{
  "schema": "anvil-serving.benchmark-decision-summary/v1",
  "campaign_id": "2026-10-16-glm-v091-trial",
  "complete": true,
  "publication_review_pending": false,
  "identity": {
    "candidate": "glm53-v091-tp2-327k-c4-vision8-nospec-r2",
    "model": "brandonmusic/GLM-5.3-Flash-tr3-4bpw",
    "model_revision": "a5fee929cf4888b1824323e33e8a19b60129e025",
    "runtime": "v0.9.1",
    "image_digest": "sha256:a9e985c09885710601b32ac86e81256aa1279b542caa345233ce331a816d7e07",
    "source_revision": "4c8b73ed0e09c70f3b4015a675cc243c4d86949a",
    "hardware": "two RTX PRO 6000 Blackwell Max-Q GPUs",
    "quantization": "EXL3 4 bpw",
    "kv_dtype": "FP8 DS-MLA",
    "configured_context_tokens": 327680,
    "configured_concurrency": 4,
    "gpu_memory_utilization": 0.94
  },
  "workload": {
    "reasoning_mode": "default with max effort",
    "oracle": "three strict 18-case rounds",
    "coding": "pinned five-case SWE suite; four resolved cases required",
    "vision": "18 image cases",
    "context": "retrieval and tool-context probes"
  },
  "gates": [
    {
      "name": "native preflight",
      "status": "passed",
      "result": "6/6"
    },
    {
      "name": "Oracle",
      "status": "passed",
      "result": "54/54"
    },
    {
      "name": "vision",
      "status": "passed",
      "result": "18/18"
    },
    {
      "name": "context",
      "status": "passed",
      "actual_prompt_tokens": [
        216307,
        110760
      ]
    },
    {
      "name": "GPU reserve",
      "status": "passed",
      "minimum_free_mib_per_gpu": [
        4789,
        5583
      ],
      "required_free_mib": 4096,
      "sampling_ms": 200,
      "oom_events": 0
    },
    {
      "name": "coding",
      "status": "failed",
      "resolved": 3,
      "attempted": 5,
      "required_resolved": 4
    }
  ],
  "not_run": [
    "capacity and stability: blocked by the failed coding gate",
    "DFlash trial: blocked by the failed coding gate",
    "performance comparison: no eligible matched population"
  ],
  "decision": {
    "evidence_labels": [
      "functional",
      "quality"
    ],
    "decision_labels": [
      "rejected",
      "no-promotion"
    ],
    "promotion_authorized": false,
    "reason": "The exact R2 profile did not meet the declared coding replacement gate. This is not a model-family or runtime defect claim."
  },
  "matched_incumbent_control": {
    "preflight": "direct 6/6; routed 6/6",
    "oracle": "54/54",
    "vision": "18/18",
    "context_actual_prompt_tokens": [
      216307,
      110760
    ],
    "coding": {
      "resolved": 3,
      "attempted": 5,
      "required_resolved": 4,
      "same_resolved_instances_as_r2": true
    },
    "reserve": {
      "minimum_free_mib_per_gpu": [
        1725,
        1745
      ],
      "historical_waiver": true,
      "challenger_gate_assessment": "not-applicable",
      "oom_events": 0
    }
  },
  "limitations": [
    "The configured 327680-token C4 profile is not a simultaneous full-window proof.",
    "200 ms sampling cannot establish behavior between samples.",
    "The five-case coding suite is a bounded quality sample, not a general model-family verdict.",
    "Matching resolved counts do not establish equivalent grader failure mechanisms, causal behavior, or performance."
  ],
  "coding_failure_detail": {
    "django": "Both fail Unicode JSON display",
    "pylint": {
      "candidate_failed": 2,
      "candidate_passed": 18,
      "original_failed": 1,
      "original_passed": 19,
      "scope": "One exact suite execution per profile; no causal or general quality claim"
    }
  }
}