{
  "schema": "anvil-serving.benchmark-decision-summary/v1",
  "campaign_id": "2026-09-17-glm53-mixed35-fix-forward",
  "complete": true,
  "evidence_labels": [
    "functional",
    "quality",
    "capacity"
  ],
  "decision_labels": [
    "no-promotion"
  ],
  "promotion_authorized": true,
  "promotion_authorization_condition": "User authorized promotion if the candidate demonstrates better performance while passing required quality and operational gates.",
  "headline_results": [
    "Bounded loader retry loaded all 120 shards and initialized CUDA on both TP ranks.",
    "Synthetic 64 MiB cgroup probe contained a 256 MiB allocation with oom_kill=1.",
    "C1 default-thinking preflight passed smoke, structured JSON, and 3/3 shared-prefix tool requests.",
    "C1 intelligence scout passed two deterministic checks once each; it is below the three-repetition qualification threshold.",
    "C4 needle passed at 216307 actual input tokens with exact ZEBRA and stop; it is one C1 request, not C4 capacity.",
    "C4 preflight and three-repetition bounded intelligence/tool checks passed.",
    "C4 strict capacity passed 4/4 then 3/4; the partial cell is performance-ineligible.",
    "The baseline strict cell passed 2/4; no complete throughput comparison is available."
  ],
  "failed_diagnostic": "Disabled-thinking 256-token JSON control reached its length limit; default-thinking/max_tokens=16384 is the retained remedy.",
  "missing_gates": [
    "Representative repository tasks and repeated-session reliability",
    "Real-client acceptance",
    "Consistent repeated strict capacity and demonstrated overall improvement",
    "Full configured 327680-token window and simultaneous long-context capacity"
  ],
  "decision": "retain baseline; no-promotion",
  "boundary": "No alias, route, or serve promotion occurred; the runnable configuration is not a rejection of the mixed-quant family, and conditional authorization does not itself promote the candidate.",
  "restoration_status": "Both C1 and C4 trials restored the exact baseline; direct and routed checks passed."
}