{
  "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": "2026-10-05T11:43:06Z",
  "finished_at_unix_ns": 1791200586664844800,
  "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": 519.0,
    "completion_tokens_mean": 380.5,
    "completion_tokens_mean_ci95_high": 462.3140563310079,
    "completion_tokens_mean_ci95_low": 298.6859436689921,
    "completion_tokens_min": 274.0,
    "completion_tokens_p25": 281.0,
    "completion_tokens_p50": 321.0,
    "completion_tokens_p75": 455.0,
    "completion_tokens_p90": 519.0,
    "completion_tokens_p95": 519.0,
    "completion_tokens_p99": 519.0,
    "completion_tokens_samples": 6,
    "completion_tokens_stddev": 102.24627132565764,
    "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": 250.04892366510725,
    "decode_tok_s_mean": 235.30873496370418,
    "decode_tok_s_mean_ci95_high": 244.88803846533187,
    "decode_tok_s_mean_ci95_low": 225.7294314620765,
    "decode_tok_s_min": 216.56510552809334,
    "decode_tok_s_p25": 229.24850371168577,
    "decode_tok_s_p50": 231.96481822725846,
    "decode_tok_s_p75": 244.27235024908848,
    "decode_tok_s_p90": 250.04892366510725,
    "decode_tok_s_p95": 250.04892366510725,
    "decode_tok_s_p99": 250.04892366510725,
    "decode_tok_s_samples": 6,
    "decode_tok_s_stddev": 11.97163554604285,
    "e2e_max_ms": 10137.64810000066,
    "e2e_mean_ci95_high_ms": 10393.49822772722,
    "e2e_mean_ci95_low_ms": 3732.780838940023,
    "e2e_mean_ms": 7063.139533333621,
    "e2e_min_ms": 1562.8556000010576,
    "e2e_p25_ms": 1861.3442999994732,
    "e2e_p50_ms": 9133.986500000901,
    "e2e_p75_ms": 10089.705200000026,
    "e2e_p90_ms": 10137.64810000066,
    "e2e_p95_ms": 10137.64810000066,
    "e2e_p99_ms": 10137.64810000066,
    "e2e_samples": 6,
    "e2e_stddev_ms": 4162.081358013212,
    "effective_prefill_tok_s_max": 1047830.055619343,
    "effective_prefill_tok_s_mean": 208981.07614165635,
    "effective_prefill_tok_s_mean_ci95_high": 542149.8880280248,
    "effective_prefill_tok_s_mean_ci95_low": -124187.73574471209,
    "effective_prefill_tok_s_min": 7641.654964622031,
    "effective_prefill_tok_s_p25": 7680.192685719447,
    "effective_prefill_tok_s_p50": 7773.402377361303,
    "effective_prefill_tok_s_p75": 175186.0721351123,
    "effective_prefill_tok_s_p90": 1047830.055619343,
    "effective_prefill_tok_s_p95": 1047830.055619343,
    "effective_prefill_tok_s_p99": 1047830.055619343,
    "effective_prefill_tok_s_samples": 6,
    "effective_prefill_tok_s_stddev": 416374.2792504682,
    "generation_max_ms": 2071.594599998207,
    "generation_mean_ci95_high_ms": 1931.2941278232697,
    "generation_mean_ci95_low_ms": 1287.4176055085081,
    "generation_mean_ms": 1609.355866665889,
    "generation_min_ms": 1117.6049999994575,
    "generation_p25_ms": 1207.079600000725,
    "generation_p50_ms": 1477.6156999978411,
    "generation_p75_ms": 1980.3836999999476,
    "generation_p90_ms": 2071.594599998207,
    "generation_p95_ms": 2071.594599998207,
    "generation_p99_ms": 2071.594599998207,
    "generation_samples": 6,
    "generation_stddev_ms": 402.3390145481918,
    "mean_inter_token_latency_ms_max": 4.617549062493254,
    "mean_inter_token_latency_ms_mean": 4.259099867239761,
    "mean_inter_token_latency_ms_mean_ci95_high": 4.4360136767270975,
    "mean_inter_token_latency_ms_mean_ci95_low": 4.082186057752424,
    "mean_inter_token_latency_ms_min": 3.999217374513913,
    "mean_inter_token_latency_ms_p25": 4.093791208789222,
    "mean_inter_token_latency_ms_p50": 4.1709643518498964,
    "mean_inter_token_latency_ms_p75": 4.362078634361118,
    "mean_inter_token_latency_ms_p90": 4.617549062493254,
    "mean_inter_token_latency_ms_p95": 4.617549062493254,
    "mean_inter_token_latency_ms_p99": 4.617549062493254,
    "mean_inter_token_latency_ms_samples": 6,
    "mean_inter_token_latency_ms_stddev": 0.22109620494690177,
    "mean_time_per_output_token_max_ms": 4.617549062493254,
    "mean_time_per_output_token_mean_ci95_high_ms": 4.4360136767270975,
    "mean_time_per_output_token_mean_ci95_low_ms": 4.082186057752424,
    "mean_time_per_output_token_mean_ms": 4.259099867239761,
    "mean_time_per_output_token_min_ms": 3.999217374513913,
    "mean_time_per_output_token_p25_ms": 4.093791208789222,
    "mean_time_per_output_token_p50_ms": 4.1709643518498964,
    "mean_time_per_output_token_p75_ms": 4.362078634361118,
    "mean_time_per_output_token_p90_ms": 4.617549062493254,
    "mean_time_per_output_token_p95_ms": 4.617549062493254,
    "mean_time_per_output_token_p99_ms": 4.617549062493254,
    "mean_time_per_output_token_samples": 6,
    "mean_time_per_output_token_stddev_ms": 0.22109620494690177,
    "output_token_sources": {
      "usage": 6
    },
    "output_tokens": 2283,
    "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": 497,
    "request_canary_passed": 6,
    "request_canary_samples": 6,
    "throughput_performance_eligible": true,
    "throughput_tok_s": 53.862812534781895,
    "time_to_first_output_max_ms": 8157.264400000713,
    "time_to_first_output_mean_ci95_high_ms": 8706.543300030713,
    "time_to_first_output_mean_ci95_low_ms": 2201.024033304752,
    "time_to_first_output_mean_ms": 5453.783666667732,
    "time_to_first_output_min_ms": 59.487700000318,
    "time_to_first_output_p25_ms": 355.7760000003327,
    "time_to_first_output_p50_ms": 8016.381500001444,
    "time_to_first_output_p75_ms": 8115.6818000017665,
    "time_to_first_output_p90_ms": 8157.264400000713,
    "time_to_first_output_p95_ms": 8157.264400000713,
    "time_to_first_output_p99_ms": 8157.264400000713,
    "time_to_first_output_samples": 6,
    "time_to_first_output_stddev_ms": 4065.1027335009135,
    "tpot_max_ms": 4.617549062493254,
    "tpot_mean_ci95_high_ms": 4.4360136767270975,
    "tpot_mean_ci95_low_ms": 4.082186057752424,
    "tpot_mean_ms": 4.259099867239761,
    "tpot_min_ms": 3.999217374513913,
    "tpot_p25_ms": 4.093791208789222,
    "tpot_p50_ms": 4.1709643518498964,
    "tpot_p75_ms": 4.362078634361118,
    "tpot_p90_ms": 4.617549062493254,
    "tpot_p95_ms": 4.617549062493254,
    "tpot_p99_ms": 4.617549062493254,
    "tpot_samples": 6,
    "tpot_stddev_ms": 0.22109620494690177,
    "ttft_max_ms": 10017.000400002871,
    "ttft_mean_ci95_high_ms": 10263.631167790554,
    "ttft_mean_ci95_low_ms": 3603.26923220975,
    "ttft_mean_ms": 6933.450200000152,
    "ttft_min_ms": 1418.298299999151,
    "ttft_p25_ms": 1752.577299997938,
    "ttft_p50_ms": 8953.168000000005,
    "ttft_p75_ms": 9974.789100000635,
    "ttft_p90_ms": 10017.000400002871,
    "ttft_p95_ms": 10017.000400002871,
    "ttft_p99_ms": 10017.000400002871,
    "ttft_samples": 6,
    "ttft_stddev_ms": 4161.859246027728
  },
  "model": "qwen38-swift15-ninfer-dflash2-c4-262k",
  "performance_eligible": true,
  "prompt_cache_mode": "unique",
  "request_canaries": true,
  "request_timings": [
    {
      "content_chunks": 9,
      "controlled_output": null,
      "decode_tok_s": 231.96481822725846,
      "e2e_ms": 1562.8556000010576,
      "effective_prefill_tok_s": 175186.0721351123,
      "generation_ms": 1207.079600000725,
      "mean_inter_token_latency_ms": 4.310998571431161,
      "mean_time_per_output_token_ms": 4.310998571431161,
      "output_token_source": "usage",
      "output_tokens": 281,
      "planned_context_tokens": 65536,
      "prompt_tokens": 62327,
      "reasoning_chunks": 59,
      "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": 355.7760000003327,
      "tpot_ms": 4.310998571431161,
      "ttft_ms": 1418.298299999151,
      "visible_content_truncated": false,
      "visible_generation_ms": 144.5573000019067
    },
    {
      "content_chunks": 5,
      "controlled_output": null,
      "decode_tok_s": 250.04892366510725,
      "e2e_ms": 10089.705200000026,
      "effective_prefill_tok_s": 7773.402377361303,
      "generation_ms": 2071.594599998207,
      "mean_inter_token_latency_ms": 3.999217374513913,
      "mean_time_per_output_token_ms": 3.999217374513913,
      "output_token_source": "usage",
      "output_tokens": 519,
      "planned_context_tokens": 65536,
      "prompt_tokens": 62328,
      "reasoning_chunks": 112,
      "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": 8018.110600001819,
      "tpot_ms": 3.999217374513913,
      "ttft_ms": 10017.000400002871,
      "visible_content_truncated": false,
      "visible_generation_ms": 72.70479999715462
    },
    {
      "content_chunks": 11,
      "controlled_output": null,
      "decode_tok_s": 244.27235024908848,
      "e2e_ms": 9133.986500000901,
      "effective_prefill_tok_s": 7775.079067779992,
      "generation_ms": 1117.6049999994575,
      "mean_inter_token_latency_ms": 4.093791208789222,
      "mean_time_per_output_token_ms": 4.093791208789222,
      "output_token_source": "usage",
      "output_tokens": 274,
      "planned_context_tokens": 65536,
      "prompt_tokens": 62328,
      "reasoning_chunks": 53,
      "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": 8016.381500001444,
      "tpot_ms": 4.093791208789222,
      "ttft_ms": 8953.168000000005,
      "visible_content_truncated": false,
      "visible_generation_ms": 180.81850000089617
    },
    {
      "content_chunks": 7,
      "controlled_output": null,
      "decode_tok_s": 216.56510552809334,
      "e2e_ms": 9593.297499999608,
      "effective_prefill_tok_s": 7680.192685719447,
      "generation_ms": 1477.6156999978411,
      "mean_inter_token_latency_ms": 4.617549062493254,
      "mean_time_per_output_token_ms": 4.617549062493254,
      "output_token_source": "usage",
      "output_tokens": 321,
      "planned_context_tokens": 65536,
      "prompt_tokens": 62330,
      "reasoning_chunks": 77,
      "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": 8115.6818000017665,
      "tpot_ms": 4.617549062493254,
      "ttft_ms": 9484.868100000313,
      "visible_content_truncated": false,
      "visible_generation_ms": 108.42939999929513
    },
    {
      "content_chunks": 7,
      "controlled_output": null,
      "decode_tok_s": 239.75270840099182,
      "e2e_ms": 1861.3442999994732,
      "effective_prefill_tok_s": 1047830.055619343,
      "generation_ms": 1801.8565999991552,
      "mean_inter_token_latency_ms": 4.1709643518498964,
      "mean_time_per_output_token_ms": 4.1709643518498964,
      "output_token_source": "usage",
      "output_tokens": 433,
      "planned_context_tokens": 65536,
      "prompt_tokens": 62333,
      "reasoning_chunks": 95,
      "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": 59.487700000318,
      "tpot_ms": 4.1709643518498964,
      "ttft_ms": 1752.577299997938,
      "visible_content_truncated": false,
      "visible_generation_ms": 108.76700000153505
    },
    {
      "content_chunks": 10,
      "controlled_output": null,
      "decode_tok_s": 229.24850371168577,
      "e2e_ms": 10137.64810000066,
      "effective_prefill_tok_s": 7641.654964622031,
      "generation_ms": 1980.3836999999476,
      "mean_inter_token_latency_ms": 4.362078634361118,
      "mean_time_per_output_token_ms": 4.362078634361118,
      "output_token_source": "usage",
      "output_tokens": 455,
      "planned_context_tokens": 65536,
      "prompt_tokens": 62335,
      "reasoning_chunks": 101,
      "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": 8157.264400000713,
      "tpot_ms": 4.362078634361118,
      "ttft_ms": 9974.789100000635,
      "visible_content_truncated": false,
      "visible_generation_ms": 162.85900000002584
    }
  ],
  "requests": 6,
  "response_words": 0,
  "responses_received": 6,
  "run_id": "benchmark-20261005T114224Z",
  "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": "2026-10-05T11:42:24Z",
  "started_at_unix_ns": 1791200544279383700,
  "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": 42385.45840000006
}