{ "base_url": "[operator identity redacted]", "clock_domain_id": null, "completed": 3, "concurrency": 1, "configuration_fingerprint": null, "context_distribution": { "8192": 3 }, "context_seed": 17, "context_tokens": 8192, "controlled_output_policy": "observe", "engine": "strata", "failed": 0, "failures": [], "finished_at": "2026-10-05T10:49:51Z", "finished_at_unix_ns": 1791197391902324100, "gpu": "rtx5090", "manifest": null, "max_context_tokens": 32768, "max_tokens": 8192, "measurement_protocol": "capacity-v4-reasoning", "metric_population": { "aggregate_throughput_eligible": true, "attempted_requests": 3, "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": 3, "e2e": 3, "effective_prefill_tok_s": 3, "tpot": 3, "ttft": 3 }, "performance_eligible_requests": 3, "successful_requests": 3 }, "metrics": { "completion_tokens_max": 302.0, "completion_tokens_mean": 265.0, "completion_tokens_mean_ci95_high": 328.90321431665234, "completion_tokens_mean_ci95_low": 201.09678568334766, "completion_tokens_min": 200.0, "completion_tokens_p25": 200.0, "completion_tokens_p50": 293.0, "completion_tokens_p75": 302.0, "completion_tokens_p90": 302.0, "completion_tokens_p95": 302.0, "completion_tokens_p99": 302.0, "completion_tokens_samples": 3, "completion_tokens_stddev": 56.471231613982, "content_chunks": 107, "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": 89.76554402975655, "decode_tok_s_mean": 83.59881581998275, "decode_tok_s_mean_ci95_high": 93.45496413906754, "decode_tok_s_mean_ci95_low": 73.74266750089795, "decode_tok_s_min": 73.63501042063028, "decode_tok_s_p25": 73.63501042063028, "decode_tok_s_p50": 87.3958930095614, "decode_tok_s_p75": 89.76554402975655, "decode_tok_s_p90": 89.76554402975655, "decode_tok_s_p95": 89.76554402975655, "decode_tok_s_p99": 89.76554402975655, "decode_tok_s_samples": 3, "decode_tok_s_stddev": 8.709872273259924, "e2e_max_ms": 31237.22160000034, "e2e_mean_ci95_high_ms": 31221.133186866824, "e2e_mean_ci95_low_ms": 29434.02134646576, "e2e_mean_ms": 30327.577266666292, "e2e_min_ms": 29818.592899999203, "e2e_p25_ms": 29818.592899999203, "e2e_p50_ms": 29926.91729999933, "e2e_p75_ms": 31237.22160000034, "e2e_p90_ms": 31237.22160000034, "e2e_p95_ms": 31237.22160000034, "e2e_p99_ms": 31237.22160000034, "e2e_samples": 3, "e2e_stddev_ms": 789.6348230567756, "effective_prefill_tok_s_max": 280.29035269323174, "effective_prefill_tok_s_mean": 272.8452533396152, "effective_prefill_tok_s_mean_ci95_high": 281.0125190973987, "effective_prefill_tok_s_mean_ci95_low": 264.67798758183176, "effective_prefill_tok_s_min": 265.8794264941047, "effective_prefill_tok_s_p25": 265.8794264941047, "effective_prefill_tok_s_p50": 272.36598083150926, "effective_prefill_tok_s_p75": 280.29035269323174, "effective_prefill_tok_s_p90": 280.29035269323174, "effective_prefill_tok_s_p95": 280.29035269323174, "effective_prefill_tok_s_p99": 280.29035269323174, "effective_prefill_tok_s_samples": 3, "effective_prefill_tok_s_stddev": 7.217407781325739, "generation_max_ms": 3353.1796999996004, "generation_mean_ci95_high_ms": 3553.4858207591897, "generation_mean_ci95_low_ms": 2711.058579241697, "generation_mean_ms": 3132.2722000004433, "generation_min_ms": 2702.518800000689, "generation_p25_ms": 2702.518800000689, "generation_p50_ms": 3341.11810000104, "generation_p75_ms": 3353.1796999996004, "generation_p90_ms": 3353.1796999996004, "generation_p95_ms": 3353.1796999996004, "generation_p99_ms": 3353.1796999996004, "generation_samples": 3, "generation_stddev_ms": 372.226220405203, "mean_inter_token_latency_ms_max": 13.580496482415525, "mean_inter_token_latency_ms_mean": 12.05427121669269, "mean_inter_token_latency_ms_mean_ci95_high": 13.559704220108333, "mean_inter_token_latency_ms_mean_ci95_low": 10.548838213277048, "mean_inter_token_latency_ms_min": 11.14013189368638, "mean_inter_token_latency_ms_p25": 11.14013189368638, "mean_inter_token_latency_ms_p50": 11.442185273976165, "mean_inter_token_latency_ms_p75": 13.580496482415525, "mean_inter_token_latency_ms_p90": 13.580496482415525, "mean_inter_token_latency_ms_p95": 13.580496482415525, "mean_inter_token_latency_ms_p99": 13.580496482415525, "mean_inter_token_latency_ms_samples": 3, "mean_inter_token_latency_ms_stddev": 1.330350229238216, "mean_time_per_output_token_max_ms": 13.580496482415525, "mean_time_per_output_token_mean_ci95_high_ms": 13.559704220108333, "mean_time_per_output_token_mean_ci95_low_ms": 10.548838213277048, "mean_time_per_output_token_mean_ms": 12.05427121669269, "mean_time_per_output_token_min_ms": 11.14013189368638, "mean_time_per_output_token_p25_ms": 11.14013189368638, "mean_time_per_output_token_p50_ms": 11.442185273976165, "mean_time_per_output_token_p75_ms": 13.580496482415525, "mean_time_per_output_token_p90_ms": 13.580496482415525, "mean_time_per_output_token_p95_ms": 13.580496482415525, "mean_time_per_output_token_p99_ms": 13.580496482415525, "mean_time_per_output_token_samples": 3, "mean_time_per_output_token_stddev_ms": 1.330350229238216, "output_token_sources": { "usage": 3 }, "output_tokens": 795, "prefix_cache_hit_avg": 0.0, "prompt_token_samples": 3, "prompt_tokens": 22250, "prompt_tokens_max": 7418.0, "prompt_tokens_mean": 7416.666666666667, "prompt_tokens_mean_ci95_high": 7418.395224189896, "prompt_tokens_mean_ci95_low": 7414.938109143438, "prompt_tokens_min": 7415.0, "prompt_tokens_p25": 7415.0, "prompt_tokens_p50": 7417.0, "prompt_tokens_p75": 7418.0, "prompt_tokens_p90": 7418.0, "prompt_tokens_p95": 7418.0, "prompt_tokens_p99": 7418.0, "prompt_tokens_samples": 3, "prompt_tokens_stddev": 1.5275252316519468, "reasoning_chunks": 679, "request_canary_passed": 3, "request_canary_samples": 3, "throughput_performance_eligible": true, "throughput_tok_s": 8.737772855257319, "time_to_first_output_max_ms": 27896.1034999993, "time_to_first_output_mean_ci95_high_ms": 28005.29626476322, "time_to_first_output_mean_ci95_low_ms": 26385.31386856848, "time_to_first_output_mean_ms": 27195.30506666585, "time_to_first_output_min_ms": 26465.413199999603, "time_to_first_output_p25_ms": 26465.413199999603, "time_to_first_output_p50_ms": 27224.398499998642, "time_to_first_output_p75_ms": 27896.1034999993, "time_to_first_output_p90_ms": 27896.1034999993, "time_to_first_output_p95_ms": 27896.1034999993, "time_to_first_output_p99_ms": 27896.1034999993, "time_to_first_output_samples": 3, "time_to_first_output_stddev_ms": 715.7887289735867, "tpot_max_ms": 13.580496482415525, "tpot_mean_ci95_high_ms": 13.559704220108333, "tpot_mean_ci95_low_ms": 10.548838213277048, "tpot_mean_ms": 12.05427121669269, "tpot_min_ms": 11.14013189368638, "tpot_p25_ms": 11.14013189368638, "tpot_p50_ms": 11.442185273976165, "tpot_p75_ms": 13.580496482415525, "tpot_p90_ms": 13.580496482415525, "tpot_p95_ms": 13.580496482415525, "tpot_p99_ms": 13.580496482415525, "tpot_samples": 3, "tpot_stddev_ms": 1.330350229238216, "ttft_max_ms": 30779.642800000147, "ttft_mean_ci95_high_ms": 30767.353223101058, "ttft_mean_ci95_low_ms": 29071.077110233393, "ttft_mean_ms": 29919.215166667225, "ttft_min_ms": 29408.396800001356, "ttft_p25_ms": 29408.396800001356, "ttft_p50_ms": 29569.60590000017, "ttft_p75_ms": 30779.642800000147, "ttft_p90_ms": 30779.642800000147, "ttft_p95_ms": 30779.642800000147, "ttft_p99_ms": 30779.642800000147, "ttft_samples": 3, "ttft_stddev_ms": 749.4990844776116 }, "model": "qwen38-flash-next-strata-iq4xs-32k-c1-r36-m50", "performance_eligible": true, "prompt_cache_mode": "unique", "request_canaries": true, "request_timings": [ { "content_chunks": 34, "controlled_output": null, "decode_tok_s": 73.63501042063028, "e2e_ms": 29926.91729999933, "effective_prefill_tok_s": 272.36598083150926, "generation_ms": 2702.518800000689, "mean_inter_token_latency_ms": 13.580496482415525, "mean_time_per_output_token_ms": 13.580496482415525, "output_token_source": "usage", "output_tokens": 200, "planned_context_tokens": 8192, "prompt_tokens": 7415, "reasoning_chunks": 163, "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": 27224.398499998642, "tpot_ms": 13.580496482415525, "ttft_ms": 29569.60590000017, "visible_content_truncated": false, "visible_generation_ms": 357.31139999916195 }, { "content_chunks": 35, "controlled_output": null, "decode_tok_s": 87.3958930095614, "e2e_ms": 31237.22160000034, "effective_prefill_tok_s": 265.8794264941047, "generation_ms": 3341.11810000104, "mean_inter_token_latency_ms": 11.442185273976165, "mean_time_per_output_token_ms": 11.442185273976165, "output_token_source": "usage", "output_tokens": 293, "planned_context_tokens": 8192, "prompt_tokens": 7417, "reasoning_chunks": 255, "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": 27896.1034999993, "tpot_ms": 11.442185273976165, "ttft_ms": 30779.642800000147, "visible_content_truncated": false, "visible_generation_ms": 457.57880000019213 }, { "content_chunks": 38, "controlled_output": null, "decode_tok_s": 89.76554402975655, "e2e_ms": 29818.592899999203, "effective_prefill_tok_s": 280.29035269323174, "generation_ms": 3353.1796999996004, "mean_inter_token_latency_ms": 11.14013189368638, "mean_time_per_output_token_ms": 11.14013189368638, "output_token_source": "usage", "output_tokens": 302, "planned_context_tokens": 8192, "prompt_tokens": 7418, "reasoning_chunks": 261, "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": 26465.413199999603, "tpot_ms": 11.14013189368638, "ttft_ms": 29408.396800001356, "visible_content_truncated": false, "visible_generation_ms": 410.19609999784734 } ], "requests": 3, "response_words": 0, "responses_received": 3, "run_id": "benchmark-20261005T104820Z", "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": "2026-10-05T10:48:20Z", "started_at_unix_ns": 1791197300918038400, "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": 90984.28319999948 }