{
  "base_url": "http://127.0.0.1:30001/v1",
  "clock_domain_id": null,
  "completed": 8,
  "concurrency": 4,
  "configuration_fingerprint": null,
  "context_distribution": {
    "8192": 8
  },
  "context_seed": 1909,
  "context_tokens": 8192,
  "controlled_output_policy": "strict",
  "engine": "vllm-v84",
  "failed": 0,
  "failures": [],
  "finished_at": "2026-09-19T11:22:05Z",
  "finished_at_unix_ns": 1789816925066307583,
  "gpu": "2x-rtx-pro-6000-max-q",
  "manifest": null,
  "max_context_tokens": 327680,
  "max_tokens": 4096,
  "measurement_protocol": "capacity-v4-reasoning",
  "metric_population": {
    "aggregate_throughput_eligible": true,
    "attempted_requests": 8,
    "excluded_requests": 0,
    "failure_policy": "failed or contract-invalid responses retain sanitized timing and metadata but are excluded from performance distributions",
    "p99_interpretation": "descriptive-nearest-rank-only",
    "p99_interpretation_by_metric": {
      "decode_tok_s": "descriptive-nearest-rank-only",
      "e2e": "descriptive-nearest-rank-only",
      "effective_prefill_tok_s": "descriptive-nearest-rank-only",
      "tpot": "descriptive-nearest-rank-only",
      "ttft": "descriptive-nearest-rank-only"
    },
    "p99_minimum_recommended_samples": 100,
    "p99_samples_by_metric": {
      "decode_tok_s": 8,
      "e2e": 8,
      "effective_prefill_tok_s": 8,
      "tpot": 8,
      "ttft": 8
    },
    "performance_eligible_requests": 8,
    "successful_requests": 8
  },
  "metrics": {
    "completion_tokens_max": 352.0,
    "completion_tokens_mean": 218.5,
    "completion_tokens_mean_ci95_high": 274.9536872135027,
    "completion_tokens_mean_ci95_low": 162.04631278649728,
    "completion_tokens_min": 121.0,
    "completion_tokens_p25": 128.0,
    "completion_tokens_p50": 187.0,
    "completion_tokens_p75": 245.0,
    "completion_tokens_p90": 352.0,
    "completion_tokens_p95": 352.0,
    "completion_tokens_p99": 352.0,
    "completion_tokens_samples": 8,
    "completion_tokens_stddev": 81.46690826867767,
    "content_chunks": 208,
    "content_chunks_s": null,
    "controlled_output_exact_adherence": 8,
    "controlled_output_exact_adherence_rate": 1.0,
    "controlled_output_samples": 8,
    "controlled_output_unobservable": 0,
    "decode_tok_s_max": 64.18702009730778,
    "decode_tok_s_mean": 36.161628188224924,
    "decode_tok_s_mean_ci95_high": 48.06354276789217,
    "decode_tok_s_mean_ci95_low": 24.259713608557675,
    "decode_tok_s_min": 14.665019691762916,
    "decode_tok_s_p25": 19.231336722171285,
    "decode_tok_s_p50": 36.553921107783516,
    "decode_tok_s_p75": 40.04670709877654,
    "decode_tok_s_p90": 64.18702009730778,
    "decode_tok_s_p95": 64.18702009730778,
    "decode_tok_s_p99": 64.18702009730778,
    "decode_tok_s_samples": 8,
    "decode_tok_s_stddev": 17.175356139562755,
    "e2e_max_ms": 17244.001089944504,
    "e2e_mean_ci95_high_ms": 13151.578274604259,
    "e2e_mean_ci95_low_ms": 6147.147893565847,
    "e2e_mean_ms": 9649.363084085053,
    "e2e_min_ms": 2605.1619799109176,
    "e2e_p25_ms": 5361.519476980902,
    "e2e_p50_ms": 7359.892691951245,
    "e2e_p75_ms": 12844.19034793973,
    "e2e_p90_ms": 17244.001089944504,
    "e2e_p95_ms": 17244.001089944504,
    "e2e_p99_ms": 17244.001089944504,
    "e2e_samples": 8,
    "e2e_stddev_ms": 5053.959408960545,
    "effective_prefill_tok_s_max": 8246.277251503414,
    "effective_prefill_tok_s_mean": 5057.968559919313,
    "effective_prefill_tok_s_mean_ci95_high": 7259.258778801149,
    "effective_prefill_tok_s_mean_ci95_low": 2856.6783410374765,
    "effective_prefill_tok_s_min": 710.4903896502088,
    "effective_prefill_tok_s_p25": 1910.4952933691902,
    "effective_prefill_tok_s_p50": 3472.4681860739324,
    "effective_prefill_tok_s_p75": 8239.250992449575,
    "effective_prefill_tok_s_p90": 8246.277251503414,
    "effective_prefill_tok_s_p95": 8246.277251503414,
    "effective_prefill_tok_s_p99": 8246.277251503414,
    "effective_prefill_tok_s_samples": 8,
    "effective_prefill_tok_s_stddev": 3176.627022716257,
    "generation_max_ms": 15498.556733014993,
    "generation_mean_ci95_high_ms": 10132.785405206238,
    "generation_mean_ci95_low_ms": 4477.392405999848,
    "generation_mean_ms": 7305.088905603043,
    "generation_min_ms": 1869.5368599146605,
    "generation_p25_ms": 4626.151846954599,
    "generation_p50_ms": 6092.885474907234,
    "generation_p75_ms": 8660.063380026259,
    "generation_p90_ms": 15498.556733014993,
    "generation_p95_ms": 15498.556733014993,
    "generation_p99_ms": 15498.556733014993,
    "generation_samples": 8,
    "generation_stddev_ms": 4080.5783061364946,
    "mean_inter_token_latency_ms_max": 68.18947543327762,
    "mean_inter_token_latency_ms_mean": 34.65463119822111,
    "mean_inter_token_latency_ms_mean_ci95_high": 47.36733280974576,
    "mean_inter_token_latency_ms_mean_ci95_low": 21.941929586696464,
    "mean_inter_token_latency_ms_min": 15.57947383262217,
    "mean_inter_token_latency_ms_p25": 18.399434531590224,
    "mean_inter_token_latency_ms_p50": 26.587079580198846,
    "mean_inter_token_latency_ms_p75": 44.155432287792,
    "mean_inter_token_latency_ms_p90": 68.18947543327762,
    "mean_inter_token_latency_ms_p95": 68.18947543327762,
    "mean_inter_token_latency_ms_p99": 68.18947543327762,
    "mean_inter_token_latency_ms_samples": 8,
    "mean_inter_token_latency_ms_stddev": 18.345382687163735,
    "mean_time_per_output_token_max_ms": 68.18947543327762,
    "mean_time_per_output_token_mean_ci95_high_ms": 47.36733280974576,
    "mean_time_per_output_token_mean_ci95_low_ms": 21.941929586696464,
    "mean_time_per_output_token_mean_ms": 34.65463119822111,
    "mean_time_per_output_token_min_ms": 15.57947383262217,
    "mean_time_per_output_token_p25_ms": 18.399434531590224,
    "mean_time_per_output_token_p50_ms": 26.587079580198846,
    "mean_time_per_output_token_p75_ms": 44.155432287792,
    "mean_time_per_output_token_p90_ms": 68.18947543327762,
    "mean_time_per_output_token_p95_ms": 68.18947543327762,
    "mean_time_per_output_token_p99_ms": 68.18947543327762,
    "mean_time_per_output_token_samples": 8,
    "mean_time_per_output_token_stddev_ms": 18.345382687163735,
    "output_token_sources": {
      "usage": 8
    },
    "output_tokens": 1748,
    "prefix_cache_hit_avg": null,
    "prompt_token_samples": 8,
    "prompt_tokens": 48488,
    "prompt_tokens_max": 6061.0,
    "prompt_tokens_mean": 6061.0,
    "prompt_tokens_mean_ci95_high": 6061.0,
    "prompt_tokens_mean_ci95_low": 6061.0,
    "prompt_tokens_min": 6061.0,
    "prompt_tokens_p25": 6061.0,
    "prompt_tokens_p50": 6061.0,
    "prompt_tokens_p75": 6061.0,
    "prompt_tokens_p90": 6061.0,
    "prompt_tokens_p95": 6061.0,
    "prompt_tokens_p99": 6061.0,
    "prompt_tokens_samples": 8,
    "prompt_tokens_stddev": 0.0,
    "reasoning_chunks": 1387,
    "request_canary_passed": 8,
    "request_canary_samples": 8,
    "throughput_performance_eligible": true,
    "throughput_tok_s": 83.6431108942441,
    "time_to_first_output_max_ms": 8530.727633042261,
    "time_to_first_output_mean_ci95_high_ms": 4182.401493395237,
    "time_to_first_output_mean_ci95_low_ms": 506.14686356878315,
    "time_to_first_output_mean_ms": 2344.27417848201,
    "time_to_first_output_min_ms": 734.9983289605007,
    "time_to_first_output_p25_ms": 735.3676300263032,
    "time_to_first_output_p50_ms": 876.3201279798523,
    "time_to_first_output_p75_ms": 2223.234475008212,
    "time_to_first_output_p90_ms": 8530.727633042261,
    "time_to_first_output_p95_ms": 8530.727633042261,
    "time_to_first_output_p99_ms": 8530.727633042261,
    "time_to_first_output_samples": 8,
    "time_to_first_output_stddev_ms": 2652.5556919578844,
    "tpot_max_ms": 68.18947543327762,
    "tpot_mean_ci95_high_ms": 47.36733280974576,
    "tpot_mean_ci95_low_ms": 21.941929586696464,
    "tpot_mean_ms": 34.65463119822111,
    "tpot_min_ms": 15.57947383262217,
    "tpot_p25_ms": 18.399434531590224,
    "tpot_p50_ms": 26.587079580198846,
    "tpot_p75_ms": 44.155432287792,
    "tpot_p90_ms": 68.18947543327762,
    "tpot_p95_ms": 68.18947543327762,
    "tpot_p99_ms": 68.18947543327762,
    "tpot_samples": 8,
    "tpot_stddev_ms": 18.345382687163735,
    "ttft_max_ms": 16748.365851934068,
    "ttft_mean_ci95_high_ms": 12591.315731967677,
    "ttft_mean_ci95_low_ms": 5450.920038964488,
    "ttft_mean_ms": 9021.117885466083,
    "ttft_min_ms": 2209.9412489915267,
    "ttft_p25_ms": 4208.901755977422,
    "ttft_p50_ms": 6207.96317292843,
    "ttft_p75_ms": 12347.814337932505,
    "ttft_p90_ms": 16748.365851934068,
    "ttft_p95_ms": 16748.365851934068,
    "ttft_p99_ms": 16748.365851934068,
    "ttft_samples": 8,
    "ttft_stddev_ms": 5152.063484569157
  },
  "model": "glm53-flash-exl3-4bpw-v84-fp8-tp2-c4-327k-r10-vision8-apc-nospec",
  "performance_eligible": true,
  "prompt_cache_mode": "shared",
  "request_canaries": true,
  "request_timings": [
    {
      "content_chunks": 26,
      "controlled_output": {
        "capture_complete": true,
        "exact_adherence": true,
        "observed_code_words": 16,
        "observed_extra_words": 0,
        "passed": true,
        "policy": "strict",
        "requested_words": 16
      },
      "decode_tok_s": 22.647270068205806,
      "e2e_ms": 17244.001089944504,
      "effective_prefill_tok_s": 3472.4681860739324,
      "generation_ms": 15498.556733014993,
      "mean_inter_token_latency_ms": 44.155432287792,
      "mean_time_per_output_token_ms": 44.155432287792,
      "output_token_source": "usage",
      "output_tokens": 352,
      "planned_context_tokens": 8192,
      "prompt_tokens": 6061,
      "reasoning_chunks": 301,
      "request_canary": {
        "capture_complete": true,
        "foreign_marker_count": 0,
        "marker": "ANVIL_REQ_1909_00000",
        "marker_at_start": true,
        "observed_marker_count": 1,
        "passed": true
      },
      "request_index": 0,
      "time_to_first_output_ms": 1745.4443569295108,
      "tpot_ms": 44.155432287792,
      "ttft_ms": 16748.365851934068,
      "visible_content_truncated": false,
      "visible_generation_ms": 495.6352380104363
    },
    {
      "content_chunks": 26,
      "controlled_output": {
        "capture_complete": true,
        "exact_adherence": true,
        "observed_code_words": 16,
        "observed_extra_words": 0,
        "passed": true,
        "policy": "strict",
        "requested_words": 16
      },
      "decode_tok_s": 14.665019691762916,
      "e2e_ms": 10883.29785503447,
      "effective_prefill_tok_s": 2726.208174680996,
      "generation_ms": 8660.063380026259,
      "mean_inter_token_latency_ms": 68.18947543327762,
      "mean_time_per_output_token_ms": 68.18947543327762,
      "output_token_source": "usage",
      "output_tokens": 128,
      "planned_context_tokens": 8192,
      "prompt_tokens": 6061,
      "reasoning_chunks": 96,
      "request_canary": {
        "capture_complete": true,
        "foreign_marker_count": 0,
        "marker": "ANVIL_REQ_1909_00001",
        "marker_at_start": true,
        "observed_marker_count": 1,
        "passed": true
      },
      "request_index": 1,
      "time_to_first_output_ms": 2223.234475008212,
      "tpot_ms": 68.18947543327762,
      "ttft_ms": 10392.801402020268,
      "visible_content_truncated": false,
      "visible_generation_ms": 490.4964530142024
    },
    {
      "content_chunks": 26,
      "controlled_output": {
        "capture_complete": true,
        "exact_adherence": true,
        "observed_code_words": 16,
        "observed_extra_words": 0,
        "passed": true,
        "policy": "strict",
        "requested_words": 16
      },
      "decode_tok_s": 19.231336722171285,
      "e2e_ms": 12844.19034793973,
      "effective_prefill_tok_s": 1910.4952933691902,
      "generation_ms": 9671.714592026547,
      "mean_inter_token_latency_ms": 51.99846554852982,
      "mean_time_per_output_token_ms": 51.99846554852982,
      "output_token_source": "usage",
      "output_tokens": 187,
      "planned_context_tokens": 8192,
      "prompt_tokens": 6061,
      "reasoning_chunks": 139,
      "request_canary": {
        "capture_complete": true,
        "foreign_marker_count": 0,
        "marker": "ANVIL_REQ_1909_00002",
        "marker_at_start": true,
        "observed_marker_count": 1,
        "passed": true
      },
      "request_index": 2,
      "time_to_first_output_ms": 3172.475755913183,
      "tpot_ms": 51.99846554852982,
      "ttft_ms": 12347.814337932505,
      "visible_content_truncated": false,
      "visible_generation_ms": 496.376010007225
    },
    {
      "content_chunks": 26,
      "controlled_output": {
        "capture_complete": true,
        "exact_adherence": true,
        "observed_code_words": 16,
        "observed_extra_words": 0,
        "passed": true,
        "policy": "strict",
        "requested_words": 16
      },
      "decode_tok_s": 40.04670709877654,
      "e2e_ms": 14623.613107949495,
      "effective_prefill_tok_s": 710.4903896502088,
      "generation_ms": 6092.885474907234,
      "mean_inter_token_latency_ms": 24.97084211027555,
      "mean_time_per_output_token_ms": 24.97084211027555,
      "output_token_source": "usage",
      "output_tokens": 245,
      "planned_context_tokens": 8192,
      "prompt_tokens": 6061,
      "reasoning_chunks": 196,
      "request_canary": {
        "capture_complete": true,
        "foreign_marker_count": 0,
        "marker": "ANVIL_REQ_1909_00003",
        "marker_at_start": true,
        "observed_marker_count": 1,
        "passed": true
      },
      "request_index": 3,
      "time_to_first_output_ms": 8530.727633042261,
      "tpot_ms": 24.97084211027555,
      "ttft_ms": 14129.15194395464,
      "visible_content_truncated": false,
      "visible_generation_ms": 494.4611639948562
    },
    {
      "content_chunks": 26,
      "controlled_output": {
        "capture_complete": true,
        "exact_adherence": true,
        "observed_code_words": 16,
        "observed_extra_words": 0,
        "passed": true,
        "policy": "strict",
        "requested_words": 16
      },
      "decode_tok_s": 36.553921107783516,
      "e2e_ms": 7359.892691951245,
      "effective_prefill_tok_s": 6916.422214302203,
      "generation_ms": 6483.572563971393,
      "mean_inter_token_latency_ms": 27.35684626148267,
      "mean_time_per_output_token_ms": 27.35684626148267,
      "output_token_source": "usage",
      "output_tokens": 238,
      "planned_context_tokens": 8192,
      "prompt_tokens": 6061,
      "reasoning_chunks": 189,
      "request_canary": {
        "capture_complete": true,
        "foreign_marker_count": 0,
        "marker": "ANVIL_REQ_1909_00004",
        "marker_at_start": true,
        "observed_marker_count": 1,
        "passed": true
      },
      "request_index": 4,
      "time_to_first_output_ms": 876.3201279798523,
      "tpot_ms": 27.35684626148267,
      "ttft_ms": 6207.96317292843,
      "visible_content_truncated": false,
      "visible_generation_ms": 1151.929519022815
    },
    {
      "content_chunks": 26,
      "controlled_output": {
        "capture_complete": true,
        "exact_adherence": true,
        "observed_code_words": 16,
        "observed_extra_words": 0,
        "passed": true,
        "policy": "strict",
        "requested_words": 16
      },
      "decode_tok_s": 37.61225436526568,
      "e2e_ms": 5361.519476980902,
      "effective_prefill_tok_s": 8242.13597732498,
      "generation_ms": 4626.151846954599,
      "mean_inter_token_latency_ms": 26.587079580198846,
      "mean_time_per_output_token_ms": 26.587079580198846,
      "output_token_source": "usage",
      "output_tokens": 175,
      "planned_context_tokens": 8192,
      "prompt_tokens": 6061,
      "reasoning_chunks": 127,
      "request_canary": {
        "capture_complete": true,
        "foreign_marker_count": 0,
        "marker": "ANVIL_REQ_1909_00005",
        "marker_at_start": true,
        "observed_marker_count": 1,
        "passed": true
      },
      "request_index": 5,
      "time_to_first_output_ms": 735.3676300263032,
      "tpot_ms": 26.587079580198846,
      "ttft_ms": 4208.901755977422,
      "visible_content_truncated": false,
      "visible_generation_ms": 1152.6177210034803
    },
    {
      "content_chunks": 26,
      "controlled_output": {
        "capture_complete": true,
        "exact_adherence": true,
        "observed_code_words": 16,
        "observed_extra_words": 0,
        "passed": true,
        "policy": "strict",
        "requested_words": 16
      },
      "decode_tok_s": 54.349496354525854,
      "e2e_ms": 6273.228122969158,
      "effective_prefill_tok_s": 8246.277251503414,
      "generation_ms": 5538.229794008657,
      "mean_inter_token_latency_ms": 18.399434531590224,
      "mean_time_per_output_token_ms": 18.399434531590224,
      "output_token_source": "usage",
      "output_tokens": 302,
      "planned_context_tokens": 8192,
      "prompt_tokens": 6061,
      "reasoning_chunks": 250,
      "request_canary": {
        "capture_complete": true,
        "foreign_marker_count": 0,
        "marker": "ANVIL_REQ_1909_00006",
        "marker_at_start": true,
        "observed_marker_count": 1,
        "passed": true
      },
      "request_index": 6,
      "time_to_first_output_ms": 734.9983289605007,
      "tpot_ms": 18.399434531590224,
      "ttft_ms": 5924.003369989805,
      "visible_content_truncated": false,
      "visible_generation_ms": 349.22475297935307
    },
    {
      "content_chunks": 26,
      "controlled_output": {
        "capture_complete": true,
        "exact_adherence": true,
        "observed_code_words": 16,
        "observed_extra_words": 0,
        "passed": true,
        "policy": "strict",
        "requested_words": 16
      },
      "decode_tok_s": 64.18702009730778,
      "e2e_ms": 2605.1619799109176,
      "effective_prefill_tok_s": 8239.250992449575,
      "generation_ms": 1869.5368599146605,
      "mean_inter_token_latency_ms": 15.57947383262217,
      "mean_time_per_output_token_ms": 15.57947383262217,
      "output_token_source": "usage",
      "output_tokens": 121,
      "planned_context_tokens": 8192,
      "prompt_tokens": 6061,
      "reasoning_chunks": 89,
      "request_canary": {
        "capture_complete": true,
        "foreign_marker_count": 0,
        "marker": "ANVIL_REQ_1909_00007",
        "marker_at_start": true,
        "observed_marker_count": 1,
        "passed": true
      },
      "request_index": 7,
      "time_to_first_output_ms": 735.6251199962571,
      "tpot_ms": 15.57947383262217,
      "ttft_ms": 2209.9412489915267,
      "visible_content_truncated": false,
      "visible_generation_ms": 395.2207309193909
    }
  ],
  "requests": 8,
  "response_words": 16,
  "responses_received": 8,
  "run_id": "benchmark-20260919T112144Z",
  "sampling": {
    "temperature": {
      "effective_request": 0.0,
      "requested": 0.0,
      "sent": true
    },
    "top_p": {
      "effective_request": 0.95,
      "requested": 0.95,
      "sent": true
    }
  },
  "schema": "anvil-serving.benchmark/v1",
  "serve_flags": {
    "no_thinking": false,
    "reasoning_effort": "max",
    "shared_prefix_burst": false,
    "thinking_mode": "default"
  },
  "shared_prefix_tokens": 8000,
  "source_recipe": null,
  "started_at": "2026-09-19T11:21:44Z",
  "started_at_unix_ns": 1789816904167990718,
  "tier": null,
  "timestamp_precision": "nanosecond-unix-fields; ISO labels are second precision",
  "timing_methodology": {
    "artifact_timestamps": "time.time_ns Unix timestamps; independent artifacts may align them only when the operator supplied the same explicit clock_domain_id",
    "clock": "client time.perf_counter",
    "controlled_output": "--response-words adherence is recorded; observe preserves historical behavior while strict rejects non-exact or truncated completions; visible response capture is capped at 8192 characters",
    "decode": "usage completion tokens after the first token divided by client-observed generation time; completion tokens may include reasoning tokens",
    "distributions": "request-level mean, sample standard deviation, normal-approximation 95% confidence interval for the mean, min, p25, p50, p75, p90, p95, p99, and max using the nearest-rank percentile method",
    "effective_prefill": "usage.prompt_tokens divided by client-observed time to first output; includes queueing, scheduling, prefill, and first-token work",
    "generation": "client-observed E2E minus time to first output",
    "mean_inter_token_latency": "legacy key for a TPOT proxy: client-observed generation time divided by completion tokens after the first token; not raw per-token ITL",
    "p99": "numeric nearest-rank p99 is always retained for compatibility; it is descriptive only below 100 eligible samples",
    "request_canaries": "when enabled, each response must begin with its own synthetic marker and contain no other ANVIL_REQ marker; bounded/truncated captures fail closed and are excluded from performance metrics",
    "time_to_first_output": "request start through the first non-empty streamed reasoning or content delta",
    "tpot": "client-observed mean time per output token after the first output; identical to the legacy mean_inter_token_latency key",
    "ttft": "request start through the first non-empty streamed content delta"
  },
  "wall_clock_ms": 20898.314054938965
}
