{
  "schema": "anvil-serving.benchmark-decision-summary/v1",
  "campaign_id": "2026-10-13-full-qualification",
  "complete": true,
  "identity": {
    "hardware": "dual RTX PRO 6000 Blackwell Max-Q",
    "topology": "TP1 and TP2 public-native lanes only"
  },
  "workload": {
    "quality": "five fixed SWE-bench Verified tasks with an independent official grader",
    "capacity": "controlled output with request-canary validation"
  },
  "measurement": {
    "path": "direct managed qualification",
    "warm_cold_state": "preserved per native artifact",
    "sample_count": 3,
    "statistics": [
      "actual prompt-token means; no stable p99 claim below 100 samples"
    ]
  },
  "gates": [
    "complete fixed-five grading",
    "at least four resolved tasks",
    "Django and Pylint sentinels resolved"
  ],
  "headline_results": [
    {
      "name": "GLM baseline SWE",
      "value": "2/5 resolved with complete grading"
    },
    {
      "name": "baseline prompt means",
      "value": "25280.5, 25281.125, 106520, 216194.67"
    },
    {
      "name": "Ignis terminal SWE",
      "value": "2/5 resolved; 2/5 graded; 3 empty"
    }
  ],
  "failures": [
    "No challenger met the complete fixed-five floor and sentinel requirement.",
    "Ignis C8 protocol preflight passed 6/13 with seven immediate 503 refusals; serial C1 passed 6/6. Separate SWE engine-full retries recovered 24 times; JSON_OBJECT is unsupported.",
    "Qwen top-k-20 was cancelled before load/query; no improvement was measured."
  ],
  "limitations": [
    "No new full-context C4 or performance ranking claim is made.",
    "R19 accepts restoration evidence but preserves the routed thinking-off failure and does not prove parser flattening without a failed-request wire capture."
  ],
  "decision": {
    "evidence_labels": [
      "functional",
      "capacity",
      "quality",
      "compatibility-only"
    ],
    "decision_labels": [
      "no-promotion",
      "rollback"
    ],
    "promotion_authorized": false,
    "next_gate": "resolve the routed thinking-off evaluation gap before any later configuration decision"
  },
  "evidence": [
    "README.md",
    "native/",
    "baseline-chart/baseline-capacity-graph-data.json"
  ],
  "restoration": {
    "status": "accepted-with-thinking-off-gap",
    "direct": "13/13",
    "openclaw_high": "9/9",
    "openclaw_off": "two failures with reasoning visible",
    "openclaw_max": "local rejection before query",
    "hindsight": "recall and retain passed",
    "no_promotion": true
  },
  "cleanup": {
    "status": "complete",
    "removed_model_cache_snapshots": 11,
    "removed_docker_images": 106,
    "remaining_safely_removable_images": 0,
    "evidence": "cleanup-summary.json",
    "accounting_caveat": "Logical repository bytes and native allocated-plan bytes are reported separately; neither is a model-quality measurement."
  }
}