{
  "base_url": "http://127.0.0.1:30000/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": "2026-09-19T11:15:58Z",
  "finished_at_unix_ns": 1789816558670837757,
  "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": 384.0,
    "completion_tokens_mean": 248.625,
    "completion_tokens_mean_ci95_high": 325.34711561864026,
    "completion_tokens_mean_ci95_low": 171.90288438135977,
    "completion_tokens_min": 75.0,
    "completion_tokens_p25": 168.0,
    "completion_tokens_p50": 187.0,
    "completion_tokens_p75": 335.0,
    "completion_tokens_p90": 384.0,
    "completion_tokens_p95": 384.0,
    "completion_tokens_p99": 384.0,
    "completion_tokens_samples": 8,
    "completion_tokens_stddev": 110.71577187942621,
    "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": 65.72490855510684,
    "decode_tok_s_mean": 35.48324551826503,
    "decode_tok_s_mean_ci95_high": 46.033109583502416,
    "decode_tok_s_mean_ci95_low": 24.933381453027643,
    "decode_tok_s_min": 19.694036022545973,
    "decode_tok_s_p25": 21.063972364684496,
    "decode_tok_s_p50": 31.965775653529583,
    "decode_tok_s_p75": 37.407516710398774,
    "decode_tok_s_p90": 65.72490855510684,
    "decode_tok_s_p95": 65.72490855510684,
    "decode_tok_s_p99": 65.72490855510684,
    "decode_tok_s_samples": 8,
    "decode_tok_s_stddev": 15.224245757399249,
    "e2e_max_ms": 15989.138527074829,
    "e2e_mean_ci95_high_ms": 12064.59187949225,
    "e2e_mean_ci95_low_ms": 7432.828347022553,
    "e2e_mean_ms": 9748.710113257403,
    "e2e_min_ms": 6076.927501009777,
    "e2e_p25_ms": 6506.786016980186,
    "e2e_p50_ms": 9383.096884936094,
    "e2e_p75_ms": 11594.539429061115,
    "e2e_p90_ms": 15989.138527074829,
    "e2e_p95_ms": 15989.138527074829,
    "e2e_p99_ms": 15989.138527074829,
    "e2e_samples": 8,
    "e2e_stddev_ms": 3341.991227206002,
    "effective_prefill_tok_s_max": 4287.5747718433795,
    "effective_prefill_tok_s_mean": 3245.3966057797384,
    "effective_prefill_tok_s_mean_ci95_high": 4231.972270176578,
    "effective_prefill_tok_s_mean_ci95_low": 2258.820941382899,
    "effective_prefill_tok_s_min": 1091.418820330701,
    "effective_prefill_tok_s_p25": 1440.2157174296183,
    "effective_prefill_tok_s_p50": 4237.562621197824,
    "effective_prefill_tok_s_p75": 4251.972869366432,
    "effective_prefill_tok_s_p90": 4287.5747718433795,
    "effective_prefill_tok_s_p95": 4287.5747718433795,
    "effective_prefill_tok_s_p99": 4287.5747718433795,
    "effective_prefill_tok_s_samples": 8,
    "effective_prefill_tok_s_stddev": 1423.702739691079,
    "generation_max_ms": 10435.816660989076,
    "generation_mean_ci95_high_ms": 9235.404797838315,
    "generation_mean_ci95_low_ms": 5337.990559396088,
    "generation_mean_ms": 7286.697678617202,
    "generation_min_ms": 2694.6208369918168,
    "generation_p25_ms": 4651.471557095647,
    "generation_p50_ms": 7952.793537988327,
    "generation_p75_ms": 8830.24325990118,
    "generation_p90_ms": 10435.816660989076,
    "generation_p95_ms": 10435.816660989076,
    "generation_p99_ms": 10435.816660989076,
    "generation_samples": 8,
    "generation_stddev_ms": 2812.130650097451,
    "mean_inter_token_latency_ms_max": 50.77679348484931,
    "mean_inter_token_latency_ms_mean": 32.44195885666397,
    "mean_inter_token_latency_ms_mean_ci95_high": 40.91867371730799,
    "mean_inter_token_latency_ms_mean_ci95_low": 23.965243996019957,
    "mean_inter_token_latency_ms_min": 15.214931781328433,
    "mean_inter_token_latency_ms_p25": 20.76447398952566,
    "mean_inter_token_latency_ms_p50": 30.87519722186117,
    "mean_inter_token_latency_ms_p75": 36.41379509448401,
    "mean_inter_token_latency_ms_p90": 50.77679348484931,
    "mean_inter_token_latency_ms_p95": 50.77679348484931,
    "mean_inter_token_latency_ms_p99": 50.77679348484931,
    "mean_inter_token_latency_ms_samples": 8,
    "mean_inter_token_latency_ms_stddev": 12.23253583703299,
    "mean_time_per_output_token_max_ms": 50.77679348484931,
    "mean_time_per_output_token_mean_ci95_high_ms": 40.91867371730799,
    "mean_time_per_output_token_mean_ci95_low_ms": 23.965243996019957,
    "mean_time_per_output_token_mean_ms": 32.44195885666397,
    "mean_time_per_output_token_min_ms": 15.214931781328433,
    "mean_time_per_output_token_p25_ms": 20.76447398952566,
    "mean_time_per_output_token_p50_ms": 30.87519722186117,
    "mean_time_per_output_token_p75_ms": 36.41379509448401,
    "mean_time_per_output_token_p90_ms": 50.77679348484931,
    "mean_time_per_output_token_p95_ms": 50.77679348484931,
    "mean_time_per_output_token_p99_ms": 50.77679348484931,
    "mean_time_per_output_token_samples": 8,
    "mean_time_per_output_token_stddev_ms": 12.23253583703299,
    "output_token_sources": {
      "usage": 8
    },
    "output_tokens": 1989,
    "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": 1638,
    "request_canary_passed": 8,
    "request_canary_samples": 8,
    "throughput_performance_eligible": true,
    "throughput_tok_s": 88.48521879288275,
    "time_to_first_output_max_ms": 5553.321866085753,
    "time_to_first_output_mean_ci95_high_ms": 3576.7026657253846,
    "time_to_first_output_mean_ci95_low_ms": 1347.322203555017,
    "time_to_first_output_mean_ms": 2462.012434640201,
    "time_to_first_output_min_ms": 1413.6196620529518,
    "time_to_first_output_p25_ms": 1424.9988020164892,
    "time_to_first_output_p50_ms": 1427.41558002308,
    "time_to_first_output_p75_ms": 2812.5869320938364,
    "time_to_first_output_p90_ms": 5553.321866085753,
    "time_to_first_output_p95_ms": 5553.321866085753,
    "time_to_first_output_p99_ms": 5553.321866085753,
    "time_to_first_output_samples": 8,
    "time_to_first_output_stddev_ms": 1608.5816761688432,
    "tpot_max_ms": 50.77679348484931,
    "tpot_mean_ci95_high_ms": 40.91867371730799,
    "tpot_mean_ci95_low_ms": 23.965243996019957,
    "tpot_mean_ms": 32.44195885666397,
    "tpot_min_ms": 15.214931781328433,
    "tpot_p25_ms": 20.76447398952566,
    "tpot_p50_ms": 30.87519722186117,
    "tpot_p75_ms": 36.41379509448401,
    "tpot_p90_ms": 50.77679348484931,
    "tpot_p95_ms": 50.77679348484931,
    "tpot_p99_ms": 50.77679348484931,
    "tpot_samples": 8,
    "tpot_stddev_ms": 12.23253583703299,
    "ttft_max_ms": 15493.262408999726,
    "ttft_mean_ci95_high_ms": 11367.907194933938,
    "ttft_mean_ci95_low_ms": 6867.981025348566,
    "ttft_mean_ms": 9117.944110141252,
    "ttft_min_ms": 5581.193133024499,
    "ttft_p25_ms": 6157.563652959652,
    "ttft_p50_ms": 8990.15463097021,
    "ttft_p75_ms": 9816.42162206117,
    "ttft_p90_ms": 15493.262408999726,
    "ttft_p95_ms": 15493.262408999726,
    "ttft_p99_ms": 15493.262408999726,
    "ttft_samples": 8,
    "ttft_stddev_ms": 3246.8656217883913
  },
  "model": "glm53-flash-exl3-4bpw-v84-fp8-tp2-c4-327k-r9-vision8-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": 21.063972364684496,
      "e2e_ms": 11642.830191995017,
      "effective_prefill_tok_s": 2154.955614292027,
      "generation_ms": 8830.24325990118,
      "mean_inter_token_latency_ms": 47.47442612850097,
      "mean_time_per_output_token_ms": 47.47442612850097,
      "output_token_source": "usage",
      "output_tokens": 187,
      "planned_context_tokens": 8192,
      "prompt_tokens": 6061,
      "reasoning_chunks": 154,
      "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": 2812.5869320938364,
      "tpot_ms": 47.47442612850097,
      "ttft_ms": 9816.42162206117,
      "visible_content_truncated": false,
      "visible_generation_ms": 1826.4085699338466
    },
    {
      "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": 19.694036022545973,
      "e2e_ms": 9893.344174022786,
      "effective_prefill_tok_s": 4287.5747718433795,
      "generation_ms": 8479.724511969835,
      "mean_inter_token_latency_ms": 50.77679348484931,
      "mean_time_per_output_token_ms": 50.77679348484931,
      "output_token_source": "usage",
      "output_tokens": 168,
      "planned_context_tokens": 8192,
      "prompt_tokens": 6061,
      "reasoning_chunks": 135,
      "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": 1413.6196620529518,
      "tpot_ms": 50.77679348484931,
      "ttft_ms": 9395.710141048767,
      "visible_content_truncated": false,
      "visible_generation_ms": 497.63403297401965
    },
    {
      "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": 27.462119710545654,
      "e2e_ms": 6903.018180979416,
      "effective_prefill_tok_s": 1440.2157174296183,
      "generation_ms": 2694.6208369918168,
      "mean_inter_token_latency_ms": 36.41379509448401,
      "mean_time_per_output_token_ms": 36.41379509448401,
      "output_token_source": "usage",
      "output_tokens": 75,
      "planned_context_tokens": 8192,
      "prompt_tokens": 6061,
      "reasoning_chunks": 46,
      "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": 4208.397343987599,
      "tpot_ms": 36.41379509448401,
      "ttft_ms": 6405.9588389936835,
      "visible_content_truncated": false,
      "visible_generation_ms": 497.0593419857323
    },
    {
      "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": 32.3884570781608,
      "e2e_ms": 15989.138527074829,
      "effective_prefill_tok_s": 1091.418820330701,
      "generation_ms": 10435.816660989076,
      "mean_inter_token_latency_ms": 30.87519722186117,
      "mean_time_per_output_token_ms": 30.87519722186117,
      "output_token_source": "usage",
      "output_tokens": 339,
      "planned_context_tokens": 8192,
      "prompt_tokens": 6061,
      "reasoning_chunks": 289,
      "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": 5553.321866085753,
      "tpot_ms": 30.87519722186117,
      "ttft_ms": 15493.262408999726,
      "visible_content_truncated": false,
      "visible_generation_ms": 495.87611807510257
    },
    {
      "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": 31.965775653529583,
      "e2e_ms": 11594.539429061115,
      "effective_prefill_tok_s": 4246.135522706008,
      "generation_ms": 10167.123849038035,
      "mean_inter_token_latency_ms": 31.283457997040106,
      "mean_time_per_output_token_ms": 31.283457997040106,
      "output_token_source": "usage",
      "output_tokens": 326,
      "planned_context_tokens": 8192,
      "prompt_tokens": 6061,
      "reasoning_chunks": 276,
      "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": 1427.41558002308,
      "tpot_ms": 31.283457997040106,
      "ttft_ms": 11103.28845307231,
      "visible_content_truncated": false,
      "visible_generation_ms": 491.2509759888053
    },
    {
      "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": 37.407516710398774,
      "e2e_ms": 6076.927501009777,
      "effective_prefill_tok_s": 4251.972869366432,
      "generation_ms": 4651.471557095647,
      "mean_inter_token_latency_ms": 26.73259515572211,
      "mean_time_per_output_token_ms": 26.73259515572211,
      "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": 1425.4559439141303,
      "tpot_ms": 26.73259515572211,
      "ttft_ms": 5581.193133024499,
      "visible_content_truncated": false,
      "visible_generation_ms": 495.73436798527837
    },
    {
      "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": 48.15917805114811,
      "e2e_ms": 9383.096884936094,
      "effective_prefill_tok_s": 4237.562621197824,
      "generation_ms": 7952.793537988327,
      "mean_inter_token_latency_ms": 20.76447398952566,
      "mean_time_per_output_token_ms": 20.76447398952566,
      "output_token_source": "usage",
      "output_tokens": 384,
      "planned_context_tokens": 8192,
      "prompt_tokens": 6061,
      "reasoning_chunks": 327,
      "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": 1430.3033469477668,
      "tpot_ms": 20.76447398952566,
      "ttft_ms": 8990.15463097021,
      "visible_content_truncated": false,
      "visible_generation_ms": 392.94225396588445
    },
    {
      "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": 65.72490855510684,
      "e2e_ms": 6506.786016980186,
      "effective_prefill_tok_s": 4253.336909071918,
      "generation_ms": 5081.787214963697,
      "mean_inter_token_latency_ms": 15.214931781328433,
      "mean_time_per_output_token_ms": 15.214931781328433,
      "output_token_source": "usage",
      "output_tokens": 335,
      "planned_context_tokens": 8192,
      "prompt_tokens": 6061,
      "reasoning_chunks": 284,
      "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": 1424.9988020164892,
      "tpot_ms": 15.214931781328433,
      "ttft_ms": 6157.563652959652,
      "visible_content_truncated": false,
      "visible_generation_ms": 349.22236402053386
    }
  ],
  "requests": 8,
  "response_words": 16,
  "responses_received": 8,
  "run_id": "benchmark-20260919T111536Z",
  "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": "2026-09-19T11:15:36Z",
  "started_at_unix_ns": 1789816536192504001,
  "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": 22478.33058598917
}
