{ "base_url": "REDACTED_LOOPBACK_ENDPOINT", "clock_domain_id": null, "completed": 0, "concurrency": 4, "configuration_fingerprint": null, "context_distribution": { "32768": 4 }, "context_seed": 0, "context_tokens": 32768, "controlled_output_policy": "strict", "engine": "vllm-00a33e231", "failed": 4, "failures": [ { "controlled_output": { "capture_complete": true, "exact_adherence": false, "observed_code_words": 136, "observed_extra_words": 1, "passed": false, "policy": "strict", "requested_words": 128 }, "elapsed_ms": 10959.90795099351, "error_message": "REDACTED_BENCHMARK_CONTENT", "error_message_truncated": false, "error_type": "ValueError", "failure_class": "request_canary_failed", "performance_eligible": false, "planned_context_tokens": 32768, "request_canary": { "capture_complete": true, "foreign_marker_count": 0, "marker": "ANVIL_REQ_0_00000", "marker_at_start": false, "observed_marker_count": 1, "passed": false }, "request_index": 0, "response_metadata": { "content_chunks": 147, "finish_reasons": [ "stop" ], "output_token_source": "usage", "reasoning_chunks": 136, "stream_terminal_observed": true, "usage": { "completion_tokens": 287, "prompt_tokens": 30385, "total_tokens": 30672 }, "visible_content_truncated": false }, "response_received": true, "response_timing": { "content_chunks": 147, "controlled_output": { "capture_complete": true, "exact_adherence": false, "observed_code_words": 136, "observed_extra_words": 1, "passed": false, "policy": "strict", "requested_words": 128 }, "decode_tok_s": 34.978785297602485, "e2e_ms": 10956.94317399466, "effective_prefill_tok_s": 10927.66030400396, "generation_ms": 8176.384559003054, "mean_inter_token_latency_ms": 28.588757199311377, "mean_time_per_output_token_ms": 28.588757199311377, "output_token_source": "usage", "output_tokens": 287, "planned_context_tokens": 32768, "prompt_tokens": 30385, "reasoning_chunks": 136, "request_canary": { "capture_complete": true, "foreign_marker_count": 0, "marker": "ANVIL_REQ_0_00000", "marker_at_start": false, "observed_marker_count": 1, "passed": false }, "request_index": 0, "time_to_first_output_ms": 2780.5586149916053, "tpot_ms": 28.588757199311377, "ttft_ms": 9598.412909996114, "visible_content_truncated": false, "visible_generation_ms": 1358.5302639985457 } }, { "controlled_output": { "capture_complete": true, "exact_adherence": false, "observed_code_words": 121, "observed_extra_words": 1, "passed": false, "policy": "strict", "requested_words": 128 }, "elapsed_ms": 11075.394348998088, "error_message": "REDACTED_BENCHMARK_CONTENT", "error_message_truncated": false, "error_type": "ValueError", "failure_class": "request_canary_failed", "performance_eligible": false, "planned_context_tokens": 32768, "request_canary": { "capture_complete": true, "foreign_marker_count": 0, "marker": "ANVIL_REQ_0_00001", "marker_at_start": false, "observed_marker_count": 1, "passed": false }, "request_index": 1, "response_metadata": { "content_chunks": 132, "finish_reasons": [ "stop" ], "output_token_source": "usage", "reasoning_chunks": 152, "stream_terminal_observed": true, "usage": { "completion_tokens": 288, "prompt_tokens": 30388, "total_tokens": 30676 }, "visible_content_truncated": false }, "response_received": true, "response_timing": { "content_chunks": 132, "controlled_output": { "capture_complete": true, "exact_adherence": false, "observed_code_words": 121, "observed_extra_words": 1, "passed": false, "policy": "strict", "requested_words": 128 }, "decode_tok_s": 73.15030164921856, "e2e_ms": 11072.709936997853, "effective_prefill_tok_s": 4250.497274474489, "generation_ms": 3923.428797002998, "mean_inter_token_latency_ms": 13.670483613250864, "mean_time_per_output_token_ms": 13.670483613250864, "output_token_source": "usage", "output_tokens": 288, "planned_context_tokens": 32768, "prompt_tokens": 30388, "reasoning_chunks": 152, "request_canary": { "capture_complete": true, "foreign_marker_count": 0, "marker": "ANVIL_REQ_0_00001", "marker_at_start": false, "observed_marker_count": 1, "passed": false }, "request_index": 1, "time_to_first_output_ms": 7149.281139994855, "tpot_ms": 13.670483613250864, "ttft_ms": 9846.476131991949, "visible_content_truncated": false, "visible_generation_ms": 1226.2338050059043 } }, { "controlled_output": { "capture_complete": true, "exact_adherence": false, "observed_code_words": 132, "observed_extra_words": 1, "passed": false, "policy": "strict", "requested_words": 128 }, "elapsed_ms": 11036.795181993512, "error_message": "REDACTED_BENCHMARK_CONTENT", "error_message_truncated": false, "error_type": "ValueError", "failure_class": "request_canary_failed", "performance_eligible": false, "planned_context_tokens": 32768, "request_canary": { "capture_complete": true, "foreign_marker_count": 0, "marker": "ANVIL_REQ_0_00002", "marker_at_start": false, "observed_marker_count": 1, "passed": false }, "request_index": 2, "response_metadata": { "content_chunks": 143, "finish_reasons": [ "stop" ], "output_token_source": "usage", "reasoning_chunks": 133, "stream_terminal_observed": true, "usage": { "completion_tokens": 280, "prompt_tokens": 30390, "total_tokens": 30670 }, "visible_content_truncated": false }, "response_received": true, "response_timing": { "content_chunks": 143, "controlled_output": { "capture_complete": true, "exact_adherence": false, "observed_code_words": 132, "observed_extra_words": 1, "passed": false, "policy": "strict", "requested_words": 128 }, "decode_tok_s": 108.27865999353952, "e2e_ms": 11034.113122004783, "effective_prefill_tok_s": 3593.2909619735715, "generation_ms": 2576.6850090003572, "mean_inter_token_latency_ms": 9.235430139786226, "mean_time_per_output_token_ms": 9.235430139786226, "output_token_source": "usage", "output_tokens": 280, "planned_context_tokens": 32768, "prompt_tokens": 30390, "reasoning_chunks": 133, "request_canary": { "capture_complete": true, "foreign_marker_count": 0, "marker": "ANVIL_REQ_0_00002", "marker_at_start": false, "observed_marker_count": 1, "passed": false }, "request_index": 2, "time_to_first_output_ms": 8457.428113004426, "tpot_ms": 9.235430139786226, "ttft_ms": 9712.899541002116, "visible_content_truncated": false, "visible_generation_ms": 1321.2135810026666 } }, { "controlled_output": { "capture_complete": true, "exact_adherence": false, "observed_code_words": 126, "observed_extra_words": 1, "passed": false, "policy": "strict", "requested_words": 128 }, "elapsed_ms": 10899.14109800884, "error_message": "REDACTED_BENCHMARK_CONTENT", "error_message_truncated": false, "error_type": "ValueError", "failure_class": "request_canary_failed", "performance_eligible": false, "planned_context_tokens": 32768, "request_canary": { "capture_complete": true, "foreign_marker_count": 0, "marker": "ANVIL_REQ_0_00003", "marker_at_start": false, "observed_marker_count": 1, "passed": false }, "request_index": 3, "response_metadata": { "content_chunks": 137, "finish_reasons": [ "stop" ], "output_token_source": "usage", "reasoning_chunks": 135, "stream_terminal_observed": true, "usage": { "completion_tokens": 276, "prompt_tokens": 30392, "total_tokens": 30668 }, "visible_content_truncated": false }, "response_received": true, "response_timing": { "content_chunks": 137, "controlled_output": { "capture_complete": true, "exact_adherence": false, "observed_code_words": 126, "observed_extra_words": 1, "passed": false, "policy": "strict", "requested_words": 128 }, "decode_tok_s": 47.81606140866262, "e2e_ms": 10897.305935999611, "effective_prefill_tok_s": 5905.831211582749, "generation_ms": 5751.20559700008, "mean_inter_token_latency_ms": 20.91347489818211, "mean_time_per_output_token_ms": 20.91347489818211, "output_token_source": "usage", "output_tokens": 276, "planned_context_tokens": 32768, "prompt_tokens": 30392, "reasoning_chunks": 135, "request_canary": { "capture_complete": true, "foreign_marker_count": 0, "marker": "ANVIL_REQ_0_00003", "marker_at_start": false, "observed_marker_count": 1, "passed": false }, "request_index": 3, "time_to_first_output_ms": 5146.100338999531, "tpot_ms": 20.91347489818211, "ttft_ms": 9636.853545991471, "visible_content_truncated": false, "visible_generation_ms": 1260.4523900081404 } } ], "finished_at": "2026-10-05T08:45:24Z", "finished_at_unix_ns": 1791189924667609290, "gpu": "dual-96GiB-class-gpus", "manifest": null, "max_context_tokens": 262144, "max_tokens": 8192, "measurement_protocol": "capacity-v3", "metric_population": { "aggregate_throughput_eligible": false, "attempted_requests": 4, "excluded_requests": 4, "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": 0, "e2e": 0, "effective_prefill_tok_s": 0, "tpot": 0, "ttft": 0 }, "performance_eligible_requests": 0, "successful_requests": 0 }, "metrics": { "completion_tokens_max": null, "completion_tokens_mean": null, "completion_tokens_mean_ci95_high": null, "completion_tokens_mean_ci95_low": null, "completion_tokens_min": null, "completion_tokens_p25": null, "completion_tokens_p50": null, "completion_tokens_p75": null, "completion_tokens_p90": null, "completion_tokens_p95": null, "completion_tokens_p99": null, "completion_tokens_samples": 0, "completion_tokens_stddev": null, "content_chunks": 0, "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": null, "decode_tok_s_mean": null, "decode_tok_s_mean_ci95_high": null, "decode_tok_s_mean_ci95_low": null, "decode_tok_s_min": null, "decode_tok_s_p25": null, "decode_tok_s_p50": null, "decode_tok_s_p75": null, "decode_tok_s_p90": null, "decode_tok_s_p95": null, "decode_tok_s_p99": null, "decode_tok_s_samples": 0, "decode_tok_s_stddev": null, "e2e_max_ms": null, "e2e_mean_ci95_high_ms": null, "e2e_mean_ci95_low_ms": null, "e2e_mean_ms": null, "e2e_min_ms": null, "e2e_p25_ms": null, "e2e_p50_ms": null, "e2e_p75_ms": null, "e2e_p90_ms": null, "e2e_p95_ms": null, "e2e_p99_ms": null, "e2e_samples": 0, "e2e_stddev_ms": null, "effective_prefill_tok_s_max": null, "effective_prefill_tok_s_mean": null, "effective_prefill_tok_s_mean_ci95_high": null, "effective_prefill_tok_s_mean_ci95_low": null, "effective_prefill_tok_s_min": null, "effective_prefill_tok_s_p25": null, "effective_prefill_tok_s_p50": null, "effective_prefill_tok_s_p75": null, "effective_prefill_tok_s_p90": null, "effective_prefill_tok_s_p95": null, "effective_prefill_tok_s_p99": null, "effective_prefill_tok_s_samples": 0, "effective_prefill_tok_s_stddev": null, "generation_max_ms": null, "generation_mean_ci95_high_ms": null, "generation_mean_ci95_low_ms": null, "generation_mean_ms": null, "generation_min_ms": null, "generation_p25_ms": null, "generation_p50_ms": null, "generation_p75_ms": null, "generation_p90_ms": null, "generation_p95_ms": null, "generation_p99_ms": null, "generation_samples": 0, "generation_stddev_ms": null, "mean_inter_token_latency_ms_max": null, "mean_inter_token_latency_ms_mean": null, "mean_inter_token_latency_ms_mean_ci95_high": null, "mean_inter_token_latency_ms_mean_ci95_low": null, "mean_inter_token_latency_ms_min": null, "mean_inter_token_latency_ms_p25": null, "mean_inter_token_latency_ms_p50": null, "mean_inter_token_latency_ms_p75": null, "mean_inter_token_latency_ms_p90": null, "mean_inter_token_latency_ms_p95": null, "mean_inter_token_latency_ms_p99": null, "mean_inter_token_latency_ms_samples": 0, "mean_inter_token_latency_ms_stddev": null, "mean_time_per_output_token_max_ms": null, "mean_time_per_output_token_mean_ci95_high_ms": null, "mean_time_per_output_token_mean_ci95_low_ms": null, "mean_time_per_output_token_mean_ms": null, "mean_time_per_output_token_min_ms": null, "mean_time_per_output_token_p25_ms": null, "mean_time_per_output_token_p50_ms": null, "mean_time_per_output_token_p75_ms": null, "mean_time_per_output_token_p90_ms": null, "mean_time_per_output_token_p95_ms": null, "mean_time_per_output_token_p99_ms": null, "mean_time_per_output_token_samples": 0, "mean_time_per_output_token_stddev_ms": null, "output_token_sources": {}, "output_tokens": 0, "prefix_cache_hit_avg": null, "prompt_token_samples": 0, "prompt_tokens": null, "prompt_tokens_max": null, "prompt_tokens_mean": null, "prompt_tokens_mean_ci95_high": null, "prompt_tokens_mean_ci95_low": null, "prompt_tokens_min": null, "prompt_tokens_p25": null, "prompt_tokens_p50": null, "prompt_tokens_p75": null, "prompt_tokens_p90": null, "prompt_tokens_p95": null, "prompt_tokens_p99": null, "prompt_tokens_samples": 0, "prompt_tokens_stddev": null, "reasoning_chunks": 0, "request_canary_passed": 0, "request_canary_samples": 0, "throughput_performance_eligible": false, "throughput_tok_s": null, "time_to_first_output_max_ms": null, "time_to_first_output_mean_ci95_high_ms": null, "time_to_first_output_mean_ci95_low_ms": null, "time_to_first_output_mean_ms": null, "time_to_first_output_min_ms": null, "time_to_first_output_p25_ms": null, "time_to_first_output_p50_ms": null, "time_to_first_output_p75_ms": null, "time_to_first_output_p90_ms": null, "time_to_first_output_p95_ms": null, "time_to_first_output_p99_ms": null, "time_to_first_output_samples": 0, "time_to_first_output_stddev_ms": null, "tpot_max_ms": null, "tpot_mean_ci95_high_ms": null, "tpot_mean_ci95_low_ms": null, "tpot_mean_ms": null, "tpot_min_ms": null, "tpot_p25_ms": null, "tpot_p50_ms": null, "tpot_p75_ms": null, "tpot_p90_ms": null, "tpot_p95_ms": null, "tpot_p99_ms": null, "tpot_samples": 0, "tpot_stddev_ms": null, "ttft_max_ms": null, "ttft_mean_ci95_high_ms": null, "ttft_mean_ci95_low_ms": null, "ttft_mean_ms": null, "ttft_min_ms": null, "ttft_p25_ms": null, "ttft_p50_ms": null, "ttft_p75_ms": null, "ttft_p90_ms": null, "ttft_p95_ms": null, "ttft_p99_ms": null, "ttft_samples": 0, "ttft_stddev_ms": null }, "model": "candidate-local-tier-r1", "performance_eligible": false, "prompt_cache_mode": "unique", "request_canaries": true, "request_timings": [], "requests": 4, "response_words": 128, "responses_received": 4, "run_id": "REDACTED_RUN_OR_WORKER_ID", "sampling": { "temperature": { "effective_request": 1.0, "requested": 1.0, "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": "medium", "shared_prefix_burst": false, "thinking_mode": "default" }, "shared_prefix_tokens": 0, "source_recipe": null, "started_at": "2026-10-05T08:45:13Z", "started_at_unix_ns": 1791189913587431566, "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": 11080.170479996013 }