{
  "conditions": "Heterogeneous runtime and evaluator versions and provider settings; historical grades unchanged. Technical and protocol exclusions are not quality failures. River recovery used serial pending cases. Descriptive comparison, not a controlled ranking.",
  "date": "2026-10-29",
  "evaluator": "Compliance-0.3.0",
  "models": [
    {
      "aborted": 0,
      "assigned": 44,
      "cost_usd": null,
      "endpoint": "responses",
      "excluded": 0,
      "goal_completed": 44,
      "model": "gpt-6.1-terra",
      "observation_count": 1,
      "provider": "openai",
      "reasoning": "none",
      "reliable": 38,
      "scored": 44,
      "wire_model": "gpt-6.1-terra"
    },
    {
      "aborted": 0,
      "assigned": 44,
      "cost_usd": null,
      "endpoint": "messages",
      "excluded": 0,
      "goal_completed": 0,
      "model": "claude-haiku-5-5",
      "observation_count": 1,
      "provider": "anthropic",
      "reasoning": "adaptive-low",
      "reliable": 0,
      "scored": 44,
      "wire_model": "claude-haiku-5-5"
    },
    {
      "aborted": 0,
      "assigned": 44,
      "cost_usd": null,
      "endpoint": "native-chat",
      "excluded": 20,
      "goal_completed": 0,
      "model": "mistral-large-4",
      "observation_count": 1,
      "provider": "mistral",
      "reasoning": "none",
      "reliable": 0,
      "scored": 24,
      "wire_model": "mistral-large-4"
    },
    {
      "aborted": 0,
      "assigned": 44,
      "cost_usd": null,
      "endpoint": "responses",
      "excluded": 29,
      "goal_completed": 15,
      "model": "grok-4.7",
      "observation_count": 1,
      "provider": "xai",
      "reasoning": "low",
      "reliable": 5,
      "scored": 15,
      "wire_model": "grok-4.7"
    },
    {
      "aborted": 0,
      "assigned": 44,
      "cost_usd": null,
      "endpoint": "http",
      "excluded": 8,
      "goal_completed": 32,
      "model": "GLM-5.3",
      "observation_count": 1,
      "provider": "river",
      "reasoning": "low",
      "reliable": 6,
      "scored": 36,
      "wire_model": "GLM-5.3"
    },
    {
      "aborted": 0,
      "assigned": 44,
      "cost_usd": null,
      "endpoint": "http",
      "excluded": 0,
      "goal_completed": 19,
      "model": "moonshotai/Kimi-K3",
      "observation_count": 1,
      "provider": "river",
      "reasoning": "none",
      "reliable": 0,
      "scored": 44,
      "wire_model": "moonshotai/Kimi-K3"
    }
  ],
  "provenance": {
    "correction_source_revision": "989c22d9c8342df26eb78eb77f798c9b245abf6e",
    "models": [
      {
        "attempt_groups": [
          {
            "basis": "admitted-attempt-contract",
            "count": 44,
            "phase": "original",
            "registry_mapping_sha256": "f6b4984dade59dec48fc5dbbeab500b235443a71cc1d1130d4efb37b07e74292",
            "source_revision": "f3bd45fb0d219929c9415b734dfeb9ef0cb6a5b8"
          }
        ],
        "distinct_attempted": 0,
        "effective_corrected": 0,
        "effective_original": 44,
        "effective_unresolved": 0,
        "model": "gpt-6.1-terra",
        "original_aborted": 0,
        "original_excluded": 0,
        "original_scored": 44,
        "technical_selected": 0,
        "terminal_attempts": 0,
        "total_attempts": 0
      },
      {
        "attempt_groups": [
          {
            "basis": "admitted-attempt-contract",
            "count": 35,
            "phase": "correction",
            "registry_mapping_sha256": "17c29ef901b74570bbd7ffb21225e38ed3f3310dcbd1a5cca2896c3b08d38492",
            "source_revision": "989c22d9c8342df26eb78eb77f798c9b245abf6e"
          },
          {
            "basis": "admitted-attempt-contract",
            "count": 44,
            "phase": "original",
            "registry_mapping_sha256": "de93ace0f2632cc1019f523a63bced3ec265a2e028b0c4d620c38d21681f9d95",
            "source_revision": "f3bd45fb0d219929c9415b734dfeb9ef0cb6a5b8"
          }
        ],
        "distinct_attempted": 31,
        "effective_corrected": 31,
        "effective_original": 13,
        "effective_unresolved": 0,
        "model": "claude-haiku-5-5",
        "original_aborted": 0,
        "original_excluded": 31,
        "original_scored": 13,
        "technical_selected": 31,
        "terminal_attempts": 35,
        "total_attempts": 35
      },
      {
        "attempt_groups": [
          {
            "basis": "admitted-attempt-contract",
            "count": 18,
            "phase": "correction",
            "registry_mapping_sha256": "1202bb6c39099ba0920e740b97f87058ac85b686001e62eead6af646b699e7cf",
            "source_revision": "989c22d9c8342df26eb78eb77f798c9b245abf6e"
          },
          {
            "basis": "admitted-attempt-contract",
            "count": 2,
            "phase": "correction",
            "registry_mapping_sha256": "1202bb6c39099ba0920e740b97f87058ac85b686001e62eead6af646b699e7cf",
            "source_revision": "f3bd45fb0d219929c9415b734dfeb9ef0cb6a5b8"
          },
          {
            "basis": "admitted-attempt-contract",
            "count": 44,
            "phase": "original",
            "registry_mapping_sha256": "1202bb6c39099ba0920e740b97f87058ac85b686001e62eead6af646b699e7cf",
            "source_revision": "f3bd45fb0d219929c9415b734dfeb9ef0cb6a5b8"
          }
        ],
        "distinct_attempted": 20,
        "effective_corrected": 0,
        "effective_original": 24,
        "effective_unresolved": 20,
        "model": "mistral-large-4",
        "original_aborted": 0,
        "original_excluded": 20,
        "original_scored": 24,
        "technical_selected": 20,
        "terminal_attempts": 20,
        "total_attempts": 20
      },
      {
        "attempt_groups": [
          {
            "basis": "admitted-attempt-contract",
            "count": 29,
            "phase": "correction",
            "registry_mapping_sha256": "db856d6a0550702c7b9874f03e6ba1c2d4b0959dc7cc0e818d3a869f63adef67",
            "source_revision": "f3bd45fb0d219929c9415b734dfeb9ef0cb6a5b8"
          },
          {
            "basis": "admitted-attempt-contract",
            "count": 44,
            "phase": "original",
            "registry_mapping_sha256": "db856d6a0550702c7b9874f03e6ba1c2d4b0959dc7cc0e818d3a869f63adef67",
            "source_revision": "f3bd45fb0d219929c9415b734dfeb9ef0cb6a5b8"
          }
        ],
        "distinct_attempted": 29,
        "effective_corrected": 0,
        "effective_original": 15,
        "effective_unresolved": 29,
        "model": "grok-4.7",
        "original_aborted": 0,
        "original_excluded": 29,
        "original_scored": 15,
        "technical_selected": 29,
        "terminal_attempts": 29,
        "total_attempts": 29
      },
      {
        "attempt_groups": [
          {
            "basis": "admitted-attempt-contract",
            "count": 15,
            "phase": "correction",
            "registry_mapping_sha256": "710024a6798780eff7c7e2c5f9551043067e319cac00f6dcb6fd13ff39415255",
            "source_revision": "989c22d9c8342df26eb78eb77f798c9b245abf6e"
          },
          {
            "basis": "admitted-attempt-contract",
            "count": 1,
            "phase": "correction",
            "registry_mapping_sha256": "131b62f069ee766e4c7f720f2136055aaea2cc31de2824a64287031c7ec5ed9c",
            "source_revision": "f3bd45fb0d219929c9415b734dfeb9ef0cb6a5b8"
          },
          {
            "basis": "admitted-attempt-contract",
            "count": 44,
            "phase": "original",
            "registry_mapping_sha256": "131b62f069ee766e4c7f720f2136055aaea2cc31de2824a64287031c7ec5ed9c",
            "source_revision": "f3bd45fb0d219929c9415b734dfeb9ef0cb6a5b8"
          }
        ],
        "distinct_attempted": 16,
        "effective_corrected": 8,
        "effective_original": 28,
        "effective_unresolved": 8,
        "model": "GLM-5.3",
        "original_aborted": 4,
        "original_excluded": 12,
        "original_scored": 28,
        "technical_selected": 16,
        "terminal_attempts": 16,
        "total_attempts": 16
      },
      {
        "attempt_groups": [
          {
            "basis": "admitted-attempt-contract",
            "count": 6,
            "phase": "correction",
            "registry_mapping_sha256": "1e14f12022400705096ab295b66433f7f7c856faa723c0c386ffb89a04333697",
            "source_revision": "989c22d9c8342df26eb78eb77f798c9b245abf6e"
          },
          {
            "basis": "admitted-attempt-contract",
            "count": 44,
            "phase": "original",
            "registry_mapping_sha256": "d6dbbfa4f96504b27bcfe74be1d8001343802b827e9d312a3821e4dc5ae1613c",
            "source_revision": "f3bd45fb0d219929c9415b734dfeb9ef0cb6a5b8"
          }
        ],
        "distinct_attempted": 6,
        "effective_corrected": 6,
        "effective_original": 38,
        "effective_unresolved": 0,
        "model": "moonshotai/Kimi-K3",
        "original_aborted": 0,
        "original_excluded": 6,
        "original_scored": 38,
        "technical_selected": 6,
        "terminal_attempts": 6,
        "total_attempts": 6
      }
    ],
    "original_source_revision": "f3bd45fb0d219929c9415b734dfeb9ef0cb6a5b8"
  },
  "schema_version": 2,
  "series": "2026-10-29 owner-run selected six",
  "source_revision": "989c22d9c8342df26eb78eb77f798c9b245abf6e"
}