{
  "schema": "anvil-serving.benchmark-decision-summary/v1",
  "campaign_id": "2026-09-24-thinkingcap-context",
  "complete": true,
  "updated_at": "2026-09-24T07:29:30.048956+00:00",
  "identity": {
    "schema": "anvil-serving.qualification-identity/v1",
    "model": "bottlecapai/ThinkingCap-Qwen3.8-27B-NVFP4A4-AWQ",
    "model_revision": "f8fe157f207a13f977bc3d620ce10a3e9ba5ab11",
    "served_model": "thinkingcap-awq-5090",
    "runtime_image": "vllm/vllm-openai:v0.29.0@sha256:082ca6f035279109041ffd3fe0695cb568b29bc580b35c4f297a66a08b216c1b",
    "engine_revision": "98dff2a81d747d1dba01a47f939f48c3526d4206",
    "recipe_semantic_digest": "c81bf6a3c79e6062c99258243d50a1c6feaa0886c8ff6724f79b2cca8f46f026",
    "private_registry_sha256": "98d19789e52066c5fef7f8886f3752b31a2cd83202818f9268531ec08728cddd",
    "public_recipe": "configurations/thinkingcap-awq-vllm029-5090-128k-mtp3-bf16-uva3.toml",
    "public_recipe_sha256": "98d19789e52066c5fef7f8886f3752b31a2cd83202818f9268531ec08728cddd",
    "hardware": {
      "gpu": "NVIDIA GeForce RTX 5090",
      "reported_memory_mib": 32607,
      "count": 1,
      "architecture": "sm_120",
      "driver": "616.92",
      "host": "Operator Node",
      "runtime": "Windows / Docker Desktop Linux / WSL2"
    },
    "measurement_path": "direct loopback online endpoint; warm process after managed startup; prefix cache disabled",
    "context_tokens": 131072,
    "concurrency": 1,
    "attention": "TRITON_ATTN",
    "kv_dtype": "bfloat16",
    "ssm_dtype": "model-default float32",
    "speculation": {
      "method": "mtp",
      "num_speculative_tokens": 3,
      "attention_backend": "TRITON_ATTN"
    },
    "vision": {
      "maximum_images": 2,
      "maximum_pixels_per_image": 1048576,
      "video": 0,
      "weight_offload_backend": "uva",
      "weight_offload_parameter_selector": null,
      "weight_offload_cap_gib": 3,
      "observed_offloaded_gib": 3.02,
      "observed_visual_offload_gib": 0.86,
      "scope": "Generic offload includes visual and language parameters; vision remains enabled"
    },
    "cache_completeness": "The exact complete local checkpoint was verified by managed load. This branch actively uses MTP3; vision and MTP weights remain in the cached snapshot.",
    "source_evidence": [
      "native/mtp-uva3/load-preview.json",
      "native/mtp-uva3/startup.log",
      "native/mtp-uva3/health.json",
      "operational-source-check.json"
    ],
    "promotion_authorized": false,
    "branch": "128k-mtp3-generic-uva3",
    "qualification_state": "All declared model gates passed: functional 27/27, thinking controls 9/9, vision 18/18, repeated quality 30/30, strict capacity 100/100, context 60/60 all stop, and synthetic agentic 2/2. Final identity and HTTP health verified; direct candidate retained.",
    "ending_evidence": "restoration.json"
  },
  "gates": [
    {
      "name": "functional",
      "status": "passed",
      "passed": 27,
      "attempts": 27
    },
    {
      "name": "vision",
      "status": "passed",
      "passed": 18,
      "attempts": 18
    },
    {
      "name": "reasoning",
      "status": "passed",
      "passed": 9,
      "attempts": 9
    },
    {
      "name": "quality-scout",
      "status": "passed",
      "passed": 10,
      "attempts": 10
    },
    {
      "name": "quality-30",
      "status": "passed",
      "passed": 30,
      "attempts": 30
    },
    {
      "name": "capacity-100",
      "status": "passed",
      "passed": 100,
      "attempts": 100
    },
    {
      "name": "context-60",
      "status": "passed",
      "passed": 60,
      "attempts": 60,
      "completed": 60,
      "normal_stops": 60
    },
    {
      "name": "agentic-2",
      "status": "passed",
      "passed": 2,
      "attempts": 2,
      "turns": 6,
      "tool_calls": 4
    }
  ],
  "failures": [
    "GPU-only fraction .95 failed CUDA reserve check before weight loading; no KV-fit conclusion.",
    "No-MTP visual-UVA1 branch failed repeated quality at 27/30 despite normal stops."
  ],
  "limitations": [
    "Diagnostic quality suite contains only ten repeated questions",
    "Text context and small-image vision are separate gates; no combined near-limit image proof",
    "No C2, SWE, video, or long-duration soak",
    "Earlier unrelated Windows Pi repository-test failure remains deferred",
    "Exact-total and one-over-limit admission were not tested",
    "Buffered five-second GPU telemetry covers full context but not the later brief agentic gate; final GPU observation retained"
  ],
  "decision": {
    "evidence_labels": [
      "functional",
      "capacity",
      "quality"
    ],
    "decision_labels": [
      "challenger",
      "no-promotion"
    ],
    "promotion_authorized": false,
    "next_gate": "None for the declared direct candidate; independent review accepted retention. Future route/client promotion remains a separate gate.",
    "retained_context_tokens": 131072
  },
  "evidence": [
    "identity-mtp-uva3.json",
    "calibration-mtp-uva3.json",
    "identity-nospec.json",
    "run-plan.json",
    "context-plan-review.json",
    "quality-summary.json",
    "capacity-summary.json",
    "benchmark-matrix.svg",
    "context-full-summary.json",
    "agentic-summary.json",
    "final-resource-summary.json",
    "restoration.json",
    "operational-source-final-check.json",
    "final-review.json",
    "validation-final.json"
  ],
  "failed_branches": [
    {
      "id": "gpu-only-095",
      "failure": "CUDA startup reserve before weight loading",
      "evidence": "native/failed-gpu-reserve/startup.log"
    },
    {
      "id": "nospec-visual-uva1",
      "failure": "Repeated quality27/30; wrong FINAL=D for CS all3, normal stop. Functional/vision/reasoning/contextscout passed. Broader tests pruned.",
      "evidence": "native/candidate/quality.json"
    }
  ],
  "independent_review": "final-review.json",
  "validation": "validation-final.json"
}