{
  "schema": "anvil-serving.benchmark-run-plan/v1",
  "scope": "bounded mixed35 candidate scouting and portal repair",
  "usage_floor_remaining_percent": 10,
  "stop_new_experiments_remaining_percent": 15,
  "priority": [
    "correct coding and tools",
    "usable context",
    "completed-work latency"
  ],
  "baseline": "GLM4bpw FP8KV C4 no-spec327680",
  "candidate": "mixed35 NVFP4KV C1 no-spec327680; MTP only if correctness passes",
  "gates": [
    "exact identity",
    "short coding JSON tools",
    "250K actual prompt plus16384 output reserve",
    "same small independent coding fixtures",
    "matched no-spec/MTP capacity only if prior gates pass"
  ],
  "restoration": "Managed unload/recreate of exact baseline; direct and authenticated routed preflight; unchanged configuration hashes.",
  "authority": "User September 17: Merge the pr and test, following explicit downtime approval request; temporary dual-GPU turnover and exact restoration authorized. No promotion.",
  "research_coverage": "lead plus one independent scout; reduced panel to preserve usage reserve"
}