{
  "base_url": "http://127.0.0.1:30002/v1",
  "clock_domain_id": null,
  "completed": 4,
  "concurrency": 4,
  "configuration_fingerprint": null,
  "context_distribution": {
    "8192": 4
  },
  "context_seed": 0,
  "context_tokens": 8192,
  "controlled_output_policy": "strict",
  "engine": "vllm",
  "failed": 0,
  "failures": [],
  "finished_at": "2026-09-17T17:21:29Z",
  "finished_at_unix_ns": 1789665689912896690,
  "gpu": "dual-sm120",
  "manifest": null,
  "max_context_tokens": 327680,
  "max_tokens": 16384,
  "measurement_protocol": "capacity-v4-reasoning",
  "metric_population": {
    "aggregate_throughput_eligible": true,
    "attempted_requests": 4,
    "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": 4,
      "e2e": 4,
      "effective_prefill_tok_s": 4,
      "tpot": 4,
      "ttft": 4
    },
    "performance_eligible_requests": 4,
    "successful_requests": 4
  },
  "metrics": {
    "completion_tokens_max": 852.0,
    "completion_tokens_mean": 630.0,
    "completion_tokens_mean_ci95_high": 825.8872675801059,
    "completion_tokens_mean_ci95_low": 434.1127324198941,
    "completion_tokens_min": 461.0,
    "completion_tokens_p25": 461.0,
    "completion_tokens_p50": 461.0,
    "completion_tokens_p75": 746.0,
    "completion_tokens_p90": 852.0,
    "completion_tokens_p95": 852.0,
    "completion_tokens_p99": 852.0,
    "completion_tokens_samples": 4,
    "completion_tokens_stddev": 199.88496691847539,
    "content_chunks": 292,
    "content_chunks_s": null,
    "controlled_output_exact_adherence": 4,
    "controlled_output_exact_adherence_rate": 1.0,
    "controlled_output_samples": 4,
    "controlled_output_unobservable": 0,
    "decode_tok_s_max": 58.53838083656459,
    "decode_tok_s_mean": 39.96323947127735,
    "decode_tok_s_mean_ci95_high": 53.7773329844465,
    "decode_tok_s_mean_ci95_low": 26.149145958108193,
    "decode_tok_s_min": 24.825611425374145,
    "decode_tok_s_p25": 24.825611425374145,
    "decode_tok_s_p50": 35.41597563396896,
    "decode_tok_s_p75": 41.072989989201695,
    "decode_tok_s_p90": 58.53838083656459,
    "decode_tok_s_p95": 58.53838083656459,
    "decode_tok_s_p99": 58.53838083656459,
    "decode_tok_s_samples": 4,
    "decode_tok_s_stddev": 14.096013788948119,
    "e2e_max_ms": 25668.013541027904,
    "e2e_mean_ci95_high_ms": 25368.622083720125,
    "e2e_mean_ci95_low_ms": 19890.144839263216,
    "e2e_mean_ms": 22629.38346149167,
    "e2e_min_ms": 20185.567106935196,
    "e2e_p25_ms": 20185.567106935196,
    "e2e_p50_ms": 20324.98452393338,
    "e2e_p75_ms": 24338.9686740702,
    "e2e_p90_ms": 25668.013541027904,
    "e2e_p95_ms": 25668.013541027904,
    "e2e_p99_ms": 25668.013541027904,
    "e2e_samples": 4,
    "e2e_stddev_ms": 2795.1414512535275,
    "effective_prefill_tok_s_max": 3662.3456952908055,
    "effective_prefill_tok_s_mean": 1803.0703708573544,
    "effective_prefill_tok_s_mean_ci95_high": 3132.819718351965,
    "effective_prefill_tok_s_mean_ci95_low": 473.321023362744,
    "effective_prefill_tok_s_min": 486.8093742554089,
    "effective_prefill_tok_s_p25": 486.8093742554089,
    "effective_prefill_tok_s_p50": 1226.1555319985946,
    "effective_prefill_tok_s_p75": 1836.9708818846084,
    "effective_prefill_tok_s_p90": 3662.3456952908055,
    "effective_prefill_tok_s_p95": 3662.3456952908055,
    "effective_prefill_tok_s_p99": 3662.3456952908055,
    "effective_prefill_tok_s_samples": 4,
    "effective_prefill_tok_s_stddev": 1356.8870892802147,
    "generation_max_ms": 21035.704556037672,
    "generation_mean_ci95_high_ms": 23130.174573605163,
    "generation_mean_ci95_low_ms": 10940.955809921743,
    "generation_mean_ms": 17035.565191763453,
    "generation_min_ms": 7858.092305017635,
    "generation_p25_ms": 7858.092305017635,
    "generation_p50_ms": 18529.25159095321,
    "generation_p75_ms": 20719.212315045297,
    "generation_p90_ms": 21035.704556037672,
    "generation_p95_ms": 21035.704556037672,
    "generation_p99_ms": 21035.704556037672,
    "generation_samples": 4,
    "generation_stddev_ms": 6218.989165144601,
    "mean_inter_token_latency_ms_max": 40.2809817194635,
    "mean_inter_token_latency_ms_mean": 27.486634151893398,
    "mean_inter_token_latency_ms_mean_ci95_high": 36.99405091155727,
    "mean_inter_token_latency_ms_mean_ci95_low": 17.979217392229533,
    "mean_inter_token_latency_ms_min": 17.08280935873399,
    "mean_inter_token_latency_ms_p25": 17.08280935873399,
    "mean_inter_token_latency_ms_p50": 24.3469004877148,
    "mean_inter_token_latency_ms_p75": 28.235845041661303,
    "mean_inter_token_latency_ms_p90": 40.2809817194635,
    "mean_inter_token_latency_ms_p95": 40.2809817194635,
    "mean_inter_token_latency_ms_p99": 40.2809817194635,
    "mean_inter_token_latency_ms_samples": 4,
    "mean_inter_token_latency_ms_stddev": 9.701445673126393,
    "mean_time_per_output_token_max_ms": 40.2809817194635,
    "mean_time_per_output_token_mean_ci95_high_ms": 36.99405091155727,
    "mean_time_per_output_token_mean_ci95_low_ms": 17.979217392229533,
    "mean_time_per_output_token_mean_ms": 27.486634151893398,
    "mean_time_per_output_token_min_ms": 17.08280935873399,
    "mean_time_per_output_token_p25_ms": 17.08280935873399,
    "mean_time_per_output_token_p50_ms": 24.3469004877148,
    "mean_time_per_output_token_p75_ms": 28.235845041661303,
    "mean_time_per_output_token_p90_ms": 40.2809817194635,
    "mean_time_per_output_token_p95_ms": 40.2809817194635,
    "mean_time_per_output_token_p99_ms": 40.2809817194635,
    "mean_time_per_output_token_samples": 4,
    "mean_time_per_output_token_stddev_ms": 9.701445673126393,
    "output_token_sources": {
      "usage": 4
    },
    "output_tokens": 2520,
    "prefix_cache_hit_avg": null,
    "prompt_token_samples": 4,
    "prompt_tokens": 24271,
    "prompt_tokens_max": 6069.0,
    "prompt_tokens_mean": 6067.75,
    "prompt_tokens_mean_ci95_high": 6068.983139624427,
    "prompt_tokens_mean_ci95_low": 6066.516860375573,
    "prompt_tokens_min": 6066.0,
    "prompt_tokens_p25": 6066.0,
    "prompt_tokens_p50": 6068.0,
    "prompt_tokens_p75": 6068.0,
    "prompt_tokens_p90": 6069.0,
    "prompt_tokens_p95": 6069.0,
    "prompt_tokens_p99": 6069.0,
    "prompt_tokens_samples": 4,
    "prompt_tokens_stddev": 1.2583057392117916,
    "reasoning_chunks": 2022,
    "request_canary_passed": 4,
    "request_canary_samples": 4,
    "throughput_performance_eligible": true,
    "throughput_tok_s": 98.17192502462154,
    "time_to_first_output_max_ms": 12466.892218915746,
    "time_to_first_output_mean_ci95_high_ms": 10273.451115319878,
    "time_to_first_output_mean_ci95_low_ms": 914.1854241365563,
    "time_to_first_output_mean_ms": 5593.818269728217,
    "time_to_first_output_min_ms": 1656.3155159819871,
    "time_to_first_output_p25_ms": 1656.3155159819871,
    "time_to_first_output_p50_ms": 3303.26411803253,
    "time_to_first_output_p75_ms": 4948.801225982606,
    "time_to_first_output_p90_ms": 12466.892218915746,
    "time_to_first_output_p95_ms": 12466.892218915746,
    "time_to_first_output_p99_ms": 12466.892218915746,
    "time_to_first_output_samples": 4,
    "time_to_first_output_stddev_ms": 4775.135556726184,
    "tpot_max_ms": 40.2809817194635,
    "tpot_mean_ci95_high_ms": 36.99405091155727,
    "tpot_mean_ci95_low_ms": 17.979217392229533,
    "tpot_mean_ms": 27.486634151893398,
    "tpot_min_ms": 17.08280935873399,
    "tpot_p25_ms": 17.08280935873399,
    "tpot_p50_ms": 24.3469004877148,
    "tpot_p75_ms": 28.235845041661303,
    "tpot_p90_ms": 40.2809817194635,
    "tpot_p95_ms": 40.2809817194635,
    "tpot_p99_ms": 40.2809817194635,
    "tpot_samples": 4,
    "tpot_stddev_ms": 9.701445673126393,
    "ttft_max_ms": 24778.88415195048,
    "ttft_mean_ci95_high_ms": 24424.010559658505,
    "ttft_mean_ci95_low_ms": 18682.853291302003,
    "ttft_mean_ms": 21553.431925480254,
    "ttft_min_ms": 18995.270304963924,
    "ttft_p25_ms": 18995.270304963924,
    "ttft_p50_ms": 19149.811200913973,
    "ttft_p75_ms": 23289.76204409264,
    "ttft_p90_ms": 24778.88415195048,
    "ttft_p95_ms": 24778.88415195048,
    "ttft_p99_ms": 24778.88415195048,
    "ttft_samples": 4,
    "ttft_stddev_ms": 2929.1618716104613
  },
  "model": "glm53-mixed35-v84-stream-r7-c4-v2",
  "performance_eligible": true,
  "prompt_cache_mode": "unique",
  "request_canaries": true,
  "request_timings": [
    {
      "content_chunks": 73,
      "controlled_output": {
        "capture_complete": true,
        "exact_adherence": true,
        "observed_code_words": 64,
        "observed_extra_words": 0,
        "passed": true,
        "policy": "strict",
        "requested_words": 64
      },
      "decode_tok_s": 24.825611425374145,
      "e2e_ms": 20185.567106935196,
      "effective_prefill_tok_s": 3662.3456952908055,
      "generation_ms": 18529.25159095321,
      "mean_inter_token_latency_ms": 40.2809817194635,
      "mean_time_per_output_token_ms": 40.2809817194635,
      "output_token_source": "usage",
      "output_tokens": 461,
      "planned_context_tokens": 8192,
      "prompt_tokens": 6066,
      "reasoning_chunks": 318,
      "request_canary": {
        "capture_complete": true,
        "foreign_marker_count": 0,
        "marker": "ANVIL_REQ_0_00000",
        "marker_at_start": true,
        "observed_marker_count": 1,
        "passed": true
      },
      "request_index": 0,
      "time_to_first_output_ms": 1656.3155159819871,
      "tpot_ms": 40.2809817194635,
      "ttft_ms": 18995.270304963924,
      "visible_content_truncated": false,
      "visible_generation_ms": 1190.2968019712716
    },
    {
      "content_chunks": 73,
      "controlled_output": {
        "capture_complete": true,
        "exact_adherence": true,
        "observed_code_words": 64,
        "observed_extra_words": 0,
        "passed": true,
        "policy": "strict",
        "requested_words": 64
      },
      "decode_tok_s": 35.41597563396896,
      "e2e_ms": 24338.9686740702,
      "effective_prefill_tok_s": 1836.9708818846084,
      "generation_ms": 21035.704556037672,
      "mean_inter_token_latency_ms": 28.235845041661303,
      "mean_time_per_output_token_ms": 28.235845041661303,
      "output_token_source": "usage",
      "output_tokens": 746,
      "planned_context_tokens": 8192,
      "prompt_tokens": 6068,
      "reasoning_chunks": 632,
      "request_canary": {
        "capture_complete": true,
        "foreign_marker_count": 0,
        "marker": "ANVIL_REQ_0_00001",
        "marker_at_start": true,
        "observed_marker_count": 1,
        "passed": true
      },
      "request_index": 1,
      "time_to_first_output_ms": 3303.26411803253,
      "tpot_ms": 28.235845041661303,
      "ttft_ms": 23289.76204409264,
      "visible_content_truncated": false,
      "visible_generation_ms": 1049.2066299775615
    },
    {
      "content_chunks": 73,
      "controlled_output": {
        "capture_complete": true,
        "exact_adherence": true,
        "observed_code_words": 64,
        "observed_extra_words": 0,
        "passed": true,
        "policy": "strict",
        "requested_words": 64
      },
      "decode_tok_s": 41.072989989201695,
      "e2e_ms": 25668.013541027904,
      "effective_prefill_tok_s": 1226.1555319985946,
      "generation_ms": 20719.212315045297,
      "mean_inter_token_latency_ms": 24.3469004877148,
      "mean_time_per_output_token_ms": 24.3469004877148,
      "output_token_source": "usage",
      "output_tokens": 852,
      "planned_context_tokens": 8192,
      "prompt_tokens": 6068,
      "reasoning_chunks": 754,
      "request_canary": {
        "capture_complete": true,
        "foreign_marker_count": 0,
        "marker": "ANVIL_REQ_0_00002",
        "marker_at_start": true,
        "observed_marker_count": 1,
        "passed": true
      },
      "request_index": 2,
      "time_to_first_output_ms": 4948.801225982606,
      "tpot_ms": 24.3469004877148,
      "ttft_ms": 24778.88415195048,
      "visible_content_truncated": false,
      "visible_generation_ms": 889.129389077425
    },
    {
      "content_chunks": 73,
      "controlled_output": {
        "capture_complete": true,
        "exact_adherence": true,
        "observed_code_words": 64,
        "observed_extra_words": 0,
        "passed": true,
        "policy": "strict",
        "requested_words": 64
      },
      "decode_tok_s": 58.53838083656459,
      "e2e_ms": 20324.98452393338,
      "effective_prefill_tok_s": 486.8093742554089,
      "generation_ms": 7858.092305017635,
      "mean_inter_token_latency_ms": 17.08280935873399,
      "mean_time_per_output_token_ms": 17.08280935873399,
      "output_token_source": "usage",
      "output_tokens": 461,
      "planned_context_tokens": 8192,
      "prompt_tokens": 6069,
      "reasoning_chunks": 318,
      "request_canary": {
        "capture_complete": true,
        "foreign_marker_count": 0,
        "marker": "ANVIL_REQ_0_00003",
        "marker_at_start": true,
        "observed_marker_count": 1,
        "passed": true
      },
      "request_index": 3,
      "time_to_first_output_ms": 12466.892218915746,
      "tpot_ms": 17.08280935873399,
      "ttft_ms": 19149.811200913973,
      "visible_content_truncated": false,
      "visible_generation_ms": 1175.1733230194077
    }
  ],
  "requests": 4,
  "response_words": 64,
  "responses_received": 4,
  "run_id": "benchmark-20260917T172104Z",
  "sampling": {
    "temperature": {
      "effective_request": 0.0,
      "requested": null,
      "sent": true
    },
    "top_p": {
      "effective_request": null,
      "requested": null,
      "sent": false
    }
  },
  "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": "2026-09-17T17:21:04Z",
  "started_at_unix_ns": 1789665664243641356,
  "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": 25669.25319400616
}