{
  "schema": "anvil-serving.benchmark-decision-summary/v1",
  "campaign_id": "2026-09-23-thinkingcap",
  "complete": true,
  "updated_at": "2026-09-24T03:37:44.165805+00:00",
  "identity": {
    "model": "bottlecapai/ThinkingCap-Qwen3.8-27B-NVFP4A4-AWQ",
    "revision": "f8fe157f207a13f977bc3d620ce10a3e9ba5ab11",
    "runtime_image": "sha256:082ca6f035279109041ffd3fe0695cb568b29bc580b35c4f297a66a08b216c1b",
    "engine": "vLLM0.29.0",
    "engine_revision": "98dff2a81d747d1dba01a47f939f48c3526d4206",
    "weights": "NVFP4 W4A4 AWQ mixed FP8/BF16; complete BF16 vision/MTP retained",
    "hardware": "one RTX5090,32607MiB,sm120,driver616.92",
    "host_runtime": "Windows with Docker Desktop Linux/WSL2",
    "selection": "explicit user instruction"
  },
  "workload": {
    "plan": "run-plan.md",
    "manifest": "workload-manifest-5090.json",
    "comparison": "100 strict short-output requests per configuration; unique cache and canaries",
    "quality": "Pinned ten-question MMLU-Pro diagnostic,three repetitions,deterministic temperature0,2048 visible+8192 reasoning accounting cap10240",
    "vision": "Six synthetic image cases,three repetitions,C1,two-image admission"
  },
  "measurement": {
    "path": "direct online loopback endpoint",
    "warm_cold_state": "warm serving process; first measured request retained; prefix caching disabled; no startup-time performance claim",
    "statistics": [
      "median and native per-request percentile distributions",
      "n=100 p99 is one observed tail order statistic,not a stable service SLA"
    ],
    "operational_source": "operational-source-hashes.json"
  },
  "gates": [
    {
      "name": "complete local snapshot",
      "status": "passed",
      "files": 26,
      "bytes": 23447947228
    },
    {
      "name": "selected 32K Triton BF16 functional",
      "status": "passed",
      "observations": 27
    },
    {
      "name": "selected 32K Triton BF16 thinking control",
      "status": "passed",
      "observations": 8
    },
    {
      "name": "selected 32K Triton BF16 image corpus",
      "status": "passed",
      "observations": 18
    },
    {
      "name": "selected 32K Triton BF16 strict capacity",
      "status": "passed",
      "observations": 100
    },
    {
      "name": "selected 32K Triton BF16 MMLU diagnostic",
      "status": "passed",
      "attempts": 30,
      "items": 10
    },
    {
      "name": "long context",
      "status": "passed",
      "observations": 60,
      "evidence": "context-result-summary.json"
    },
    {
      "name": "bounded synthetic agentic",
      "status": "passed",
      "cases": 2,
      "evidence": "agentic-result-summary.json"
    },
    {
      "name": "ending direct serve health",
      "status": "passed",
      "evidence": "restoration.json"
    },
    {
      "name": "full repository Python suite",
      "status": "failed",
      "passed": 7923,
      "failed": 1,
      "skipped": 417,
      "scope": "unrelated Windows Pi smoke, deferred with ticket"
    }
  ],
  "headline_results": [
    {
      "configuration": "eager",
      "evidence": "native/candidate/eager/eager-short32-n100.json",
      "attempted": 100,
      "completed": 100,
      "failed": 0,
      "aggregate_eligible": true,
      "median_decode_tokens_per_second": 15.673049361660688,
      "median_ttft_ms": 482.38480000873096,
      "median_e2e_ms": 3290.0751999986824,
      "aggregate_tokens_per_second": 13.630349996162131,
      "claim_boundary": "Nominal4K input, strict32-word output with canary, C1, unique cache. Failed population is excluded from clean comparison; valid-request medians remain descriptive."
    },
    {
      "configuration": "graphs",
      "evidence": "native/candidate/graphs/graph-short32-n100-r1.json",
      "attempted": 100,
      "completed": 100,
      "failed": 0,
      "aggregate_eligible": true,
      "median_decode_tokens_per_second": 67.86759402722755,
      "median_ttft_ms": 397.1048000094015,
      "median_e2e_ms": 1045.5440999940038,
      "aggregate_tokens_per_second": 42.96751096543126,
      "claim_boundary": "Nominal4K input, strict32-word output with canary, C1, unique cache. Failed population is excluded from clean comparison; valid-request medians remain descriptive."
    },
    {
      "configuration": "mtp_flash",
      "evidence": "native/candidate/mtp3/mtp3-short32-n100-r1.json",
      "attempted": 100,
      "completed": 99,
      "failed": 1,
      "aggregate_eligible": false,
      "median_decode_tokens_per_second": 152.92534012186695,
      "median_ttft_ms": 433.57300001662225,
      "median_e2e_ms": 721.3507000124082,
      "aggregate_tokens_per_second": null,
      "claim_boundary": "Nominal4K input, strict32-word output with canary, C1, unique cache. Failed population is excluded from clean comparison; valid-request medians remain descriptive."
    },
    {
      "configuration": "mtp_triton_bf16",
      "evidence": "native/candidate/triton-bf16/triton-bf16-short32-n100-r1.json",
      "attempted": 100,
      "completed": 100,
      "failed": 0,
      "aggregate_eligible": true,
      "median_decode_tokens_per_second": 158.49590269115166,
      "median_ttft_ms": 453.6534999788273,
      "median_e2e_ms": 731.3178999756929,
      "aggregate_tokens_per_second": 61.419572061303256,
      "claim_boundary": "Nominal4K input, strict32-word output with canary, C1, unique cache. Failed population is excluded from clean comparison; valid-request medians remain descriptive."
    },
    {
      "configuration": "mtp_triton_fp8",
      "evidence": "native/candidate/triton-fp8/triton-fp8-short32-n100-r1.json",
      "attempted": 100,
      "completed": 100,
      "failed": 0,
      "aggregate_eligible": true,
      "median_decode_tokens_per_second": 137.99137867361082,
      "median_ttft_ms": 494.9528999859467,
      "aggregate_tokens_per_second": 55.17134253052424,
      "quality_gate_passed": false,
      "claim_boundary": "Capacity protocol passed but recommendation rejected:27/30 quality,3 reasoning-budget exhaustions. Diagnostic only."
    },
    {
      "configuration": "selected_mtp_triton_bf16_32k",
      "evidence": "native/candidate/triton-bf16-32k/triton-bf16-32k-short32-n100-r1.json",
      "configured_max_context": 32768,
      "attempted": 100,
      "completed": 100,
      "failed": 0,
      "aggregate_eligible": true,
      "median_decode_tokens_per_second": 158.3551791480093,
      "median_ttft_ms": 455.44210000662133,
      "median_e2e_ms": 733.4115999983624,
      "aggregate_tokens_per_second": 61.14585265214693,
      "quality_gate_passed": true,
      "claim_boundary": "Nominal4K input, actual3609-3687 tokens; 32 code words plus unique canary,45 completion tokens,C1,thinking disabled. Controlled short output does not estimate general-chat performance."
    }
  ],
  "failures": [
    "Initial WSL UVA startup failure resolved with explicit pin-memory recipe",
    "Incumbent and eager256-word scouts failed exact-output contracts",
    "Initial eager preflight duplicate overwrote first artifact; retained preflight is functional-only",
    "Graph MMLU27/30: computer-science budget exhaustion; Flash-MTP MMLU27/30: computer-science wrong answer",
    "Flash-MTP strict100 cell failed one request(34 versus32 words) and is aggregate-ineligible",
    "Triton FP8 quality27/30: computer-science reasoning budget exhausted in all3 repeats; not advanced to larger context"
  ],
  "limitations": [
    "This is a bounded local campaign,not full public benchmark scores or a universal model ranking",
    "No matched BF16-weight reference or direct quantization-quality comparison",
    "Quality runner fixes temperature0; publisher sampling protocol is not reproduced",
    "No SWE isolated worker,video evaluation or long-duration endurance claim",
    "Native capacity distributions describe short controlled output,not general-chat throughput",
    "32K is the configured limit; actual near-limit inputs tested29578-29629 tokens with2048 total completion cap. Larger context and C2 remain unmeasured.",
    "No matched Triton no-speculation control; whole-recipe comparison is not isolated speculative-decoding speedup.",
    "Required full repository Python gate failed in unrelated Windows Pi smoke; source fix deferred, no clean full-suite claim.",
    "Selected-profile minimum sampled GPU headroom3342MiB; initial4GiB reserve assumption was not maintained. Five-second samples may miss peaks."
  ],
  "decision": {
    "evidence_labels": [
      "compatibility-only",
      "functional",
      "capacity",
      "quality"
    ],
    "decision_labels": [
      "challenger",
      "no-promotion"
    ],
    "promotion_authorized": false,
    "best_completed_configuration": "32K,C1,explicit TRITON_ATTN,MTP3,BF16 KV; two-image admission",
    "next_gate": null,
    "disposition": "selected direct candidate retained under explicit user authorization; no route/client alias changes"
  },
  "evidence": [
    "triton-bf16-result-summary.json",
    "mtp3-result-summary.json",
    "baseline-summary.json",
    "benchmark-graph-data.json",
    "restoration.json",
    "triton-fp8-result-summary.json",
    "cache-capacity-boundary.json",
    "triton-bf16-32k-result-summary.json",
    "context-result-summary.json",
    "agentic-result-summary.json",
    "telemetry-summary.json",
    "validation-final.json",
    "final-independent-review.json"
  ],
  "model_work_complete": true,
  "context_results": {
    "passed": 60,
    "completed": 60,
    "finish_reason_counts": {
      "stop": 60
    },
    "actual_prompt_token_ranges": {
      "8192": {
        "minimum": 8198,
        "maximum": 8229,
        "samples": 30
      },
      "29696": {
        "minimum": 29578,
        "maximum": 29629,
        "samples": 30
      }
    }
  },
  "agentic_results": {
    "synthetic_cases": 2,
    "passed_cases": 2,
    "api_turns": 6,
    "tool_calls": 4,
    "finish_reason_counts": {
      "tool_calls": 4,
      "stop": 2
    },
    "reasoning_disabled_observed_zero": true,
    "safe_error_retry": "passed"
  },
  "completed_at": "2026-09-24T03:50:10.093850+00:00",
  "completion_scope": "Bounded RTX5090 direct model campaign; full repository gate failed and deferred separately"
}