{
  "schema": "anvil-serving-vram-policy-reclassification/v1",
  "observed_at": "2026-09-02",
  "recipe": "glm53-flash-ormandj-sglang-sm120-tp2-393k-c1-adaptive-mtp",
  "configured_context_tokens": 393216,
  "concurrency": 1,
  "default_reserve_mib_per_gpu": 3072,
  "effective_reserve_mib_per_gpu": 0,
  "waiver": true,
  "waiver_scope": "this profile on the model-only dual-RTX-PRO exclusive pair",
  "waiver_rationale": "the two GPUs host model workloads only; the operator accepted a zero standing reserve while retaining runtime safety gates",
  "qualification_post_workload_free_mib_per_gpu": [2101, 2101],
  "post_promotion_client_workload_free_mib_per_gpu": [2543, 2543],
  "runtime_safety_gates_retained": [
    "functional",
    "capacity",
    "quality",
    "post-workload telemetry",
    "oom",
    "crash",
    "cuda-error",
    "restart",
    "shared-memory residue",
    "rollback"
  ],
  "previous_status": "policy-infeasible",
  "new_status": "verified",
  "promotion_decision": "human-approved-current",
  "finding": "../2026-09-02-glm53-sglang-sm120-393k-promotion.md"
}