{
  "campaign": "glm53-nvidia-fp4-ab-20261017",
  "authorized": "Controlled A/B using NVIDIA recipe with only slight modifications; no promotion",
  "model": "nvidia/GLM-5.3-Flash-FP4-Local",
  "revision": "76348ddb2be2e35e39ecfc00e75dc07703f4f3d5",
  "runtime_image": "ghcr.io/local-inference-lab/vllm:release-10152026@sha256:745feb916ee816f5409fd58386114be4f22fbede7f9c0e0a7b8c4fe8dac0fb29",
  "objectives": [
    "independent functional coding/tools/JSON/vision/context correctness",
    "matched natural-output C1/C4 latency and completion evidence",
    "preserve and restore exact original r11"
  ],
  "planned_stages": [
    "offline harness",
    "storage and bounded feasibility",
    "download pinned model",
    "baseline controls",
    "quiesce/drain exact primary",
    "retain stopped incumbent",
    "load isolated candidate",
    "direct gates",
    "matched timing and coding",
    "matched no-spec diagnostic if appropriate",
    "restore exact r11",
    "publish evidence"
  ],
  "budget": {
    "startup_seconds": 3600,
    "configuration_search_max_versions": 3,
    "candidate_live_hours": 3,
    "restoration_reserve_minutes": 70,
    "candidate_total_seconds": 10800,
    "restoration_reserve_seconds": 4200,
    "live_requests_deadline_seconds_after_candidate_load_start": 6600,
    "per_stage_timeout_seconds": 900,
    "abort_rule": "Stop dispatch on failed correctness/runtime gate or deadline. Cancel active durable job with eval benchmark {suite} cancel; unload candidate through recipes; restore exact incumbent via independently reviewed managed path; readmit only after health, identity, JSON and routed checks.",
    "lifecycle_timeout_seconds": 3600,
    "stage_cap_scope": "900 seconds applies only to benchmark/preflight stages; managed load has3600 seconds, plus bounded reconciliation after CLI timeout."
  },
  "candidate_modifications": [
    "immutable revision and image",
    "managed loopback endpoint and unique served name",
    "RAM 52 GiB, zero effective swap in r2, and 12 GiB host reserve; r1 1 MiB swap admission refused before container creation",
    "comparison context 327680 and maximum sequences 4",
    "GPU-only prefix cache using direct upstream command, no LMCache process"
  ],
  "promotion_authorized": false,
  "created_at": "2026-10-17T07:10:18.907649+00:00",
  "reserve_policy": {
    "default_mib_per_gpu": 4096,
    "effective": "Published .985 recipe; fixed KV bytes independently set",
    "scope": "provisional isolated startup and bounded scout only",
    "owner_waiver": true,
    "reason": "Dedicated model-only GPUs, user requested upstream recipe with minimal changes. Startup/workload per-device and cgroup telemetry required; does not prove safety or authorize promotion."
  },
  "restoration_contract": "Restart the exact retained original container through managed serves mode restore-retained, with full 10-manifest/32-serve scope, immutable ID/image/registry/manifest bindings, explicit native owner, and fresh original-source hashes. Preserve original image, entrypoint, flags, cgroup limits, UUID order, served identity, port and router config. Readmit only after direct checks.",
  "restoration_risk": "Retained restore has a 600-second readiness bound and stops/retains the same container on failure. Historical r11 startup took about 121 seconds. New-create replay was superseded because free host swap was below its 1 MiB allowance.",
  "matched_stability_workload": {
    "scenario": "baseline-stability-scout-r3.json",
    "anchor_tokens": 32000,
    "contender_tokens": 8192,
    "anchor_max_tokens": 16384,
    "contender_max_tokens": 8192,
    "reasoning_effort": "max",
    "temperature": 0.0,
    "top_p": 0.95,
    "rounds": 1,
    "purpose": "bounded overlap/retrieval diagnostic; not scheduler proof, endurance, or performance ranking"
  },
  "candidate_version": "r2",
  "version_history": [
    {
      "version": "r1",
      "result": "host swap admission refusal before creation"
    },
    {
      "version": "r2",
      "delta": "zero swap; served identity/version suffix",
      "result": "startup, preflight, vision and agentic passed; context failed17/18"
    }
  ],
  "stopping_result": {
    "gate": "candidate context",
    "passed": 17,
    "attempted": 18,
    "failed_case": "native-needle-318464-0.900-r0",
    "actual_input_tokens": 316249,
    "decision": "stop dependent overlap/timing; restore exact incumbent",
    "no_follow_on_requests": true
  }
}