{
  "candidates": {
    "reasoning_v1": {
      "checkpoint": "checkpoints/reasoning-v1-decisions",
      "audit_sha256": "795fa56a751bedf6175f9d9b8608ce28b16aea63c946442b3d714f5990dca491",
      "accepted": false,
      "checks": {
        "overall_accuracy": {
          "actual": 0.7096153846153846,
          "minimum": 0.7,
          "passed": true
        },
        "semantic_accuracy": {
          "actual": 0.9965277777777778,
          "minimum": 0.95,
          "passed": true
        },
        "balanced:data/semantic_v1/train.jsonl:action": {
          "actual": 1.0,
          "minimum": 0.6,
          "passed": true
        },
        "balanced:data/semantic_v1/train.jsonl:eligible": {
          "actual": 1.0,
          "minimum": 0.6,
          "passed": true
        },
        "balanced:data/semantic_v1/train.jsonl:permission": {
          "actual": 0.9895833333333333,
          "minimum": 0.6,
          "passed": true
        },
        "balanced:data/generated/judgment_v1/cases.jsonl:action": {
          "actual": 0.40277777777777785,
          "minimum": 0.6,
          "passed": false
        },
        "balanced:data/generated/judgment_v1/cases.jsonl:evidence_quality": {
          "actual": 0.5,
          "minimum": 0.6,
          "passed": false
        },
        "balanced:data/generated/judgment_v1/cases.jsonl:matches_requirement": {
          "actual": 0.5,
          "minimum": 0.6,
          "passed": false
        },
        "balanced:data/generated/support_v1/cases.jsonl:department": {
          "actual": 0.6,
          "minimum": 0.6,
          "passed": true
        },
        "balanced:data/generated/support_v1/cases.jsonl:action": {
          "actual": 0.7791666666666667,
          "minimum": 0.6,
          "passed": true
        }
      }
    },
    "reasoning_v1_balanced": {
      "checkpoint": "checkpoints/reasoning-v1-balanced-decisions",
      "audit_sha256": "f7dadd3cf4315ac486c736940fb3f6d93030e8bc9e83841358f5a9d3e60e2e77",
      "accepted": false,
      "checks": {
        "overall_accuracy": {
          "actual": 0.6392307692307693,
          "minimum": 0.7,
          "passed": false
        },
        "semantic_accuracy": {
          "actual": 0.9131944444444444,
          "minimum": 0.95,
          "passed": false
        },
        "balanced:data/semantic_v1/train.jsonl:action": {
          "actual": 0.90625,
          "minimum": 0.6,
          "passed": true
        },
        "balanced:data/semantic_v1/train.jsonl:eligible": {
          "actual": 0.9166666666666666,
          "minimum": 0.6,
          "passed": true
        },
        "balanced:data/semantic_v1/train.jsonl:permission": {
          "actual": 0.9166666666666666,
          "minimum": 0.6,
          "passed": true
        },
        "balanced:data/generated/judgment_v1/cases.jsonl:action": {
          "actual": 0.44246031746031744,
          "minimum": 0.6,
          "passed": false
        },
        "balanced:data/generated/judgment_v1/cases.jsonl:evidence_quality": {
          "actual": 0.7023809523809523,
          "minimum": 0.6,
          "passed": true
        },
        "balanced:data/generated/judgment_v1/cases.jsonl:matches_requirement": {
          "actual": 0.5952380952380952,
          "minimum": 0.6,
          "passed": false
        },
        "balanced:data/generated/support_v1/cases.jsonl:department": {
          "actual": 0.79375,
          "minimum": 0.6,
          "passed": true
        },
        "balanced:data/generated/support_v1/cases.jsonl:action": {
          "actual": 0.7708333333333334,
          "minimum": 0.6,
          "passed": true
        }
      }
    }
  },
  "followup_test_opened": false,
  "additional_paid_calls": 0,
  "default_checkpoint_unchanged": "checkpoints/curriculum-v3",
  "limits": "Existing default is validated only on the narrow semantic task; neither broader candidate passes the revised development gates."
}