{
  "base_url": "http://127.0.0.1:30000/v1",
  "clock_domain_id": null,
  "completed": 8,
  "concurrency": 4,
  "configuration_fingerprint": null,
  "context_distribution": {
    "32768": 8
  },
  "context_seed": 1910,
  "context_tokens": 32768,
  "controlled_output_policy": "strict",
  "engine": "vllm-v84",
  "failed": 0,
  "failures": [],
  "finished_at": "2026-09-19T11:17:10Z",
  "finished_at_unix_ns": 1789816630949391934,
  "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": 182.0,
    "completion_tokens_mean": 141.25,
    "completion_tokens_mean_ci95_high": 163.65492133438545,
    "completion_tokens_mean_ci95_low": 118.84507866561457,
    "completion_tokens_min": 100.0,
    "completion_tokens_p25": 122.0,
    "completion_tokens_p50": 122.0,
    "completion_tokens_p75": 175.0,
    "completion_tokens_p90": 182.0,
    "completion_tokens_p95": 182.0,
    "completion_tokens_p99": 182.0,
    "completion_tokens_samples": 8,
    "completion_tokens_stddev": 32.33198328060224,
    "content_chunks": 216,
    "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": 56.096582189245304,
    "decode_tok_s_mean": 15.148340086443465,
    "decode_tok_s_mean_ci95_high": 26.805392643455846,
    "decode_tok_s_mean_ci95_low": 3.491287529431082,
    "decode_tok_s_min": 6.43833559045937,
    "decode_tok_s_p25": 6.676405951937544,
    "decode_tok_s_p50": 8.716046365979,
    "decode_tok_s_p75": 12.234340649698371,
    "decode_tok_s_p90": 56.096582189245304,
    "decode_tok_s_p95": 56.096582189245304,
    "decode_tok_s_p99": 56.096582189245304,
    "decode_tok_s_samples": 8,
    "decode_tok_s_stddev": 16.82200186063559,
    "e2e_max_ms": 43626.68369093444,
    "e2e_mean_ci95_high_ms": 33526.79808139783,
    "e2e_mean_ci95_low_ms": 17559.210713091485,
    "e2e_mean_ms": 25543.00439724466,
    "e2e_min_ms": 7742.640117998235,
    "e2e_p25_ms": 13923.70419704821,
    "e2e_p50_ms": 24881.578658008948,
    "e2e_p75_ms": 29179.43447898142,
    "e2e_p90_ms": 43626.68369093444,
    "e2e_p95_ms": 43626.68369093444,
    "e2e_p99_ms": 43626.68369093444,
    "e2e_samples": 8,
    "e2e_stddev_ms": 11521.213578895997,
    "effective_prefill_tok_s_max": 4241.55915357036,
    "effective_prefill_tok_s_mean": 3092.23069662452,
    "effective_prefill_tok_s_mean_ci95_high": 4017.9079377679473,
    "effective_prefill_tok_s_mean_ci95_low": 2166.553455481093,
    "effective_prefill_tok_s_min": 1085.5643265822866,
    "effective_prefill_tok_s_p25": 1418.0918203139013,
    "effective_prefill_tok_s_p50": 3375.0576030755806,
    "effective_prefill_tok_s_p75": 4152.990921959005,
    "effective_prefill_tok_s_p90": 4241.55915357036,
    "effective_prefill_tok_s_p95": 4241.55915357036,
    "effective_prefill_tok_s_p99": 4241.55915357036,
    "effective_prefill_tok_s_samples": 8,
    "effective_prefill_tok_s_stddev": 1335.821743678719,
    "generation_max_ms": 20334.65079998132,
    "generation_mean_ci95_high_ms": 19609.85174458406,
    "generation_mean_ci95_low_ms": 10336.812589632562,
    "generation_mean_ms": 14973.33216710831,
    "generation_min_ms": 1764.8134010378271,
    "generation_p25_ms": 7962.924426072277,
    "generation_p50_ms": 17464.034666889347,
    "generation_p75_ms": 18872.42940394208,
    "generation_p90_ms": 20334.65079998132,
    "generation_p95_ms": 20334.65079998132,
    "generation_p99_ms": 20334.65079998132,
    "generation_samples": 8,
    "generation_stddev_ms": 6690.845784361812,
    "mean_inter_token_latency_ms_max": 155.31964526388577,
    "mean_inter_token_latency_ms_mean": 105.2351975651948,
    "mean_inter_token_latency_ms_mean_ci95_high": 138.27075244862704,
    "mean_inter_token_latency_ms_mean_ci95_low": 72.19964268176254,
    "mean_inter_token_latency_ms_min": 17.826397990281084,
    "mean_inter_token_latency_ms_p25": 65.80929277745683,
    "mean_inter_token_latency_ms_p50": 112.34613701647137,
    "mean_inter_token_latency_ms_p75": 144.33086501561445,
    "mean_inter_token_latency_ms_p90": 155.31964526388577,
    "mean_inter_token_latency_ms_p95": 155.31964526388577,
    "mean_inter_token_latency_ms_p99": 155.31964526388577,
    "mean_inter_token_latency_ms_samples": 8,
    "mean_inter_token_latency_ms_stddev": 47.67278546599044,
    "mean_time_per_output_token_max_ms": 155.31964526388577,
    "mean_time_per_output_token_mean_ci95_high_ms": 138.27075244862704,
    "mean_time_per_output_token_mean_ci95_low_ms": 72.19964268176254,
    "mean_time_per_output_token_mean_ms": 105.2351975651948,
    "mean_time_per_output_token_min_ms": 17.826397990281084,
    "mean_time_per_output_token_p25_ms": 65.80929277745683,
    "mean_time_per_output_token_p50_ms": 112.34613701647137,
    "mean_time_per_output_token_p75_ms": 144.33086501561445,
    "mean_time_per_output_token_p90_ms": 155.31964526388577,
    "mean_time_per_output_token_p95_ms": 155.31964526388577,
    "mean_time_per_output_token_p99_ms": 155.31964526388577,
    "mean_time_per_output_token_samples": 8,
    "mean_time_per_output_token_stddev_ms": 47.67278546599044,
    "output_token_sources": {
      "usage": 8
    },
    "output_tokens": 1130,
    "prefix_cache_hit_avg": null,
    "prompt_token_samples": 8,
    "prompt_tokens": 202271,
    "prompt_tokens_max": 25285.0,
    "prompt_tokens_mean": 25283.875,
    "prompt_tokens_mean_ci95_high": 25284.655272388336,
    "prompt_tokens_mean_ci95_low": 25283.094727611664,
    "prompt_tokens_min": 25282.0,
    "prompt_tokens_p25": 25283.0,
    "prompt_tokens_p50": 25284.0,
    "prompt_tokens_p75": 25285.0,
    "prompt_tokens_p90": 25285.0,
    "prompt_tokens_p95": 25285.0,
    "prompt_tokens_p99": 25285.0,
    "prompt_tokens_samples": 8,
    "prompt_tokens_stddev": 1.1259916264596033,
    "reasoning_chunks": 823,
    "request_canary_passed": 8,
    "request_canary_samples": 8,
    "throughput_performance_eligible": true,
    "throughput_tok_s": 21.996275007218443,
    "time_to_first_output_max_ms": 23292.032890953124,
    "time_to_first_output_mean_ci95_high_ms": 15157.733066162185,
    "time_to_first_output_mean_ci95_low_ms": 5981.6113941105095,
    "time_to_first_output_mean_ms": 10569.672230136348,
    "time_to_first_output_min_ms": 5960.7797709759325,
    "time_to_first_output_p25_ms": 5977.826716960408,
    "time_to_first_output_p50_ms": 6201.711195986718,
    "time_to_first_output_p75_ms": 11715.399812092073,
    "time_to_first_output_p90_ms": 23292.032890953124,
    "time_to_first_output_p95_ms": 23292.032890953124,
    "time_to_first_output_p99_ms": 23292.032890953124,
    "time_to_first_output_samples": 8,
    "time_to_first_output_stddev_ms": 6620.916182959777,
    "tpot_max_ms": 155.31964526388577,
    "tpot_mean_ci95_high_ms": 138.27075244862704,
    "tpot_mean_ci95_low_ms": 72.19964268176254,
    "tpot_mean_ms": 105.2351975651948,
    "tpot_min_ms": 17.826397990281084,
    "tpot_p25_ms": 65.80929277745683,
    "tpot_p50_ms": 112.34613701647137,
    "tpot_p75_ms": 144.33086501561445,
    "tpot_p90_ms": 155.31964526388577,
    "tpot_p95_ms": 155.31964526388577,
    "tpot_p99_ms": 155.31964526388577,
    "tpot_samples": 8,
    "tpot_stddev_ms": 47.67278546599044,
    "ttft_max_ms": 43113.76795498654,
    "ttft_mean_ci95_high_ms": 30354.141793471375,
    "ttft_mean_ci95_low_ms": 15922.207614993364,
    "ttft_mean_ms": 23138.17470423237,
    "ttft_min_ms": 7327.906672959216,
    "ttft_p25_ms": 13408.903065952472,
    "ttft_p50_ms": 24369.76901604794,
    "ttft_p75_ms": 24934.718614909798,
    "ttft_p90_ms": 43113.76795498654,
    "ttft_p95_ms": 43113.76795498654,
    "ttft_p99_ms": 43113.76795498654,
    "ttft_samples": 8,
    "ttft_stddev_ms": 10413.182166571134
  },
  "model": "glm53-flash-exl3-4bpw-v84-fp8-tp2-c4-327k-r9-vision8-nospec",
  "performance_eligible": true,
  "prompt_cache_mode": "unique",
  "request_canaries": true,
  "request_timings": [
    {
      "content_chunks": 27,
      "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": 6.43833559045937,
      "e2e_ms": 24881.578658008948,
      "effective_prefill_tok_s": 4152.990921959005,
      "generation_ms": 18793.67707693018,
      "mean_inter_token_latency_ms": 155.31964526388577,
      "mean_time_per_output_token_ms": 155.31964526388577,
      "output_token_source": "usage",
      "output_tokens": 122,
      "planned_context_tokens": 32768,
      "prompt_tokens": 25283,
      "reasoning_chunks": 90,
      "request_canary": {
        "capture_complete": true,
        "foreign_marker_count": 0,
        "marker": "ANVIL_REQ_1910_00000",
        "marker_at_start": true,
        "observed_marker_count": 1,
        "passed": true
      },
      "request_index": 0,
      "time_to_first_output_ms": 6087.901581078768,
      "tpot_ms": 155.31964526388577,
      "ttft_ms": 24369.76901604794,
      "visible_content_truncated": false,
      "visible_generation_ms": 511.8096419610083
    },
    {
      "content_chunks": 27,
      "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": 6.928524954740727,
      "e2e_ms": 29179.43447898142,
      "effective_prefill_tok_s": 2158.184987754585,
      "generation_ms": 17464.034666889347,
      "mean_inter_token_latency_ms": 144.33086501561445,
      "mean_time_per_output_token_ms": 144.33086501561445,
      "output_token_source": "usage",
      "output_tokens": 122,
      "planned_context_tokens": 32768,
      "prompt_tokens": 25284,
      "reasoning_chunks": 90,
      "request_canary": {
        "capture_complete": true,
        "foreign_marker_count": 0,
        "marker": "ANVIL_REQ_1910_00001",
        "marker_at_start": true,
        "observed_marker_count": 1,
        "passed": true
      },
      "request_index": 1,
      "time_to_first_output_ms": 11715.399812092073,
      "tpot_ms": 144.33086501561445,
      "ttft_ms": 24597.238133079372,
      "visible_content_truncated": false,
      "visible_generation_ms": 4582.196345902048
    },
    {
      "content_chunks": 27,
      "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": 6.676405951937544,
      "e2e_ms": 36702.727581956424,
      "effective_prefill_tok_s": 1418.0918203139013,
      "generation_ms": 18872.42940394208,
      "mean_inter_token_latency_ms": 149.78118574557206,
      "mean_time_per_output_token_ms": 149.78118574557206,
      "output_token_source": "usage",
      "output_tokens": 127,
      "planned_context_tokens": 32768,
      "prompt_tokens": 25285,
      "reasoning_chunks": 94,
      "request_canary": {
        "capture_complete": true,
        "foreign_marker_count": 0,
        "marker": "ANVIL_REQ_1910_00002",
        "marker_at_start": true,
        "observed_marker_count": 1,
        "passed": true
      },
      "request_index": 2,
      "time_to_first_output_ms": 17830.298178014345,
      "tpot_ms": 149.78118574557206,
      "ttft_ms": 24934.718614909798,
      "visible_content_truncated": false,
      "visible_generation_ms": 11768.008967046626
    },
    {
      "content_chunks": 27,
      "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": 8.901062613780724,
      "e2e_ms": 43626.68369093444,
      "effective_prefill_tok_s": 1085.5643265822866,
      "generation_ms": 20334.65079998132,
      "mean_inter_token_latency_ms": 112.34613701647137,
      "mean_time_per_output_token_ms": 112.34613701647137,
      "output_token_source": "usage",
      "output_tokens": 182,
      "planned_context_tokens": 32768,
      "prompt_tokens": 25285,
      "reasoning_chunks": 133,
      "request_canary": {
        "capture_complete": true,
        "foreign_marker_count": 0,
        "marker": "ANVIL_REQ_1910_00003",
        "marker_at_start": true,
        "observed_marker_count": 1,
        "passed": true
      },
      "request_index": 3,
      "time_to_first_output_ms": 23292.032890953124,
      "tpot_ms": 112.34613701647137,
      "ttft_ms": 43113.76795498654,
      "visible_content_truncated": false,
      "visible_generation_ms": 512.9157359478995
    },
    {
      "content_chunks": 27,
      "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": 8.716046365979,
      "e2e_ms": 26164.890910033137,
      "effective_prefill_tok_s": 4077.1005293446356,
      "generation_ms": 19963.17971404642,
      "mean_inter_token_latency_ms": 114.7309178968185,
      "mean_time_per_output_token_ms": 114.7309178968185,
      "output_token_source": "usage",
      "output_tokens": 175,
      "planned_context_tokens": 32768,
      "prompt_tokens": 25285,
      "reasoning_chunks": 126,
      "request_canary": {
        "capture_complete": true,
        "foreign_marker_count": 0,
        "marker": "ANVIL_REQ_1910_00004",
        "marker_at_start": true,
        "observed_marker_count": 1,
        "passed": true
      },
      "request_index": 4,
      "time_to_first_output_ms": 6201.711195986718,
      "tpot_ms": 114.7309178968185,
      "ttft_ms": 25669.02820195537,
      "visible_content_truncated": false,
      "visible_generation_ms": 495.862708077766
    },
    {
      "content_chunks": 27,
      "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": 12.234340649698371,
      "e2e_ms": 22122.37554299645,
      "effective_prefill_tok_s": 3375.0576030755806,
      "generation_ms": 14630.947847967036,
      "mean_inter_token_latency_ms": 81.73713881545831,
      "mean_time_per_output_token_ms": 81.73713881545831,
      "output_token_source": "usage",
      "output_tokens": 180,
      "planned_context_tokens": 32768,
      "prompt_tokens": 25284,
      "reasoning_chunks": 131,
      "request_canary": {
        "capture_complete": true,
        "foreign_marker_count": 0,
        "marker": "ANVIL_REQ_1910_00005",
        "marker_at_start": true,
        "observed_marker_count": 1,
        "passed": true
      },
      "request_index": 5,
      "time_to_first_output_ms": 7491.427695029415,
      "tpot_ms": 81.73713881545831,
      "ttft_ms": 21684.065973968245,
      "visible_content_truncated": false,
      "visible_generation_ms": 438.30956902820617
    },
    {
      "content_chunks": 27,
      "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": 15.195422375706686,
      "e2e_ms": 13923.70419704821,
      "effective_prefill_tok_s": 4241.55915357036,
      "generation_ms": 7962.924426072277,
      "mean_inter_token_latency_ms": 65.80929277745683,
      "mean_time_per_output_token_ms": 65.80929277745683,
      "output_token_source": "usage",
      "output_tokens": 122,
      "planned_context_tokens": 32768,
      "prompt_tokens": 25283,
      "reasoning_chunks": 90,
      "request_canary": {
        "capture_complete": true,
        "foreign_marker_count": 0,
        "marker": "ANVIL_REQ_1910_00006",
        "marker_at_start": true,
        "observed_marker_count": 1,
        "passed": true
      },
      "request_index": 6,
      "time_to_first_output_ms": 5960.7797709759325,
      "tpot_ms": 65.80929277745683,
      "ttft_ms": 13408.903065952472,
      "visible_content_truncated": false,
      "visible_generation_ms": 514.8011310957372
    },
    {
      "content_chunks": 27,
      "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": 56.096582189245304,
      "e2e_ms": 7742.640117998235,
      "effective_prefill_tok_s": 4229.2962303958075,
      "generation_ms": 1764.8134010378271,
      "mean_inter_token_latency_ms": 17.826397990281084,
      "mean_time_per_output_token_ms": 17.826397990281084,
      "output_token_source": "usage",
      "output_tokens": 100,
      "planned_context_tokens": 32768,
      "prompt_tokens": 25282,
      "reasoning_chunks": 69,
      "request_canary": {
        "capture_complete": true,
        "foreign_marker_count": 0,
        "marker": "ANVIL_REQ_1910_00007",
        "marker_at_start": true,
        "observed_marker_count": 1,
        "passed": true
      },
      "request_index": 7,
      "time_to_first_output_ms": 5977.826716960408,
      "tpot_ms": 17.826397990281084,
      "ttft_ms": 7327.906672959216,
      "visible_content_truncated": false,
      "visible_generation_ms": 414.733445039019
    }
  ],
  "requests": 8,
  "response_words": 16,
  "responses_received": 8,
  "run_id": "benchmark-20260919T111619Z",
  "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": 0,
  "source_recipe": null,
  "started_at": "2026-09-19T11:16:19Z",
  "started_at_unix_ns": 1789816579577054467,
  "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": 51372.33461707365
}
