{ "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": "2026-10-05T11:08:08Z", "finished_at_unix_ns": 1791198488423998800, "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": 330.0, "completion_tokens_mean": 280.0, "completion_tokens_mean_ci95_high": 353.15291381756435, "completion_tokens_mean_ci95_low": 206.84708618243565, "completion_tokens_min": 207.0, "completion_tokens_p25": 207.0, "completion_tokens_p50": 303.0, "completion_tokens_p75": 330.0, "completion_tokens_p90": 330.0, "completion_tokens_p95": 330.0, "completion_tokens_p99": 330.0, "completion_tokens_samples": 3, "completion_tokens_stddev": 64.64518543557594, "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": 77.96346420377165, "decode_tok_s_mean": 73.47703484991278, "decode_tok_s_mean_ci95_high": 78.43192545575485, "decode_tok_s_mean_ci95_low": 68.52214424407072, "decode_tok_s_min": 69.21481105563299, "decode_tok_s_p25": 69.21481105563299, "decode_tok_s_p50": 73.25282929033372, "decode_tok_s_p75": 77.96346420377165, "decode_tok_s_p90": 77.96346420377165, "decode_tok_s_p95": 77.96346420377165, "decode_tok_s_p99": 77.96346420377165, "decode_tok_s_samples": 3, "decode_tok_s_stddev": 4.378633813910309, "e2e_max_ms": 30648.735099999612, "e2e_mean_ci95_high_ms": 30979.772727160867, "e2e_mean_ci95_low_ms": 26790.754139507793, "e2e_mean_ms": 28885.26343333433, "e2e_min_ms": 26957.813500001066, "e2e_p25_ms": 26957.813500001066, "e2e_p50_ms": 29049.241700002312, "e2e_p75_ms": 30648.735099999612, "e2e_p90_ms": 30648.735099999612, "e2e_p95_ms": 30648.735099999612, "e2e_p99_ms": 30648.735099999612, "e2e_samples": 3, "e2e_stddev_ms": 1850.916588690191, "effective_prefill_tok_s_max": 309.19574542151054, "effective_prefill_tok_s_mean": 295.79931473322455, "effective_prefill_tok_s_mean_ci95_high": 310.352167399351, "effective_prefill_tok_s_mean_ci95_low": 281.2464620670981, "effective_prefill_tok_s_min": 283.55220222039975, "effective_prefill_tok_s_p25": 283.55220222039975, "effective_prefill_tok_s_p50": 294.6499965577635, "effective_prefill_tok_s_p75": 309.19574542151054, "effective_prefill_tok_s_p90": 309.19574542151054, "effective_prefill_tok_s_p95": 309.19574542151054, "effective_prefill_tok_s_p99": 309.19574542151054, "effective_prefill_tok_s_samples": 3, "effective_prefill_tok_s_stddev": 12.860347047344515, "generation_max_ms": 4491.294100000232, "generation_mean_ci95_high_ms": 4642.45832717983, "generation_mean_ci95_low_ms": 2918.305006154932, "generation_mean_ms": 3780.381666667381, "generation_min_ms": 2976.241600001231, "generation_p25_ms": 2976.241600001231, "generation_p50_ms": 3873.6093000006804, "generation_p75_ms": 4491.294100000232, "generation_p90_ms": 4491.294100000232, "generation_p95_ms": 4491.294100000232, "generation_p99_ms": 4491.294100000232, "generation_samples": 3, "generation_stddev_ms": 761.8166204218716, "mean_inter_token_latency_ms_max": 14.447774757287528, "mean_inter_token_latency_ms_mean": 13.64188182208065, "mean_inter_token_latency_ms_mean_ci95_high": 14.559239494088132, "mean_inter_token_latency_ms_mean_ci95_low": 12.724524150073169, "mean_inter_token_latency_ms_min": 12.826520860929405, "mean_inter_token_latency_ms_p25": 12.826520860929405, "mean_inter_token_latency_ms_p50": 13.65134984802502, "mean_inter_token_latency_ms_p75": 14.447774757287528, "mean_inter_token_latency_ms_p90": 14.447774757287528, "mean_inter_token_latency_ms_p95": 14.447774757287528, "mean_inter_token_latency_ms_p99": 14.447774757287528, "mean_inter_token_latency_ms_samples": 3, "mean_inter_token_latency_ms_stddev": 0.8106684166479914, "mean_time_per_output_token_max_ms": 14.447774757287528, "mean_time_per_output_token_mean_ci95_high_ms": 14.559239494088132, "mean_time_per_output_token_mean_ci95_low_ms": 12.724524150073169, "mean_time_per_output_token_mean_ms": 13.64188182208065, "mean_time_per_output_token_min_ms": 12.826520860929405, "mean_time_per_output_token_p25_ms": 12.826520860929405, "mean_time_per_output_token_p50_ms": 13.65134984802502, "mean_time_per_output_token_p75_ms": 14.447774757287528, "mean_time_per_output_token_p90_ms": 14.447774757287528, "mean_time_per_output_token_p95_ms": 14.447774757287528, "mean_time_per_output_token_p99_ms": 14.447774757287528, "mean_time_per_output_token_samples": 3, "mean_time_per_output_token_stddev_ms": 0.8106684166479914, "output_token_sources": { "usage": 3 }, "output_tokens": 840, "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": 730, "request_canary_passed": 3, "request_canary_samples": 3, "throughput_performance_eligible": true, "throughput_tok_s": 9.693324570521726, "time_to_first_output_max_ms": 26157.44099999938, "time_to_first_output_mean_ci95_high_ms": 26337.946530326437, "time_to_first_output_mean_ci95_low_ms": 23871.81700300746, "time_to_first_output_mean_ms": 25104.88176666695, "time_to_first_output_min_ms": 23981.571899999835, "time_to_first_output_p25_ms": 23981.571899999835, "time_to_first_output_p50_ms": 25175.632400001632, "time_to_first_output_p75_ms": 26157.44099999938, "time_to_first_output_p90_ms": 26157.44099999938, "time_to_first_output_p95_ms": 26157.44099999938, "time_to_first_output_p99_ms": 26157.44099999938, "time_to_first_output_samples": 3, "time_to_first_output_stddev_ms": 1089.658581469972, "tpot_max_ms": 14.447774757287528, "tpot_mean_ci95_high_ms": 14.559239494088132, "tpot_mean_ci95_low_ms": 12.724524150073169, "tpot_mean_ms": 13.64188182208065, "tpot_min_ms": 12.826520860929405, "tpot_p25_ms": 12.826520860929405, "tpot_p50_ms": 13.65134984802502, "tpot_p75_ms": 14.447774757287528, "tpot_p90_ms": 14.447774757287528, "tpot_p95_ms": 14.447774757287528, "tpot_p99_ms": 14.447774757287528, "tpot_samples": 3, "tpot_stddev_ms": 0.8106684166479914, "ttft_max_ms": 30198.69739999922, "ttft_mean_ci95_high_ms": 30538.23609717252, "ttft_mean_ci95_low_ms": 26327.972036161285, "ttft_mean_ms": 28433.104066666903, "ttft_min_ms": 26490.712800001347, "ttft_p25_ms": 26490.712800001347, "ttft_p50_ms": 28609.902000000147, "ttft_p75_ms": 30198.69739999922, "ttft_p90_ms": 30198.69739999922, "ttft_p95_ms": 30198.69739999922, "ttft_p99_ms": 30198.69739999922, "ttft_samples": 3, "ttft_stddev_ms": 1860.3038946308002 }, "model": "qwen38-flash-next-strata-iq4xs-32k-c1-r36-m50", "performance_eligible": true, "prompt_cache_mode": "unique", "request_canaries": true, "request_timings": [ { "content_chunks": 33, "controlled_output": null, "decode_tok_s": 69.21481105563299, "e2e_ms": 26957.813500001066, "effective_prefill_tok_s": 309.19574542151054, "generation_ms": 2976.241600001231, "mean_inter_token_latency_ms": 14.447774757287528, "mean_time_per_output_token_ms": 14.447774757287528, "output_token_source": "usage", "output_tokens": 207, "planned_context_tokens": 8192, "prompt_tokens": 7415, "reasoning_chunks": 171, "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": 23981.571899999835, "tpot_ms": 14.447774757287528, "ttft_ms": 26490.712800001347, "visible_content_truncated": false, "visible_generation_ms": 467.1006999997189 }, { "content_chunks": 35, "controlled_output": null, "decode_tok_s": 73.25282929033372, "e2e_ms": 30648.735099999612, "effective_prefill_tok_s": 283.55220222039975, "generation_ms": 4491.294100000232, "mean_inter_token_latency_ms": 13.65134984802502, "mean_time_per_output_token_ms": 13.65134984802502, "output_token_source": "usage", "output_tokens": 330, "planned_context_tokens": 8192, "prompt_tokens": 7417, "reasoning_chunks": 292, "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": 26157.44099999938, "tpot_ms": 13.65134984802502, "ttft_ms": 30198.69739999922, "visible_content_truncated": false, "visible_generation_ms": 450.03770000039367 }, { "content_chunks": 33, "controlled_output": null, "decode_tok_s": 77.96346420377165, "e2e_ms": 29049.241700002312, "effective_prefill_tok_s": 294.6499965577635, "generation_ms": 3873.6093000006804, "mean_inter_token_latency_ms": 12.826520860929405, "mean_time_per_output_token_ms": 12.826520860929405, "output_token_source": "usage", "output_tokens": 303, "planned_context_tokens": 8192, "prompt_tokens": 7418, "reasoning_chunks": 267, "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": 25175.632400001632, "tpot_ms": 12.826520860929405, "ttft_ms": 28609.902000000147, "visible_content_truncated": false, "visible_generation_ms": 439.33970000216505 } ], "requests": 3, "response_words": 0, "responses_received": 3, "run_id": "benchmark-20261005T110641Z", "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": "2026-10-05T11:06:41Z", "started_at_unix_ns": 1791198401766420500, "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": 86657.57490000033 }