{
  "caveats": [
    "Finite lab cohort; configured 262144 context differs from measured inputs. Aggregate uses first-output to final-completion window. Whole-host memory sampled during speed only; missing counters remain unassessed. Broader coding quality and delivery acceptance unassessed."
  ],
  "engine": {
    "name": "NVIDIA NGC vLLM",
    "recipe": "plain",
    "version": "0.29.0+5013de39.dev"
  },
  "format": "olir/0.1",
  "history": {
    "corrections": [],
    "supersedes": []
  },
  "id": "34822c9e0d3840f6a0fb632d",
  "measured_at": "2026-10-07T04:04:27.356047+00:00",
  "metrics": [
    {
      "basis": null,
      "n": null,
      "name": "wall_seconds",
      "scope": "machine_total",
      "source_field": "wall_seconds",
      "statistic": "single",
      "unit": "s",
      "value": 17.233130931854248
    },
    {
      "basis": null,
      "n": 1,
      "name": "ttft_median_s",
      "scope": "per_person",
      "source_field": "ttft_median_s",
      "statistic": "median",
      "unit": "s",
      "value": 9.77921269997023
    },
    {
      "basis": null,
      "n": null,
      "name": "needle_total",
      "scope": "machine_total",
      "source_field": "needle_total",
      "statistic": "single",
      "unit": "count",
      "value": 3
    },
    {
      "basis": null,
      "n": 1,
      "name": "ttft_max_s",
      "scope": "per_person",
      "source_field": "ttft_max_s",
      "statistic": "max",
      "unit": "s",
      "value": 9.77921269997023
    },
    {
      "basis": null,
      "n": null,
      "name": "requests_ok",
      "scope": "machine_total",
      "source_field": "requests_ok",
      "statistic": "single",
      "unit": "count",
      "value": 1
    },
    {
      "basis": null,
      "n": null,
      "name": "needle_correct",
      "scope": "machine_total",
      "source_field": "needle_correct",
      "statistic": "single",
      "unit": "count",
      "value": 3
    },
    {
      "basis": null,
      "n": null,
      "name": "preemptions",
      "scope": "machine_total",
      "source_field": "preemptions",
      "statistic": "single",
      "unit": "count",
      "value": 0
    },
    {
      "basis": null,
      "n": null,
      "name": "decode_overlap",
      "scope": "machine_total",
      "source_field": "decode_overlap",
      "statistic": "single",
      "unit": "count",
      "value": 1
    },
    {
      "basis": null,
      "n": null,
      "name": "aggregate_concurrency",
      "scope": "machine_total",
      "source_field": "aggregate_concurrency",
      "statistic": "single",
      "unit": "count",
      "value": 1
    },
    {
      "basis": null,
      "n": null,
      "name": "requests_total",
      "scope": "machine_total",
      "source_field": "requests_total",
      "statistic": "single",
      "unit": "count",
      "value": 1
    },
    {
      "basis": null,
      "n": null,
      "name": "correctness_probes_passed",
      "scope": "machine_total",
      "source_field": "correctness_probes_passed",
      "statistic": "single",
      "unit": "count",
      "value": 1
    },
    {
      "basis": null,
      "n": null,
      "name": "aggregate_decode_tps",
      "scope": "machine_total",
      "source_field": "aggregate_tps",
      "statistic": "single",
      "unit": "tok/s",
      "value": 68.55454876389933
    },
    {
      "basis": null,
      "n": null,
      "name": "aggregate_e2e_tps",
      "scope": "machine_total",
      "source_field": "aggregate_end_to_end_tps",
      "statistic": "single",
      "unit": "tok/s",
      "value": 29.710213543007644
    },
    {
      "basis": null,
      "n": null,
      "name": "correctness_probes_total",
      "scope": "machine_total",
      "source_field": "correctness_probes_total",
      "statistic": "single",
      "unit": "count",
      "value": 1
    },
    {
      "basis": null,
      "n": null,
      "name": "single_user_decode_tps",
      "scope": "per_person",
      "source_field": "single_user_decode_tps",
      "statistic": "single",
      "unit": "tok/s",
      "value": 68.55455903337788
    },
    {
      "basis": null,
      "n": 1,
      "name": "decode_tps",
      "scope": "per_person",
      "source_field": "per_request_decode_tps",
      "statistic": "mean",
      "unit": "tok/s",
      "value": 68.55455903337788
    },
    {
      "basis": null,
      "n": null,
      "name": "forced_output_tokens",
      "scope": "machine_total",
      "source_field": "forced_output_tokens",
      "statistic": "single",
      "unit": "tokens",
      "value": 512
    },
    {
      "basis": null,
      "n": null,
      "name": "peak_host_unavailable_gib",
      "scope": "machine_total",
      "source_field": "peak_host_unavailable_gib",
      "statistic": "max",
      "unit": "GiB",
      "value": 92.44107818603516
    },
    {
      "basis": null,
      "n": null,
      "name": "ttft_s",
      "scope": "per_person",
      "source_field": "team_load_requests[0].ttft_s",
      "statistic": "single",
      "unit": "s",
      "value": 9.77921269997023
    },
    {
      "basis": null,
      "n": null,
      "name": "total_s",
      "scope": "per_person",
      "source_field": "team_load_requests[0].total_s",
      "statistic": "single",
      "unit": "s",
      "value": 17.233129801985342
    },
    {
      "basis": null,
      "n": null,
      "name": "decode_tps",
      "scope": "per_person",
      "source_field": "team_load_requests[0].decode_tps",
      "statistic": "single",
      "unit": "tok/s",
      "value": 68.55455903337788
    },
    {
      "basis": null,
      "n": null,
      "name": "user_number",
      "scope": "per_person",
      "source_field": "team_load_requests[0].user_number",
      "statistic": "single",
      "unit": "count",
      "value": 0
    },
    {
      "basis": null,
      "n": null,
      "name": "prompt_tokens",
      "scope": "per_person",
      "source_field": "team_load_requests[0].prompt_tokens",
      "statistic": "single",
      "unit": "tokens",
      "value": 40101
    },
    {
      "basis": null,
      "n": null,
      "name": "output_tokens",
      "scope": "per_person",
      "source_field": "team_load_requests[0].output_tokens",
      "statistic": "single",
      "unit": "tokens",
      "value": 512
    },
    {
      "basis": null,
      "n": null,
      "name": "needle_correct",
      "scope": "per_person",
      "source_field": "team_load_requests[0].needle_correct",
      "statistic": "single",
      "unit": "count",
      "value": 3
    },
    {
      "basis": null,
      "n": null,
      "name": "needle_total",
      "scope": "per_person",
      "source_field": "team_load_requests[0].needle_total",
      "statistic": "single",
      "unit": "count",
      "value": 3
    },
    {
      "basis": null,
      "n": 1,
      "name": "decode_tps",
      "scope": "per_person",
      "source_field": "per_person_decode_median_tps",
      "statistic": "median",
      "unit": "tok/s",
      "value": 68.55455903337788
    }
  ],
  "model": {
    "artifact_sha256": [
      "07141c2db92e47bc08777132cd1a0323faf300eab3a7d7c111bc2bf075fda050",
      "6dea9c759a0f941cf9e1cc1501216b0e107966a66c90425050e212efbd053f02",
      "9758875fc55e49561165f4a44342b654c123c3e25c6811b34abb83553fb1a164"
    ],
    "publisher_repo": "Qwen/Qwen3.6-35B-A3B",
    "publisher_revision": "995ad96eacd98c81ed38be0c5b274b04031597b0",
    "variant": "NVFP4 variant",
    "weights_dtype": "nvfp4",
    "weights_format": "safetensors"
  },
  "outcome": {
    "gates": [
      {
        "met": true,
        "name": "requests",
        "observed": "1/1"
      },
      {
        "met": true,
        "name": "hidden-facts",
        "observed": "3/3"
      },
      {
        "met": true,
        "name": "correctness-probes",
        "observed": "1/1"
      }
    ],
    "reasons": [],
    "status": "pass"
  },
  "publication_group": "pre-standard",
  "published": null,
  "publisher": {
    "name": "Skilak Keep"
  },
  "quality": [
    {
      "passed": 3,
      "skipped": null,
      "suite": "hidden-facts",
      "total": 3,
      "version": "1"
    },
    {
      "passed": 1,
      "skipped": null,
      "suite": "correctness-probes",
      "total": 1,
      "version": "1"
    }
  ],
  "record_date": "2026-10-07",
  "reproduce": {
    "driver": null,
    "fixture_sha256": "e2097959b1ca0142f2b93edc3ec411e474c0a6622800dd247af7a71bd9c39e83",
    "harness": {
      "name": "keep-team-load-v1",
      "version": "1"
    },
    "harness_sha256": null,
    "os": null,
    "workload_name": null,
    "workload_version": null
  },
  "requests": [
    {
      "cached_prompt_tokens": null,
      "checks_passed": true,
      "completed": true,
      "decode_tps": 68.55455903337788,
      "effective_prefill_tps": null,
      "needle_correct": 3,
      "needle_total": 3,
      "output_tokens": 512,
      "prompt_tokens": 40101,
      "total_s": 17.233129801985342,
      "ttft_s": 9.77921269997023,
      "user_number": 0
    }
  ],
  "scenario": {
    "id": "team-load",
    "version": "1"
  },
  "settings": {
    "concurrency": 1,
    "configured_context_tokens": 262144,
    "context_tokens": 262144,
    "kv_dtype": "fp8",
    "max_output_tokens": 512,
    "speculative_decoding": "plain",
    "thinking": false
  },
  "setup_complete": false,
  "setup_id": "a4b96c03f67c006a12a64bbc7734ee71",
  "standard": "pre-standard",
  "system": {
    "driver": null,
    "hardware": "NVIDIA GB10, 128 GB unified memory",
    "os": null
  },
  "workload": {
    "cache": null,
    "content_tokens": null,
    "document": null,
    "input_tokens": {
      "max": 40101,
      "median": 40101,
      "min": 40101
    },
    "output_tokens": 512,
    "repetitions": null,
    "users": 1,
    "warmup": null
  }
}