{
  "schema": "anvil-serving.benchmark-decision-summary/v1",
  "campaign_id": "2026-09-23-glm-runtime-stability",
  "complete": true,
  "identity": {
    "hardware": "2x RTX PRO 6000 Blackwell Max-Q 96 GB, PCIe, native Linux",
    "driver": "615.71.09",
    "weights_revision": "a5fee929cf4888b1824323e33e8a19b60129e025",
    "image_digest": "sha256:0f1cdcc8891f1cc3a444121eb61d366289a1cbba285f0892dcbb24bc94961692",
    "runner_revision": "fdc53d1f834b06a8ca75324e3b84407030a91506",
    "source_attestation": "Partial; downstream kpool patch absent from the imported public Git revision."
  },
  "measurement": {
    "path": "online direct; routed restoration acceptance",
    "warm_cold_state": "warm weights and compiled cache; fresh run/round prefixes; cache metrics retained separately",
    "performance_eligible": false
  },
  "profiles": {
    "dcp2": {
      "large_serial_passed": 1,
      "large_serial_attempted": 1,
      "large_overlap_passed": 0,
      "large_overlap_attempted": 1,
      "measured_startup_kv_tokens": 1416244,
      "failure": "CUDA illegal memory access, EngineDeadError, two Xid31 virtual-write PDE faults"
    },
    "dcp1": {
      "large_overlap_passed": 3,
      "large_overlap_attempted": 3,
      "measured_startup_kv_tokens": 825268,
      "configured_context": 327680,
      "configured_max_sequences": 4,
      "measured_large_concurrency": 2,
      "single_compute_delta": "decode_context_parallel_size: 2 to 1"
    }
  },
  "request_policy_json": {
    "disabled_passed": 0,
    "enabled_passed": 3,
    "attempted_per_arm": 3,
    "limits": "One repeated deterministic JSON fixture; separate enabled full-preflight pass excluded from these counts; no CUDA attribution."
  },
  "trial_matrix": [
    {
      "artifact": "baseline-small.json",
      "round": 0,
      "status": "completed",
      "mode": "overlap",
      "prefix_mode": "unique",
      "runtime_passed": true,
      "coverage_passed": true,
      "retrieval_passed": true,
      "actual_prompt_tokens": {
        "anchor": 4089,
        "contender": 1017
      },
      "actual_output_tokens": {
        "anchor": 2048,
        "contender": 145
      },
      "output_caps": {
        "anchor": 2048,
        "contender": 1024
      }
    },
    {
      "artifact": "baseline-175k-serial.json",
      "round": 0,
      "status": "completed",
      "mode": "serial",
      "prefix_mode": "unique",
      "runtime_passed": true,
      "coverage_passed": true,
      "retrieval_passed": true,
      "actual_prompt_tokens": {
        "anchor": 174798,
        "contender": 31820
      },
      "actual_output_tokens": {
        "anchor": 4096,
        "contender": 220
      },
      "output_caps": {
        "anchor": 4096,
        "contender": 1024
      }
    },
    {
      "artifact": "baseline-175k-overlap.json",
      "round": 0,
      "status": "incomplete",
      "mode": "overlap",
      "prefix_mode": "unique",
      "runtime_passed": false,
      "coverage_passed": false,
      "retrieval_passed": false,
      "actual_prompt_tokens": {
        "anchor": 174798,
        "contender": 31816
      },
      "actual_output_tokens": {
        "anchor": null,
        "contender": null
      },
      "output_caps": {
        "anchor": 4096,
        "contender": 1024
      }
    },
    {
      "artifact": "dcp1-small.json",
      "round": 0,
      "status": "incomplete",
      "mode": "overlap",
      "prefix_mode": "unique",
      "runtime_passed": false,
      "coverage_passed": false,
      "retrieval_passed": false,
      "actual_prompt_tokens": {
        "anchor": 4092,
        "contender": 1021
      },
      "actual_output_tokens": {
        "anchor": null,
        "contender": 161
      },
      "output_caps": {
        "anchor": 512,
        "contender": 512
      }
    },
    {
      "artifact": "dcp1-small-matched.json",
      "round": 0,
      "status": "completed",
      "mode": "overlap",
      "prefix_mode": "unique",
      "runtime_passed": true,
      "coverage_passed": true,
      "retrieval_passed": true,
      "actual_prompt_tokens": {
        "anchor": 4088,
        "contender": 1016
      },
      "actual_output_tokens": {
        "anchor": 2048,
        "contender": 133
      },
      "output_caps": {
        "anchor": 2048,
        "contender": 1024
      }
    },
    {
      "artifact": "dcp1-175k-overlap.json",
      "round": 0,
      "status": "completed",
      "mode": "overlap",
      "prefix_mode": "unique",
      "runtime_passed": true,
      "coverage_passed": true,
      "retrieval_passed": true,
      "actual_prompt_tokens": {
        "anchor": 174800,
        "contender": 31815
      },
      "actual_output_tokens": {
        "anchor": 4096,
        "contender": 214
      },
      "output_caps": {
        "anchor": 4096,
        "contender": 1024
      }
    },
    {
      "artifact": "dcp1-175k-repeat2.json",
      "round": 0,
      "status": "completed",
      "mode": "overlap",
      "prefix_mode": "unique",
      "runtime_passed": true,
      "coverage_passed": true,
      "retrieval_passed": true,
      "actual_prompt_tokens": {
        "anchor": 174798,
        "contender": 31822
      },
      "actual_output_tokens": {
        "anchor": 4096,
        "contender": 145
      },
      "output_caps": {
        "anchor": 4096,
        "contender": 1024
      }
    },
    {
      "artifact": "dcp1-175k-repeat2.json",
      "round": 1,
      "status": "completed",
      "mode": "overlap",
      "prefix_mode": "unique",
      "runtime_passed": true,
      "coverage_passed": true,
      "retrieval_passed": true,
      "actual_prompt_tokens": {
        "anchor": 174797,
        "contender": 31816
      },
      "actual_output_tokens": {
        "anchor": 4096,
        "contender": 153
      },
      "output_caps": {
        "anchor": 4096,
        "contender": 1024
      }
    }
  ],
  "decision": {
    "evidence_labels": [
      "functional"
    ],
    "decision_labels": [
      "no-promotion",
      "challenger"
    ],
    "promotion_authorized": false,
    "result": "Bounded DCP1 mitigation candidate; exact baseline restored and direct/routed acceptance passed.",
    "next_gate": "Repeat larger incident shape and cache-state exposure, independent quality/modality/client gates, and first-fault diagnosis before root-cause closure"
  },
  "limitations": [
    "Kernel first fault unresolved; no CUDA core dump, sanitizer or reduced kernel probe.",
    "No model reliability-rate or speed claim.",
    "201K request time guard refused before traffic; zero 201K generation requests.",
    "310K, full-window C4, repeated-prefix/cache-eviction/cancellation probes unrun.",
    "No fresh candidate vision/OCR, broad agentic/SWE, full-output coherence or real-client qualification.",
    "One mismatched 512-cap small failure retained; exact budget exhaustion unproved.",
    "Known-crashing parent reverse-order repetition omitted.",
    "Source attestation partial; successful scheduler batch not measured."
  ],
  "restoration": "restoration.json",
  "scope": "Diagnostic campaign closed at the time limit. Replacement qualification and hardware envelope remain incomplete."
}
