{
  "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": "2026-10-05T11:34:39Z",
  "finished_at_unix_ns": 1791200079066181600,
  "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": 627.0,
    "completion_tokens_mean": 384.0,
    "completion_tokens_mean_ci95_high": 509.9320972588006,
    "completion_tokens_mean_ci95_low": 258.0679027411994,
    "completion_tokens_min": 213.0,
    "completion_tokens_p25": 265.0,
    "completion_tokens_p50": 339.0,
    "completion_tokens_p75": 517.0,
    "completion_tokens_p90": 627.0,
    "completion_tokens_p95": 627.0,
    "completion_tokens_p99": 627.0,
    "completion_tokens_samples": 6,
    "completion_tokens_stddev": 157.38233700132935,
    "content_chunks": 47,
    "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": 255.2619173904583,
    "decode_tok_s_mean": 242.8673737520467,
    "decode_tok_s_mean_ci95_high": 249.60868869784272,
    "decode_tok_s_mean_ci95_low": 236.12605880625065,
    "decode_tok_s_min": 230.59615989474375,
    "decode_tok_s_p25": 236.9169867813405,
    "decode_tok_s_p50": 243.8923895041408,
    "decode_tok_s_p75": 246.43002725059497,
    "decode_tok_s_p90": 255.2619173904583,
    "decode_tok_s_p95": 255.2619173904583,
    "decode_tok_s_p99": 255.2619173904583,
    "decode_tok_s_samples": 6,
    "decode_tok_s_stddev": 8.424888679897087,
    "e2e_max_ms": 10556.770799998048,
    "e2e_mean_ci95_high_ms": 10064.660231990165,
    "e2e_mean_ci95_low_ms": 9087.306068009748,
    "e2e_mean_ms": 9575.983149999956,
    "e2e_min_ms": 8927.579600000172,
    "e2e_p25_ms": 9129.797699999472,
    "e2e_p50_ms": 9324.149000000034,
    "e2e_p75_ms": 10035.54190000068,
    "e2e_p90_ms": 10556.770799998048,
    "e2e_p95_ms": 10556.770799998048,
    "e2e_p99_ms": 10556.770799998048,
    "e2e_samples": 6,
    "e2e_stddev_ms": 610.7191325858317,
    "effective_prefill_tok_s_max": 8034.712536072016,
    "effective_prefill_tok_s_mean": 7806.7861580731915,
    "effective_prefill_tok_s_mean_ci95_high": 7950.519895543658,
    "effective_prefill_tok_s_mean_ci95_low": 7663.052420602725,
    "effective_prefill_tok_s_min": 7562.392070262102,
    "effective_prefill_tok_s_p25": 7697.855669861424,
    "effective_prefill_tok_s_p50": 7753.669666628207,
    "effective_prefill_tok_s_p75": 7993.599580355352,
    "effective_prefill_tok_s_p90": 8034.712536072016,
    "effective_prefill_tok_s_p95": 8034.712536072016,
    "effective_prefill_tok_s_p99": 8034.712536072016,
    "effective_prefill_tok_s_samples": 6,
    "effective_prefill_tok_s_stddev": 179.62975287030463,
    "generation_max_ms": 2564.4517000000633,
    "generation_mean_ci95_high_ms": 2127.828285466904,
    "generation_mean_ci95_low_ms": 1048.9129478675163,
    "generation_mean_ms": 1588.3706166672102,
    "generation_min_ms": 830.5194999993546,
    "generation_p25_ms": 1082.4446000005992,
    "generation_p50_ms": 1371.5861000018776,
    "generation_p75_ms": 2237.678199999209,
    "generation_p90_ms": 2564.4517000000633,
    "generation_p95_ms": 2564.4517000000633,
    "generation_p99_ms": 2564.4517000000633,
    "generation_samples": 6,
    "generation_stddev_ms": 674.1816461176405,
    "mean_inter_token_latency_ms_max": 4.3365856589131955,
    "mean_inter_token_latency_ms_mean": 4.121617063536678,
    "mean_inter_token_latency_ms_mean_ci95_high": 4.236380591343077,
    "mean_inter_token_latency_ms_mean_ci95_low": 4.006853535730278,
    "mean_inter_token_latency_ms_min": 3.9175448113177103,
    "mean_inter_token_latency_ms_p25": 4.057947041425673,
    "mean_inter_token_latency_ms_p50": 4.096568210862721,
    "mean_inter_token_latency_ms_p75": 4.220887719304556,
    "mean_inter_token_latency_ms_p90": 4.3365856589131955,
    "mean_inter_token_latency_ms_p95": 4.3365856589131955,
    "mean_inter_token_latency_ms_p99": 4.3365856589131955,
    "mean_inter_token_latency_ms_samples": 6,
    "mean_inter_token_latency_ms_stddev": 0.14342453275887065,
    "mean_time_per_output_token_max_ms": 4.3365856589131955,
    "mean_time_per_output_token_mean_ci95_high_ms": 4.236380591343077,
    "mean_time_per_output_token_mean_ci95_low_ms": 4.006853535730278,
    "mean_time_per_output_token_mean_ms": 4.121617063536678,
    "mean_time_per_output_token_min_ms": 3.9175448113177103,
    "mean_time_per_output_token_p25_ms": 4.057947041425673,
    "mean_time_per_output_token_p50_ms": 4.096568210862721,
    "mean_time_per_output_token_p75_ms": 4.220887719304556,
    "mean_time_per_output_token_p90_ms": 4.3365856589131955,
    "mean_time_per_output_token_p95_ms": 4.3365856589131955,
    "mean_time_per_output_token_p99_ms": 4.3365856589131955,
    "mean_time_per_output_token_samples": 6,
    "mean_time_per_output_token_stddev_ms": 0.14342453275887065,
    "output_token_sources": {
      "usage": 6
    },
    "output_tokens": 2304,
    "prefix_cache_hit_avg": 0.0,
    "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": 491,
    "request_canary_passed": 6,
    "request_canary_samples": 6,
    "throughput_performance_eligible": true,
    "throughput_tok_s": 40.09576546410587,
    "time_to_first_output_max_ms": 8241.704399999435,
    "time_to_first_output_mean_ci95_high_ms": 8134.191270054355,
    "time_to_first_output_mean_ci95_low_ms": 7841.033796611137,
    "time_to_first_output_mean_ms": 7987.612533332746,
    "time_to_first_output_min_ms": 7758.211599997594,
    "time_to_first_output_p25_ms": 7797.863700001471,
    "time_to_first_output_p50_ms": 7992.319099997985,
    "time_to_first_output_p75_ms": 8097.060100000817,
    "time_to_first_output_p90_ms": 8241.704399999435,
    "time_to_first_output_p95_ms": 8241.704399999435,
    "time_to_first_output_p99_ms": 8241.704399999435,
    "time_to_first_output_samples": 6,
    "time_to_first_output_stddev_ms": 183.1852612804575,
    "tpot_max_ms": 4.3365856589131955,
    "tpot_mean_ci95_high_ms": 4.236380591343077,
    "tpot_mean_ci95_low_ms": 4.006853535730278,
    "tpot_mean_ms": 4.121617063536678,
    "tpot_min_ms": 3.9175448113177103,
    "tpot_p25_ms": 4.057947041425673,
    "tpot_p50_ms": 4.096568210862721,
    "tpot_p75_ms": 4.220887719304556,
    "tpot_p90_ms": 4.3365856589131955,
    "tpot_p95_ms": 4.3365856589131955,
    "tpot_p99_ms": 4.3365856589131955,
    "tpot_samples": 6,
    "tpot_stddev_ms": 0.14342453275887065,
    "ttft_max_ms": 10374.570799998764,
    "ttft_mean_ci95_high_ms": 9930.04668515914,
    "ttft_mean_ci95_low_ms": 8967.798614841438,
    "ttft_mean_ms": 9448.92265000029,
    "ttft_min_ms": 8800.625100000616,
    "ttft_p25_ms": 8966.76590000061,
    "ttft_p50_ms": 9233.508999997866,
    "ttft_p75_ms": 9944.86660000257,
    "ttft_p90_ms": 10374.570799998764,
    "ttft_p95_ms": 10374.570799998764,
    "ttft_p99_ms": 10374.570799998764,
    "ttft_samples": 6,
    "ttft_stddev_ms": 601.2797903714594
  },
  "model": "qwen38-swift15-ninfer-dflash2-c4-262k",
  "performance_eligible": true,
  "prompt_cache_mode": "unique",
  "request_canaries": true,
  "request_timings": [
    {
      "content_chunks": 6,
      "controlled_output": null,
      "decode_tok_s": 243.8923895041408,
      "e2e_ms": 9324.149000000034,
      "effective_prefill_tok_s": 7562.392070262102,
      "generation_ms": 1082.4446000005992,
      "mean_inter_token_latency_ms": 4.100168939396209,
      "mean_time_per_output_token_ms": 4.100168939396209,
      "output_token_source": "usage",
      "output_tokens": 265,
      "planned_context_tokens": 65536,
      "prompt_tokens": 62327,
      "reasoning_chunks": 56,
      "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": 8241.704399999435,
      "tpot_ms": 4.100168939396209,
      "ttft_ms": 9233.508999997866,
      "visible_content_truncated": false,
      "visible_generation_ms": 90.64000000216765
    },
    {
      "content_chunks": 10,
      "controlled_output": null,
      "decode_tok_s": 244.10676169100185,
      "e2e_ms": 10556.770799998048,
      "effective_prefill_tok_s": 7798.487425260049,
      "generation_ms": 2564.4517000000633,
      "mean_inter_token_latency_ms": 4.096568210862721,
      "mean_time_per_output_token_ms": 4.096568210862721,
      "output_token_source": "usage",
      "output_tokens": 627,
      "planned_context_tokens": 65536,
      "prompt_tokens": 62328,
      "reasoning_chunks": 133,
      "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": 7992.319099997985,
      "tpot_ms": 4.096568210862721,
      "ttft_ms": 10374.570799998764,
      "visible_content_truncated": false,
      "visible_generation_ms": 182.19999999928405
    },
    {
      "content_chunks": 7,
      "controlled_output": null,
      "decode_tok_s": 236.9169867813405,
      "e2e_ms": 9482.059900001332,
      "effective_prefill_tok_s": 7753.669666628207,
      "generation_ms": 1443.5436000021582,
      "mean_inter_token_latency_ms": 4.220887719304556,
      "mean_time_per_output_token_ms": 4.220887719304556,
      "output_token_source": "usage",
      "output_tokens": 343,
      "planned_context_tokens": 65536,
      "prompt_tokens": 62328,
      "reasoning_chunks": 75,
      "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": 8038.5162999991735,
      "tpot_ms": 4.220887719304556,
      "ttft_ms": 9373.198500001308,
      "visible_content_truncated": false,
      "visible_generation_ms": 108.86140000002342
    },
    {
      "content_chunks": 8,
      "controlled_output": null,
      "decode_tok_s": 255.2619173904583,
      "e2e_ms": 8927.579600000172,
      "effective_prefill_tok_s": 7697.855669861424,
      "generation_ms": 830.5194999993546,
      "mean_inter_token_latency_ms": 3.9175448113177103,
      "mean_time_per_output_token_ms": 3.9175448113177103,
      "output_token_source": "usage",
      "output_tokens": 213,
      "planned_context_tokens": 65536,
      "prompt_tokens": 62330,
      "reasoning_chunks": 40,
      "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": 8097.060100000817,
      "tpot_ms": 3.9175448113177103,
      "ttft_ms": 8800.625100000616,
      "visible_content_truncated": false,
      "visible_generation_ms": 126.95449999955599
    },
    {
      "content_chunks": 6,
      "controlled_output": null,
      "decode_tok_s": 230.59615989474375,
      "e2e_ms": 10035.54190000068,
      "effective_prefill_tok_s": 7993.599580355352,
      "generation_ms": 2237.678199999209,
      "mean_inter_token_latency_ms": 4.3365856589131955,
      "mean_time_per_output_token_ms": 4.3365856589131955,
      "output_token_source": "usage",
      "output_tokens": 517,
      "planned_context_tokens": 65536,
      "prompt_tokens": 62333,
      "reasoning_chunks": 120,
      "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": 7797.863700001471,
      "tpot_ms": 4.3365856589131955,
      "ttft_ms": 9944.86660000257,
      "visible_content_truncated": false,
      "visible_generation_ms": 90.67529999811086
    },
    {
      "content_chunks": 10,
      "controlled_output": null,
      "decode_tok_s": 246.43002725059497,
      "e2e_ms": 9129.797699999472,
      "effective_prefill_tok_s": 8034.712536072016,
      "generation_ms": 1371.5861000018776,
      "mean_inter_token_latency_ms": 4.057947041425673,
      "mean_time_per_output_token_ms": 4.057947041425673,
      "output_token_source": "usage",
      "output_tokens": 339,
      "planned_context_tokens": 65536,
      "prompt_tokens": 62335,
      "reasoning_chunks": 67,
      "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": 7758.211599997594,
      "tpot_ms": 4.057947041425673,
      "ttft_ms": 8966.76590000061,
      "visible_content_truncated": false,
      "visible_generation_ms": 163.03179999886197
    }
  ],
  "requests": 6,
  "response_words": 0,
  "responses_received": 6,
  "run_id": "benchmark-20261005T113341Z",
  "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": "2026-10-05T11:33:41Z",
  "started_at_unix_ns": 1791200021603751300,
  "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": 57462.42710000297
}