",
              "no",
              "yes",
              "yes",
              "yes"
            ]
          },
          {
            "criterion_id": "produces-cfo-grokkable-report",
            "blocker": true,
            "method": "judge",
            "verdict": "no",
            "samples": 5,
            "agreement": 1,
            "sample_verdicts": [
              "no",
              "no",
              "no",
              "no",
              "no"
            ]
          },
          {
            "criterion_id": "splits-confirmed-vs-estimated",
            "blocker": false,
            "method": "judge",
            "verdict": "yes",
            "samples": 5,
            "agreement": 1,
            "sample_verdicts": [
              "yes",
              "yes",
              "yes",
              "yes",
              "yes"
            ]
          },
          {
            "criterion_id": "triggers-on-cost-question",
            "blocker": true,
            "method": "judge",
            "verdict": "yes",
            "samples": 5,
            "agreement": 1,
            "sample_verdicts": [
              "yes",
              "yes",
              "yes",
              "yes",
              "yes"
            ]
          },
          {
            "criterion_id": "right-sizing-grounded-and-hedged",
            "blocker": false,
            "method": "judge",
            "verdict": "no",
            "samples": 5,
            "agreement": 1,
            "sample_verdicts": [
              "no",
              "no",
              "no",
              "no",
              "no"
            ]
          },
          {
            "criterion_id": "checks-metric-access-upfront",
            "blocker": false,
            "method": "judge",
            "verdict": "yes",
            "samples": 5,
            "agreement": 1,
            "sample_verdicts": [
              "yes",
              "yes",
              "yes",
              "yes",
              "yes"
            ]
          }
        ],
        "stability": {
          "min_blocker_agreement": 0.8,
          "unstable_blocker_failures": 0
        }
      },
      "failure_mode": "2 blocker criteria failed",
      "replay_fidelity_level": "RF-1"
    }
  }
]
