{
  "base_url": "[operator identity redacted]",
  "clock_domain_id": null,
  "completed": 6,
  "concurrency": 1,
  "configuration_fingerprint": null,
  "context_distribution": {
    "65536": 6
  },
  "context_seed": 1717,
  "context_tokens": 65536,
  "controlled_output_policy": "observe",
  "engine": "ninfer",
  "failed": 0,
  "failures": [],
  "finished_at": "2026-10-05T11:36:08Z",
  "finished_at_unix_ns": 1791200168588395400,
  "gpu": "rtx5090",
  "manifest": null,
  "max_context_tokens": 262144,
  "max_tokens": 8192,
  "measurement_protocol": "capacity-v4-reasoning",
  "metric_population": {
    "aggregate_throughput_eligible": true,
    "attempted_requests": 6,
    "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": 6,
      "e2e": 6,
      "effective_prefill_tok_s": 6,
      "tpot": 6,
      "ttft": 6
    },
    "performance_eligible_requests": 6,
    "successful_requests": 6
  },
  "metrics": {
    "completion_tokens_max": 1541.0,
    "completion_tokens_mean": 485.1666666666667,
    "completion_tokens_mean_ci95_high": 902.2008788416925,
    "completion_tokens_mean_ci95_low": 68.13245449164083,
    "completion_tokens_min": 185.0,
    "completion_tokens_p25": 253.0,
    "completion_tokens_p50": 265.0,
    "completion_tokens_p75": 383.0,
    "completion_tokens_p90": 1541.0,
    "completion_tokens_p95": 1541.0,
    "completion_tokens_p99": 1541.0,
    "completion_tokens_samples": 6,
    "completion_tokens_stddev": 521.1841964859129,
    "content_chunks": 49,
    "content_chunks_s": null,
    "controlled_output_exact_adherence": 0,
    "controlled_output_exact_adherence_rate": null,
    "controlled_output_samples": 0,
    "controlled_output_unobservable": 0,
    "decode_tok_s_max": 265.4504622811552,
    "decode_tok_s_mean": 235.1343474577394,
    "decode_tok_s_mean_ci95_high": 249.19904320270254,
    "decode_tok_s_mean_ci95_low": 221.06965171277628,
    "decode_tok_s_min": 217.58443602396997,
    "decode_tok_s_p25": 218.22228849690063,
    "decode_tok_s_p50": 232.7502654846772,
    "decode_tok_s_p75": 239.93284061666253,
    "decode_tok_s_p90": 265.4504622811552,
    "decode_tok_s_p95": 265.4504622811552,
    "decode_tok_s_p99": 265.4504622811552,
    "decode_tok_s_samples": 6,
    "decode_tok_s_stddev": 17.577208144210935,
    "e2e_max_ms": 13788.482299998577,
    "e2e_mean_ci95_high_ms": 11256.298733029222,
    "e2e_mean_ci95_low_ms": 3406.479800302605,
    "e2e_mean_ms": 7331.389266665913,
    "e2e_min_ms": 1259.8532000010891,
    "e2e_p25_ms": 1677.395599999727,
    "e2e_p50_ms": 8637.916799998493,
    "e2e_p75_ms": 9644.11669999754,
    "e2e_p90_ms": 13788.482299998577,
    "e2e_p95_ms": 13788.482299998577,
    "e2e_p99_ms": 13788.482299998577,
    "e2e_samples": 6,
    "e2e_stddev_ms": 4905.11504041302,
    "effective_prefill_tok_s_max": 593266.7982657194,
    "effective_prefill_tok_s_mean": 131688.44423660848,
    "effective_prefill_tok_s_mean_ci95_high": 319524.7077547941,
    "effective_prefill_tok_s_mean_ci95_low": -56147.81928157713,
    "effective_prefill_tok_s_min": 7788.203450847613,
    "effective_prefill_tok_s_p25": 7803.908720335857,
    "effective_prefill_tok_s_p50": 7909.380589082243,
    "effective_prefill_tok_s_p75": 165432.82819853612,
    "effective_prefill_tok_s_p90": 593266.7982657194,
    "effective_prefill_tok_s_p95": 593266.7982657194,
    "effective_prefill_tok_s_p99": 593266.7982657194,
    "effective_prefill_tok_s_samples": 6,
    "effective_prefill_tok_s_stddev": 234746.42898495597,
    "generation_max_ms": 5801.459099999192,
    "generation_mean_ci95_high_ms": 3484.14268613866,
    "generation_mean_ci95_low_ms": 440.9415138602558,
    "generation_mean_ms": 1962.5420999994578,
    "generation_min_ms": 776.81119999761,
    "generation_p25_ms": 1100.307900000189,
    "generation_p50_ms": 1154.7858000012639,
    "generation_p75_ms": 1641.2440999993123,
    "generation_p90_ms": 5801.459099999192,
    "generation_p95_ms": 5801.459099999192,
    "generation_p99_ms": 5801.459099999192,
    "generation_samples": 6,
    "generation_stddev_ms": 1901.6046063065544,
    "mean_inter_token_latency_ms_max": 4.595916961127845,
    "mean_inter_token_latency_ms_mean": 4.271944167720583,
    "mean_inter_token_latency_ms_mean_ci95_high": 4.517305249196211,
    "mean_inter_token_latency_ms_mean_ci95_low": 4.026583086244955,
    "mean_inter_token_latency_ms_min": 3.7671812337657093,
    "mean_inter_token_latency_ms_p25": 4.167832954546171,
    "mean_inter_token_latency_ms_p50": 4.22179999998701,
    "mean_inter_token_latency_ms_p75": 4.582483333338349,
    "mean_inter_token_latency_ms_p90": 4.595916961127845,
    "mean_inter_token_latency_ms_p95": 4.595916961127845,
    "mean_inter_token_latency_ms_p99": 4.595916961127845,
    "mean_inter_token_latency_ms_samples": 6,
    "mean_inter_token_latency_ms_stddev": 0.3066374756901724,
    "mean_time_per_output_token_max_ms": 4.595916961127845,
    "mean_time_per_output_token_mean_ci95_high_ms": 4.517305249196211,
    "mean_time_per_output_token_mean_ci95_low_ms": 4.026583086244955,
    "mean_time_per_output_token_mean_ms": 4.271944167720583,
    "mean_time_per_output_token_min_ms": 3.7671812337657093,
    "mean_time_per_output_token_p25_ms": 4.167832954546171,
    "mean_time_per_output_token_p50_ms": 4.22179999998701,
    "mean_time_per_output_token_p75_ms": 4.582483333338349,
    "mean_time_per_output_token_p90_ms": 4.595916961127845,
    "mean_time_per_output_token_p95_ms": 4.595916961127845,
    "mean_time_per_output_token_p99_ms": 4.595916961127845,
    "mean_time_per_output_token_samples": 6,
    "mean_time_per_output_token_stddev_ms": 0.3066374756901724,
    "output_token_sources": {
      "usage": 6
    },
    "output_tokens": 2911,
    "prefix_cache_hit_avg": 0.33330124579846615,
    "prompt_token_samples": 6,
    "prompt_tokens": 373981,
    "prompt_tokens_max": 62335.0,
    "prompt_tokens_mean": 62330.166666666664,
    "prompt_tokens_mean_ci95_high": 62332.718014894126,
    "prompt_tokens_mean_ci95_low": 62327.6153184392,
    "prompt_tokens_min": 62327.0,
    "prompt_tokens_p25": 62328.0,
    "prompt_tokens_p50": 62328.0,
    "prompt_tokens_p75": 62333.0,
    "prompt_tokens_p90": 62335.0,
    "prompt_tokens_p95": 62335.0,
    "prompt_tokens_p99": 62335.0,
    "prompt_tokens_samples": 6,
    "prompt_tokens_stddev": 3.1885210782848317,
    "reasoning_chunks": 614,
    "request_canary_passed": 6,
    "request_canary_samples": 6,
    "throughput_performance_eligible": true,
    "throughput_tok_s": 66.16777392560181,
    "time_to_first_output_max_ms": 8002.872599998227,
    "time_to_first_output_mean_ci95_high_ms": 8548.236592503858,
    "time_to_first_output_mean_ci95_low_ms": 2189.4577408290543,
    "time_to_first_output_mean_ms": 5368.847166666455,
    "time_to_first_output_min_ms": 105.06739999982528,
    "time_to_first_output_p25_ms": 376.751100000547,
    "time_to_first_output_p50_ms": 7861.105600000883,
    "time_to_first_output_p75_ms": 7987.023199999385,
    "time_to_first_output_p90_ms": 8002.872599998227,
    "time_to_first_output_p95_ms": 8002.872599998227,
    "time_to_first_output_p99_ms": 8002.872599998227,
    "time_to_first_output_samples": 6,
    "time_to_first_output_stddev_ms": 3973.409074950006,
    "tpot_max_ms": 4.595916961127845,
    "tpot_mean_ci95_high_ms": 4.517305249196211,
    "tpot_mean_ci95_low_ms": 4.026583086244955,
    "tpot_mean_ms": 4.271944167720583,
    "tpot_min_ms": 3.7671812337657093,
    "tpot_p25_ms": 4.167832954546171,
    "tpot_p50_ms": 4.22179999998701,
    "tpot_p75_ms": 4.582483333338349,
    "tpot_p90_ms": 4.595916961127845,
    "tpot_p95_ms": 4.595916961127845,
    "tpot_p99_ms": 4.595916961127845,
    "tpot_samples": 6,
    "tpot_stddev_ms": 0.3066374756901724,
    "ttft_max_ms": 13643.54209999874,
    "ttft_mean_ci95_high_ms": 11131.885485915413,
    "ttft_mean_ci95_low_ms": 3271.256180751561,
    "ttft_mean_ms": 7201.570833333487,
    "ttft_min_ms": 1150.9905000020808,
    "ttft_p25_ms": 1496.525800001109,
    "ttft_p50_ms": 8510.93359999868,
    "ttft_p75_ms": 9553.487699999096,
    "ttft_p90_ms": 13643.54209999874,
    "ttft_p95_ms": 13643.54209999874,
    "ttft_p99_ms": 13643.54209999874,
    "ttft_samples": 6,
    "ttft_stddev_ms": 4911.870116025437
  },
  "model": "qwen38-swift15-ninfer-dflash2-c4-262k",
  "performance_eligible": true,
  "prompt_cache_mode": "unique",
  "request_canaries": true,
  "request_timings": [
    {
      "content_chunks": 11,
      "controlled_output": null,
      "decode_tok_s": 217.58443602396997,
      "e2e_ms": 1677.395599999727,
      "effective_prefill_tok_s": 165432.82819853612,
      "generation_ms": 1300.64449999918,
      "mean_inter_token_latency_ms": 4.595916961127845,
      "mean_time_per_output_token_ms": 4.595916961127845,
      "output_token_source": "usage",
      "output_tokens": 284,
      "planned_context_tokens": 65536,
      "prompt_tokens": 62327,
      "reasoning_chunks": 63,
      "request_canary": {
        "capture_complete": true,
        "foreign_marker_count": 0,
        "marker": "ANVIL_REQ_1717_00000",
        "marker_at_start": true,
        "observed_marker_count": 1,
        "passed": true
      },
      "request_index": 0,
      "time_to_first_output_ms": 376.751100000547,
      "tpot_ms": 4.595916961127845,
      "ttft_ms": 1496.525800001109,
      "visible_content_truncated": false,
      "visible_generation_ms": 180.86979999861796
    },
    {
      "content_chunks": 8,
      "controlled_output": null,
      "decode_tok_s": 239.93284061666253,
      "e2e_ms": 8980.571000000054,
      "effective_prefill_tok_s": 7909.380589082243,
      "generation_ms": 1100.307900000189,
      "mean_inter_token_latency_ms": 4.167832954546171,
      "mean_time_per_output_token_ms": 4.167832954546171,
      "output_token_source": "usage",
      "output_tokens": 265,
      "planned_context_tokens": 65536,
      "prompt_tokens": 62328,
      "reasoning_chunks": 55,
      "request_canary": {
        "capture_complete": true,
        "foreign_marker_count": 0,
        "marker": "ANVIL_REQ_1717_00001",
        "marker_at_start": true,
        "observed_marker_count": 1,
        "passed": true
      },
      "request_index": 1,
      "time_to_first_output_ms": 7880.2630999998655,
      "tpot_ms": 4.167832954546171,
      "ttft_ms": 8853.945300001214,
      "visible_content_truncated": false,
      "visible_generation_ms": 126.62569999883999
    },
    {
      "content_chunks": 6,
      "controlled_output": null,
      "decode_tok_s": 232.7502654846772,
      "e2e_ms": 9644.11669999754,
      "effective_prefill_tok_s": 7788.203450847613,
      "generation_ms": 1641.2440999993123,
      "mean_inter_token_latency_ms": 4.296450523558409,
      "mean_time_per_output_token_ms": 4.296450523558409,
      "output_token_source": "usage",
      "output_tokens": 383,
      "planned_context_tokens": 65536,
      "prompt_tokens": 62328,
      "reasoning_chunks": 87,
      "request_canary": {
        "capture_complete": true,
        "foreign_marker_count": 0,
        "marker": "ANVIL_REQ_1717_00002",
        "marker_at_start": true,
        "observed_marker_count": 1,
        "passed": true
      },
      "request_index": 2,
      "time_to_first_output_ms": 8002.872599998227,
      "tpot_ms": 4.296450523558409,
      "ttft_ms": 9553.487699999096,
      "visible_content_truncated": false,
      "visible_generation_ms": 90.62899999844376
    },
    {
      "content_chunks": 9,
      "controlled_output": null,
      "decode_tok_s": 265.4504622811552,
      "e2e_ms": 13788.482299998577,
      "effective_prefill_tok_s": 7803.908720335857,
      "generation_ms": 5801.459099999192,
      "mean_inter_token_latency_ms": 3.7671812337657093,
      "mean_time_per_output_token_ms": 3.7671812337657093,
      "output_token_source": "usage",
      "output_tokens": 1541,
      "planned_context_tokens": 65536,
      "prompt_tokens": 62330,
      "reasoning_chunks": 314,
      "request_canary": {
        "capture_complete": true,
        "foreign_marker_count": 0,
        "marker": "ANVIL_REQ_1717_00003",
        "marker_at_start": true,
        "observed_marker_count": 1,
        "passed": true
      },
      "request_index": 3,
      "time_to_first_output_ms": 7987.023199999385,
      "tpot_ms": 3.7671812337657093,
      "ttft_ms": 13643.54209999874,
      "visible_content_truncated": false,
      "visible_generation_ms": 144.9401999998372
    },
    {
      "content_chunks": 7,
      "controlled_output": null,
      "decode_tok_s": 218.22228849690063,
      "e2e_ms": 1259.8532000010891,
      "effective_prefill_tok_s": 593266.7982657194,
      "generation_ms": 1154.7858000012639,
      "mean_inter_token_latency_ms": 4.582483333338349,
      "mean_time_per_output_token_ms": 4.582483333338349,
      "output_token_source": "usage",
      "output_tokens": 253,
      "planned_context_tokens": 65536,
      "prompt_tokens": 62333,
      "reasoning_chunks": 58,
      "request_canary": {
        "capture_complete": true,
        "foreign_marker_count": 0,
        "marker": "ANVIL_REQ_1717_00004",
        "marker_at_start": true,
        "observed_marker_count": 1,
        "passed": true
      },
      "request_index": 4,
      "time_to_first_output_ms": 105.06739999982528,
      "tpot_ms": 4.582483333338349,
      "ttft_ms": 1150.9905000020808,
      "visible_content_truncated": false,
      "visible_generation_ms": 108.86269999900833
    },
    {
      "content_chunks": 8,
      "controlled_output": null,
      "decode_tok_s": 236.8657918430709,
      "e2e_ms": 8637.916799998493,
      "effective_prefill_tok_s": 7929.546195129728,
      "generation_ms": 776.81119999761,
      "mean_inter_token_latency_ms": 4.22179999998701,
      "mean_time_per_output_token_ms": 4.22179999998701,
      "output_token_source": "usage",
      "output_tokens": 185,
      "planned_context_tokens": 65536,
      "prompt_tokens": 62335,
      "reasoning_chunks": 37,
      "request_canary": {
        "capture_complete": true,
        "foreign_marker_count": 0,
        "marker": "ANVIL_REQ_1717_00005",
        "marker_at_start": true,
        "observed_marker_count": 1,
        "passed": true
      },
      "request_index": 5,
      "time_to_first_output_ms": 7861.105600000883,
      "tpot_ms": 4.22179999998701,
      "ttft_ms": 8510.93359999868,
      "visible_content_truncated": false,
      "visible_generation_ms": 126.98319999981322
    }
  ],
  "requests": 6,
  "response_words": 0,
  "responses_received": 6,
  "run_id": "benchmark-20261005T113524Z",
  "sampling": {
    "temperature": {
      "effective_request": 0.9,
      "requested": 0.9,
      "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": null,
    "shared_prefix_burst": false,
    "thinking_mode": "default"
  },
  "shared_prefix_tokens": 0,
  "source_recipe": null,
  "started_at": "2026-10-05T11:35:24Z",
  "started_at_unix_ns": 1791200124594166500,
  "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": 43994.22600000253
}