{
  "schema": "anvil-serving.recipe-feasibility-screen/v1",
  "recorded_at": "2026-09-24T04:55:18.847858+00:00",
  "classification": "benchmark-survivor, not fit proof",
  "model_and_numerics": "same exact checkpoint/runtime; BF16 KV and default FP32 SSM; MTP3 restored",
  "context": 131072,
  "concurrency": 1,
  "memory_fraction": 0.93,
  "cpu_offload": "generic UVA3GiB; selector absent",
  "source_values": {
    "gpu_capacity_gib": 31.82,
    "measured_prior_32k_mtp_fraction": 0.87,
    "measured_prior_32k_mtp_kv_gib": 4.73,
    "estimated_fixed_residency_gib": 22.9534,
    "planned_budget_gib": 29.5926,
    "optimistic_pool_gib": 9.639200000000002,
    "reviewer_128k_mtp_pool_requirement_approx_gib": 9.13
  },
  "unknowns": [
    "128K activation/graph residency",
    "actual tensor-rounded offload",
    "packed-weight UVA kernel compatibility",
    "PCIe throughput cost"
  ],
  "authority": "Existing autonomous model optimization and managed candidate lifecycle request; no route/client promotion",
  "gates": [
    "native health and exact identity/cache >=131072",
    "sampled GPU free>=1024MiB and healthy host RAM",
    "functional and vision",
    "quality scout then full unchanged30-attempt suite",
    "strict capacity scout then100 if qualified",
    "context/agentic if all preceding gates pass"
  ],
  "stop_on": [
    "engine compatibility error",
    "insufficient cache capacity",
    "OOM/CUDA error/restart",
    "failed quality gate"
  ],
  "evidence": [
    "../2026-09-23-thinkingcap-evidence/native/candidate/triton-bf16-32k/triton-bf16-32k-startup.log",
    "context-plan-review.json",
    "native/candidate/quality.json"
  ]
}