{ "base_url": "[operator identity redacted]", "clock_domain_id": null, "completed": 12, "concurrency": 4, "configuration_fingerprint": null, "context_distribution": { "24576": 12 }, "context_seed": 17, "context_tokens": 24576, "controlled_output_policy": "observe", "engine": "ninfer", "failed": 0, "failures": [], "finished_at": "2026-10-05T09:25:55Z", "finished_at_unix_ns": 1791192355674000200, "gpu": "rtx5090", "manifest": null, "max_context_tokens": 262144, "max_tokens": 8192, "measurement_protocol": "capacity-v4-reasoning", "metric_population": { "aggregate_throughput_eligible": true, "attempted_requests": 12, "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": 12, "e2e": 12, "effective_prefill_tok_s": 12, "tpot": 12, "ttft": 12 }, "performance_eligible_requests": 12, "successful_requests": 12 }, "metrics": { "completion_tokens_max": 470.0, "completion_tokens_mean": 311.9166666666667, "completion_tokens_mean_ci95_high": 360.41246603697977, "completion_tokens_mean_ci95_low": 263.4208672963536, "completion_tokens_min": 217.0, "completion_tokens_p25": 236.0, "completion_tokens_p50": 272.0, "completion_tokens_p75": 343.0, "completion_tokens_p90": 426.0, "completion_tokens_p95": 470.0, "completion_tokens_p99": 470.0, "completion_tokens_samples": 12, "completion_tokens_stddev": 85.71141679902958, "content_chunks": 101, "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": 165.51054486067062, "decode_tok_s_mean": 60.283177405330186, "decode_tok_s_mean_ci95_high": 80.80615823812788, "decode_tok_s_mean_ci95_low": 39.76019657253249, "decode_tok_s_min": 32.200979150074744, "decode_tok_s_p25": 39.18720596376553, "decode_tok_s_p50": 45.96016735705395, "decode_tok_s_p75": 69.82067979450966, "decode_tok_s_p90": 77.16141875533732, "decode_tok_s_p95": 165.51054486067062, "decode_tok_s_p99": 165.51054486067062, "decode_tok_s_samples": 12, "decode_tok_s_stddev": 36.27229135221207, "e2e_max_ms": 17084.096099999442, "e2e_mean_ci95_high_ms": 11906.437321493066, "e2e_mean_ci95_low_ms": 8080.926645173569, "e2e_mean_ms": 9993.681983333317, "e2e_min_ms": 3398.2049000005645, "e2e_p25_ms": 7715.1405999993585, "e2e_p50_ms": 9814.59679999898, "e2e_p75_ms": 10298.539700001129, "e2e_p90_ms": 13622.960000000603, "e2e_p95_ms": 17084.096099999442, "e2e_p99_ms": 17084.096099999442, "e2e_samples": 12, "e2e_stddev_ms": 3380.601457287014, "effective_prefill_tok_s_max": 10746.949540837277, "effective_prefill_tok_s_mean": 8070.9740793952815, "effective_prefill_tok_s_mean_ci95_high": 9876.723734014198, "effective_prefill_tok_s_mean_ci95_low": 6265.224424776365, "effective_prefill_tok_s_min": 2520.588950548575, "effective_prefill_tok_s_p25": 4839.813107860371, "effective_prefill_tok_s_p50": 8919.06773632783, "effective_prefill_tok_s_p75": 10646.396833586527, "effective_prefill_tok_s_p90": 10691.451018016009, "effective_prefill_tok_s_p95": 10746.949540837277, "effective_prefill_tok_s_p99": 10746.949540837277, "effective_prefill_tok_s_samples": 12, "effective_prefill_tok_s_stddev": 3191.4797423978744, "generation_max_ms": 14564.774499998748, "generation_mean_ci95_high_ms": 8351.730099603981, "generation_mean_ci95_low_ms": 4546.079833729539, "generation_mean_ms": 6448.90496666676, "generation_min_ms": 1305.0528000003396, "generation_p25_ms": 3512.1178999997937, "generation_p50_ms": 6087.021800000002, "generation_p75_ms": 7705.567200000587, "generation_p90_ms": 8178.363800001534, "generation_p95_ms": 14564.774499998748, "generation_p99_ms": 14564.774499998748, "generation_samples": 12, "generation_stddev_ms": 3363.0508246594586, "mean_inter_token_latency_ms_max": 31.05495628997601, "mean_inter_token_latency_ms_mean": 20.042835600475524, "mean_inter_token_latency_ms_mean_ci95_high": 24.10450971761452, "mean_inter_token_latency_ms_mean_ci95_low": 15.981161483336528, "mean_inter_token_latency_ms_min": 6.041911111112683, "mean_inter_token_latency_ms_p25": 13.892010638293405, "mean_inter_token_latency_ms_p50": 20.291307327585763, "mean_inter_token_latency_ms_p75": 24.69733076923265, "mean_inter_token_latency_ms_p90": 28.36106616541656, "mean_inter_token_latency_ms_p95": 31.05495628997601, "mean_inter_token_latency_ms_p99": 31.05495628997601, "mean_inter_token_latency_ms_samples": 12, "mean_inter_token_latency_ms_stddev": 7.178597892522656, "mean_time_per_output_token_max_ms": 31.05495628997601, "mean_time_per_output_token_mean_ci95_high_ms": 24.10450971761452, "mean_time_per_output_token_mean_ci95_low_ms": 15.981161483336528, "mean_time_per_output_token_mean_ms": 20.042835600475524, "mean_time_per_output_token_min_ms": 6.041911111112683, "mean_time_per_output_token_p25_ms": 13.892010638293405, "mean_time_per_output_token_p50_ms": 20.291307327585763, "mean_time_per_output_token_p75_ms": 24.69733076923265, "mean_time_per_output_token_p90_ms": 28.36106616541656, "mean_time_per_output_token_p95_ms": 31.05495628997601, "mean_time_per_output_token_p99_ms": 31.05495628997601, "mean_time_per_output_token_samples": 12, "mean_time_per_output_token_stddev_ms": 7.178597892522656, "output_token_sources": { "usage": 12 }, "output_tokens": 3743, "prefix_cache_hit_avg": 0.0, "prompt_token_samples": 12, "prompt_tokens": 269768, "prompt_tokens_max": 22495.0, "prompt_tokens_mean": 22480.666666666668, "prompt_tokens_mean_ci95_high": 22485.915558805667, "prompt_tokens_mean_ci95_low": 22475.41777452767, "prompt_tokens_min": 22467.0, "prompt_tokens_p25": 22472.0, "prompt_tokens_p50": 22479.0, "prompt_tokens_p75": 22487.0, "prompt_tokens_p90": 22493.0, "prompt_tokens_p95": 22495.0, "prompt_tokens_p99": 22495.0, "prompt_tokens_samples": 12, "prompt_tokens_stddev": 9.276885579794873, "reasoning_chunks": 984, "request_canary_passed": 12, "request_canary_samples": 12, "throughput_performance_eligible": true, "throughput_tok_s": 123.53141958950683, "time_to_first_output_max_ms": 8915.376700000706, "time_to_first_output_mean_ci95_high_ms": 4808.2225771928925, "time_to_first_output_mean_ci95_low_ms": 2281.331456140223, "time_to_first_output_mean_ms": 3544.7770166665578, "time_to_first_output_min_ms": 2093.152100000225, "time_to_first_output_p25_ms": 2109.0295999983937, "time_to_first_output_p50_ms": 2124.799199998961, "time_to_first_output_p75_ms": 4203.022699999565, "time_to_first_output_p90_ms": 6776.663599999665, "time_to_first_output_p95_ms": 8915.376700000706, "time_to_first_output_p99_ms": 8915.376700000706, "time_to_first_output_samples": 12, "time_to_first_output_stddev_ms": 2233.0121463560718, "tpot_max_ms": 31.05495628997601, "tpot_mean_ci95_high_ms": 24.10450971761452, "tpot_mean_ci95_low_ms": 15.981161483336528, "tpot_mean_ms": 20.042835600475524, "tpot_min_ms": 6.041911111112683, "tpot_p25_ms": 13.892010638293405, "tpot_p50_ms": 20.291307327585763, "tpot_p75_ms": 24.69733076923265, "tpot_p90_ms": 28.36106616541656, "tpot_p95_ms": 31.05495628997601, "tpot_p99_ms": 31.05495628997601, "tpot_samples": 12, "tpot_stddev_ms": 7.178597892522656, "ttft_max_ms": 16897.40239999992, "ttft_mean_ci95_high_ms": 11444.49355242162, "ttft_mean_ci95_low_ms": 7808.6233975779, "ttft_mean_ms": 9626.55847499976, "ttft_min_ms": 3238.9356000003318, "ttft_p25_ms": 7572.877699998571, "ttft_p50_ms": 9502.766399999018, "ttft_p75_ms": 10131.436900001063, "ttft_p90_ms": 12188.387399999556, "ttft_p95_ms": 16897.40239999992, "ttft_p99_ms": 16897.40239999992, "ttft_samples": 12, "ttft_stddev_ms": 3213.0162438329826 }, "model": "qwen38-swift15-ninfer-dflash2-c4-262k", "performance_eligible": true, "prompt_cache_mode": "unique", "request_canaries": true, "request_timings": [ { "content_chunks": 7, "controlled_output": null, "decode_tok_s": 45.064025306898166, "e2e_ms": 12231.323600000906, "effective_prefill_tok_s": 4839.813107860371, "generation_ms": 7589.202200000727, "mean_inter_token_latency_ms": 22.190649707604464, "mean_time_per_output_token_ms": 22.190649707604464, "output_token_source": "usage", "output_tokens": 343, "planned_context_tokens": 24576, "prompt_tokens": 22467, "reasoning_chunks": 77, "request_canary": { "capture_complete": true, "foreign_marker_count": 0, "marker": "ANVIL_REQ_17_00000", "marker_at_start": true, "observed_marker_count": 1, "passed": true }, "request_index": 0, "time_to_first_output_ms": 4642.121400000178, "tpot_ms": 22.190649707604464, "ttft_ms": 11055.967899999814, "visible_content_truncated": false, "visible_generation_ms": 1175.3557000010915 }, { "content_chunks": 10, "controlled_output": null, "decode_tok_s": 32.200979150074744, "e2e_ms": 17084.096099999442, "effective_prefill_tok_s": 8919.06773632783, "generation_ms": 14564.774499998748, "mean_inter_token_latency_ms": 31.05495628997601, "mean_time_per_output_token_ms": 31.05495628997601, "output_token_source": "usage", "output_tokens": 470, "planned_context_tokens": 24576, "prompt_tokens": 22470, "reasoning_chunks": 131, "request_canary": { "capture_complete": true, "foreign_marker_count": 0, "marker": "ANVIL_REQ_17_00001", "marker_at_start": true, "observed_marker_count": 1, "passed": true }, "request_index": 1, "time_to_first_output_ms": 2519.321600000694, "tpot_ms": 31.05495628997601, "ttft_ms": 16897.40239999992, "visible_content_truncated": false, "visible_generation_ms": 186.69369999952323 }, { "content_chunks": 8, "controlled_output": null, "decode_tok_s": 49.28218689194625, "e2e_ms": 13622.960000000603, "effective_prefill_tok_s": 2520.588950548575, "generation_ms": 4707.583299999897, "mean_inter_token_latency_ms": 20.291307327585763, "mean_time_per_output_token_ms": 20.291307327585763, "output_token_source": "usage", "output_tokens": 233, "planned_context_tokens": 24576, "prompt_tokens": 22472, "reasoning_chunks": 65, "request_canary": { "capture_complete": true, "foreign_marker_count": 0, "marker": "ANVIL_REQ_17_00002", "marker_at_start": true, "observed_marker_count": 1, "passed": true }, "request_index": 2, "time_to_first_output_ms": 8915.376700000706, "tpot_ms": 20.291307327585763, "ttft_ms": 12188.387399999556, "visible_content_truncated": false, "visible_generation_ms": 1434.5726000010472 }, { "content_chunks": 6, "controlled_output": null, "decode_tok_s": 71.98382048769056, "e2e_ms": 10041.286099998615, "effective_prefill_tok_s": 3316.2336699140724, "generation_ms": 3264.62249999895, "mean_inter_token_latency_ms": 13.892010638293405, "mean_time_per_output_token_ms": 13.892010638293405, "output_token_source": "usage", "output_tokens": 236, "planned_context_tokens": 24576, "prompt_tokens": 22473, "reasoning_chunks": 58, "request_canary": { "capture_complete": true, "foreign_marker_count": 0, "marker": "ANVIL_REQ_17_00003", "marker_at_start": true, "observed_marker_count": 1, "passed": true }, "request_index": 3, "time_to_first_output_ms": 6776.663599999665, "tpot_ms": 13.892010638293405, "ttft_ms": 9936.503099999754, "visible_content_truncated": false, "visible_generation_ms": 104.78299999886076 }, { "content_chunks": 9, "controlled_output": null, "decode_tok_s": 35.25960533949878, "e2e_ms": 9668.842799999766, "effective_prefill_tok_s": 10577.940729651533, "generation_ms": 7544.043600000805, "mean_inter_token_latency_ms": 28.36106616541656, "mean_time_per_output_token_ms": 28.36106616541656, "output_token_source": "usage", "output_tokens": 267, "planned_context_tokens": 24576, "prompt_tokens": 22476, "reasoning_chunks": 78, "request_canary": { "capture_complete": true, "foreign_marker_count": 0, "marker": "ANVIL_REQ_17_00004", "marker_at_start": true, "observed_marker_count": 1, "passed": true }, "request_index": 4, "time_to_first_output_ms": 2124.799199998961, "tpot_ms": 28.36106616541656, "ttft_ms": 9502.766399999018, "visible_content_truncated": false, "visible_generation_ms": 166.07640000074753 }, { "content_chunks": 9, "controlled_output": null, "decode_tok_s": 51.477289381516755, "e2e_ms": 10298.539700001129, "effective_prefill_tok_s": 10602.42218582161, "generation_ms": 8178.363800001534, "mean_inter_token_latency_ms": 19.42604228028868, "mean_time_per_output_token_ms": 19.42604228028868, "output_token_source": "usage", "output_tokens": 422, "planned_context_tokens": 24576, "prompt_tokens": 22479, "reasoning_chunks": 109, "request_canary": { "capture_complete": true, "foreign_marker_count": 0, "marker": "ANVIL_REQ_17_00005", "marker_at_start": true, "observed_marker_count": 1, "passed": true }, "request_index": 5, "time_to_first_output_ms": 2120.1758999995945, "tpot_ms": 19.42604228028868, "ttft_ms": 10131.436900001063, "visible_content_truncated": false, "visible_generation_ms": 167.10280000006605 }, { "content_chunks": 6, "controlled_output": null, "decode_tok_s": 69.82067979450966, "e2e_ms": 8905.056899999181, "effective_prefill_tok_s": 7977.899210697039, "generation_ms": 6087.021800000002, "mean_inter_token_latency_ms": 14.322404235294123, "mean_time_per_output_token_ms": 14.322404235294123, "output_token_source": "usage", "output_tokens": 426, "planned_context_tokens": 24576, "prompt_tokens": 22482, "reasoning_chunks": 103, "request_canary": { "capture_complete": true, "foreign_marker_count": 0, "marker": "ANVIL_REQ_17_00006", "marker_at_start": true, "observed_marker_count": 1, "passed": true }, "request_index": 6, "time_to_first_output_ms": 2818.035099999179, "tpot_ms": 14.322404235294123, "ttft_ms": 8800.865499999418, "visible_content_truncated": false, "visible_generation_ms": 104.19139999976323 }, { "content_chunks": 11, "controlled_output": null, "decode_tok_s": 40.490205574999884, "e2e_ms": 9814.59679999898, "effective_prefill_tok_s": 10661.3012923181, "generation_ms": 7705.567200000587, "mean_inter_token_latency_ms": 24.69733076923265, "mean_time_per_output_token_ms": 24.69733076923265, "output_token_source": "usage", "output_tokens": 313, "planned_context_tokens": 24576, "prompt_tokens": 22485, "reasoning_chunks": 83, "request_canary": { "capture_complete": true, "foreign_marker_count": 0, "marker": "ANVIL_REQ_17_00007", "marker_at_start": true, "observed_marker_count": 1, "passed": true }, "request_index": 7, "time_to_first_output_ms": 2109.0295999983937, "tpot_ms": 24.69733076923265, "ttft_ms": 9328.157099998862, "visible_content_truncated": false, "visible_generation_ms": 486.43970000011905 }, { "content_chunks": 8, "controlled_output": null, "decode_tok_s": 39.18720596376553, "e2e_ms": 9818.7668999999, "effective_prefill_tok_s": 10646.396833586527, "generation_ms": 7706.5968999995675, "mean_inter_token_latency_ms": 25.518532781455523, "mean_time_per_output_token_ms": 25.518532781455523, "output_token_source": "usage", "output_tokens": 303, "planned_context_tokens": 24576, "prompt_tokens": 22487, "reasoning_chunks": 86, "request_canary": { "capture_complete": true, "foreign_marker_count": 0, "marker": "ANVIL_REQ_17_00008", "marker_at_start": true, "observed_marker_count": 1, "passed": true }, "request_index": 8, "time_to_first_output_ms": 2112.170000000333, "tpot_ms": 25.518532781455523, "ttft_ms": 9673.633899999913, "visible_content_truncated": false, "visible_generation_ms": 145.13299999998708 }, { "content_chunks": 8, "controlled_output": null, "decode_tok_s": 45.96016735705395, "e2e_ms": 7325.369400001364, "effective_prefill_tok_s": 10691.451018016009, "generation_ms": 5221.913100000165, "mean_inter_token_latency_ms": 21.75797125000069, "mean_time_per_output_token_ms": 21.75797125000069, "output_token_source": "usage", "output_tokens": 241, "planned_context_tokens": 24576, "prompt_tokens": 22489, "reasoning_chunks": 60, "request_canary": { "capture_complete": true, "foreign_marker_count": 0, "marker": "ANVIL_REQ_17_00009", "marker_at_start": true, "observed_marker_count": 1, "passed": true }, "request_index": 9, "time_to_first_output_ms": 2103.4563000011985, "tpot_ms": 21.75797125000069, "ttft_ms": 7191.767800000889, "visible_content_truncated": false, "visible_generation_ms": 133.60160000047472 }, { "content_chunks": 9, "controlled_output": null, "decode_tok_s": 77.16141875533732, "e2e_ms": 7715.1405999993585, "effective_prefill_tok_s": 5351.624677164444, "generation_ms": 3512.1178999997937, "mean_inter_token_latency_ms": 12.959844649445733, "mean_time_per_output_token_ms": 12.959844649445733, "output_token_source": "usage", "output_tokens": 272, "planned_context_tokens": 24576, "prompt_tokens": 22493, "reasoning_chunks": 74, "request_canary": { "capture_complete": true, "foreign_marker_count": 0, "marker": "ANVIL_REQ_17_00010", "marker_at_start": true, "observed_marker_count": 1, "passed": true }, "request_index": 10, "time_to_first_output_ms": 4203.022699999565, "tpot_ms": 12.959844649445733, "ttft_ms": 7572.877699998571, "visible_content_truncated": false, "visible_generation_ms": 142.26290000078734 }, { "content_chunks": 10, "controlled_output": null, "decode_tok_s": 165.51054486067062, "e2e_ms": 3398.2049000005645, "effective_prefill_tok_s": 10746.949540837277, "generation_ms": 1305.0528000003396, "mean_inter_token_latency_ms": 6.041911111112683, "mean_time_per_output_token_ms": 6.041911111112683, "output_token_source": "usage", "output_tokens": 217, "planned_context_tokens": 24576, "prompt_tokens": 22495, "reasoning_chunks": 60, "request_canary": { "capture_complete": true, "foreign_marker_count": 0, "marker": "ANVIL_REQ_17_00011", "marker_at_start": true, "observed_marker_count": 1, "passed": true }, "request_index": 11, "time_to_first_output_ms": 2093.152100000225, "tpot_ms": 6.041911111112683, "ttft_ms": 3238.9356000003318, "visible_content_truncated": false, "visible_generation_ms": 159.26930000023276 } ], "requests": 12, "response_words": 0, "responses_received": 12, "run_id": "benchmark-20261005T092525Z", "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": "2026-10-05T09:25:25Z", "started_at_unix_ns": 1791192325374013800, "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": 30299.983700000666 }