{ "base_url": "http://127.0.0.1:8080/v1", "clock_domain_id": null, "completed": 3, "concurrency": 1, "configuration_fingerprint": null, "context_distribution": { "154624": 3 }, "context_seed": 73, "context_tokens": 154624, "controlled_output_policy": "strict", "engine": "ninfer", "failed": 0, "failures": [], "finished_at": "2026-09-20T01:17:08Z", "finished_at_unix_ns": 1789867028224810100, "gpu": null, "manifest": null, "max_context_tokens": 163840, "max_tokens": 512, "measurement_protocol": "capacity-v3", "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": 46.0, "completion_tokens_mean": 46.0, "completion_tokens_mean_ci95_high": 46.0, "completion_tokens_mean_ci95_low": 46.0, "completion_tokens_min": 46.0, "completion_tokens_p25": 46.0, "completion_tokens_p50": 46.0, "completion_tokens_p75": 46.0, "completion_tokens_p90": 46.0, "completion_tokens_p95": 46.0, "completion_tokens_p99": 46.0, "completion_tokens_samples": 3, "completion_tokens_stddev": 0.0, "content_chunks": 39, "content_chunks_s": null, "controlled_output_exact_adherence": 3, "controlled_output_exact_adherence_rate": 1.0, "controlled_output_samples": 3, "controlled_output_unobservable": 0, "decode_tok_s_max": 163.38338673804404, "decode_tok_s_mean": 158.78908656683458, "decode_tok_s_mean_ci95_high": 167.11426560902405, "decode_tok_s_mean_ci95_low": 150.46390752464512, "decode_tok_s_min": 150.30371371878837, "decode_tok_s_p25": 150.30371371878837, "decode_tok_s_p50": 162.6801592436713, "decode_tok_s_p75": 163.38338673804404, "decode_tok_s_p90": 163.38338673804404, "decode_tok_s_p95": 163.38338673804404, "decode_tok_s_p99": 163.38338673804404, "decode_tok_s_samples": 3, "decode_tok_s_stddev": 7.356955654683547, "e2e_max_ms": 46472.61749999598, "e2e_mean_ci95_high_ms": 46498.79222861245, "e2e_mean_ci95_low_ms": 46308.745837990165, "e2e_mean_ms": 46403.76903330131, "e2e_min_ms": 46310.21609995514, "e2e_p25_ms": 46310.21609995514, "e2e_p50_ms": 46428.473499952815, "e2e_p75_ms": 46472.61749999598, "e2e_p90_ms": 46472.61749999598, "e2e_p95_ms": 46472.61749999598, "e2e_p99_ms": 46472.61749999598, "e2e_samples": 3, "e2e_stddev_ms": 83.97193988593634, "effective_prefill_tok_s_max": 3237.303585423008, "effective_prefill_tok_s_mean": 3229.645448103935, "effective_prefill_tok_s_mean_ci95_high": 3237.3758530271825, "effective_prefill_tok_s_mean_ci95_low": 3221.9150431806875, "effective_prefill_tok_s_min": 3224.178667986695, "effective_prefill_tok_s_p25": 3224.178667986695, "effective_prefill_tok_s_p50": 3227.4540909021034, "effective_prefill_tok_s_p75": 3237.303585423008, "effective_prefill_tok_s_p90": 3237.303585423008, "effective_prefill_tok_s_p95": 3237.303585423008, "effective_prefill_tok_s_p99": 3237.303585423008, "effective_prefill_tok_s_samples": 3, "effective_prefill_tok_s_stddev": 6.831354127625315, "generation_max_ms": 299.3937999708578, "generation_mean_ci95_high_ms": 299.097015727036, "generation_mean_ci95_low_ms": 268.5269842516901, "generation_mean_ms": 283.8119999893631, "generation_min_ms": 275.42580000590533, "generation_p25_ms": 275.42580000590533, "generation_p50_ms": 276.61639999132603, "generation_p75_ms": 299.3937999708578, "generation_p90_ms": 299.3937999708578, "generation_p95_ms": 299.3937999708578, "generation_p99_ms": 299.3937999708578, "generation_samples": 3, "generation_stddev_ms": 13.507359108234413, "mean_inter_token_latency_ms_max": 6.653195554907951, "mean_inter_token_latency_ms_mean": 6.306933333096957, "mean_inter_token_latency_ms_mean_ci95_high": 6.646600349489689, "mean_inter_token_latency_ms_mean_ci95_low": 5.967266316704225, "mean_inter_token_latency_ms_min": 6.120573333464563, "mean_inter_token_latency_ms_p25": 6.120573333464563, "mean_inter_token_latency_ms_p50": 6.147031110918356, "mean_inter_token_latency_ms_p75": 6.653195554907951, "mean_inter_token_latency_ms_p90": 6.653195554907951, "mean_inter_token_latency_ms_p95": 6.653195554907951, "mean_inter_token_latency_ms_p99": 6.653195554907951, "mean_inter_token_latency_ms_samples": 3, "mean_inter_token_latency_ms_stddev": 0.3001635357385424, "mean_time_per_output_token_max_ms": 6.653195554907951, "mean_time_per_output_token_mean_ci95_high_ms": 6.646600349489689, "mean_time_per_output_token_mean_ci95_low_ms": 5.967266316704225, "mean_time_per_output_token_mean_ms": 6.306933333096957, "mean_time_per_output_token_min_ms": 6.120573333464563, "mean_time_per_output_token_p25_ms": 6.120573333464563, "mean_time_per_output_token_p50_ms": 6.147031110918356, "mean_time_per_output_token_p75_ms": 6.653195554907951, "mean_time_per_output_token_p90_ms": 6.653195554907951, "mean_time_per_output_token_p95_ms": 6.653195554907951, "mean_time_per_output_token_p99_ms": 6.653195554907951, "mean_time_per_output_token_samples": 3, "mean_time_per_output_token_stddev_ms": 0.3001635357385424, "output_token_sources": { "usage": 3 }, "output_tokens": 138, "prefix_cache_hit_avg": null, "prompt_token_samples": 3, "prompt_tokens": 446852, "prompt_tokens_max": 148953.0, "prompt_tokens_mean": 148950.66666666666, "prompt_tokens_mean_ci95_high": 148953.5144806431, "prompt_tokens_mean_ci95_low": 148947.81885269022, "prompt_tokens_min": 148948.0, "prompt_tokens_p25": 148948.0, "prompt_tokens_p50": 148951.0, "prompt_tokens_p75": 148953.0, "prompt_tokens_p90": 148953.0, "prompt_tokens_p95": 148953.0, "prompt_tokens_p99": 148953.0, "prompt_tokens_samples": 3, "prompt_tokens_stddev": 2.516611478423583, "reasoning_chunks": 0, "request_canary_passed": 3, "request_canary_samples": 3, "throughput_performance_eligible": true, "throughput_tok_s": 0.9912278625483528, "time_to_first_output_max_ms": 46197.19169999007, "time_to_first_output_mean_ci95_high_ms": 46229.94196369342, "time_to_first_output_mean_ci95_low_ms": 46009.97210293048, "time_to_first_output_mean_ms": 46119.95703331195, "time_to_first_output_min_ms": 46010.822299984284, "time_to_first_output_p25_ms": 46010.822299984284, "time_to_first_output_p50_ms": 46151.85709996149, "time_to_first_output_p75_ms": 46197.19169999007, "time_to_first_output_p90_ms": 46197.19169999007, "time_to_first_output_p95_ms": 46197.19169999007, "time_to_first_output_p99_ms": 46197.19169999007, "time_to_first_output_samples": 3, "time_to_first_output_stddev_ms": 97.19361606511612, "tpot_max_ms": 6.653195554907951, "tpot_mean_ci95_high_ms": 6.646600349489689, "tpot_mean_ci95_low_ms": 5.967266316704225, "tpot_mean_ms": 6.306933333096957, "tpot_min_ms": 6.120573333464563, "tpot_p25_ms": 6.120573333464563, "tpot_p50_ms": 6.147031110918356, "tpot_p75_ms": 6.653195554907951, "tpot_p90_ms": 6.653195554907951, "tpot_p95_ms": 6.653195554907951, "tpot_p99_ms": 6.653195554907951, "tpot_samples": 3, "tpot_stddev_ms": 0.3001635357385424, "ttft_max_ms": 46197.19299999997, "ttft_mean_ci95_high_ms": 46229.94294756114, "ttft_mean_ci95_low_ms": 46009.97278574211, "ttft_mean_ms": 46119.957866651624, "ttft_min_ms": 46010.823099990375, "ttft_p25_ms": 46010.823099990375, "ttft_p50_ms": 46151.857499964535, "ttft_p75_ms": 46197.19299999997, "ttft_p90_ms": 46197.19299999997, "ttft_p95_ms": 46197.19299999997, "ttft_p99_ms": 46197.19299999997, "ttft_samples": 3, "ttft_stddev_ms": 97.19374908665733 }, "model": "qwen38-huihui-ninfer-schemafix-mtp3-160k", "performance_eligible": true, "prompt_cache_mode": "unique", "request_canaries": true, "request_timings": [ { "content_chunks": 13, "controlled_output": { "capture_complete": true, "exact_adherence": true, "observed_code_words": 32, "observed_extra_words": 0, "passed": true, "policy": "strict", "requested_words": 32 }, "decode_tok_s": 163.38338673804404, "e2e_ms": 46472.61749999598, "effective_prefill_tok_s": 3224.178667986695, "generation_ms": 275.42580000590533, "mean_inter_token_latency_ms": 6.120573333464563, "mean_time_per_output_token_ms": 6.120573333464563, "output_token_source": "usage", "output_tokens": 46, "planned_context_tokens": 154624, "prompt_tokens": 148948, "reasoning_chunks": 0, "request_canary": { "capture_complete": true, "foreign_marker_count": 0, "marker": "ANVIL_REQ_73_00000", "marker_at_start": true, "observed_marker_count": 1, "passed": true }, "request_index": 0, "time_to_first_output_ms": 46197.19169999007, "tpot_ms": 6.120573333464563, "ttft_ms": 46197.19299999997, "visible_content_truncated": false, "visible_generation_ms": 275.4244999960065 }, { "content_chunks": 13, "controlled_output": { "capture_complete": true, "exact_adherence": true, "observed_code_words": 32, "observed_extra_words": 0, "passed": true, "policy": "strict", "requested_words": 32 }, "decode_tok_s": 150.30371371878837, "e2e_ms": 46310.21609995514, "effective_prefill_tok_s": 3237.303585423008, "generation_ms": 299.3937999708578, "mean_inter_token_latency_ms": 6.653195554907951, "mean_time_per_output_token_ms": 6.653195554907951, "output_token_source": "usage", "output_tokens": 46, "planned_context_tokens": 154624, "prompt_tokens": 148951, "reasoning_chunks": 0, "request_canary": { "capture_complete": true, "foreign_marker_count": 0, "marker": "ANVIL_REQ_73_00001", "marker_at_start": true, "observed_marker_count": 1, "passed": true }, "request_index": 1, "time_to_first_output_ms": 46010.822299984284, "tpot_ms": 6.653195554907951, "ttft_ms": 46010.823099990375, "visible_content_truncated": false, "visible_generation_ms": 299.3929999647662 }, { "content_chunks": 13, "controlled_output": { "capture_complete": true, "exact_adherence": true, "observed_code_words": 32, "observed_extra_words": 0, "passed": true, "policy": "strict", "requested_words": 32 }, "decode_tok_s": 162.6801592436713, "e2e_ms": 46428.473499952815, "effective_prefill_tok_s": 3227.4540909021034, "generation_ms": 276.61639999132603, "mean_inter_token_latency_ms": 6.147031110918356, "mean_time_per_output_token_ms": 6.147031110918356, "output_token_source": "usage", "output_tokens": 46, "planned_context_tokens": 154624, "prompt_tokens": 148953, "reasoning_chunks": 0, "request_canary": { "capture_complete": true, "foreign_marker_count": 0, "marker": "ANVIL_REQ_73_00002", "marker_at_start": true, "observed_marker_count": 1, "passed": true }, "request_index": 2, "time_to_first_output_ms": 46151.85709996149, "tpot_ms": 6.147031110918356, "ttft_ms": 46151.857499964535, "visible_content_truncated": false, "visible_generation_ms": 276.61599998828024 } ], "requests": 3, "response_words": 32, "responses_received": 3, "run_id": "benchmark-20260920T011449Z", "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": null, "shared_prefix_burst": false, "thinking_mode": "unsupported" }, "shared_prefix_tokens": 0, "source_recipe": "REDACTED_LOCAL_PATH/.anvil-serving\\evidence\\2026-09-19-qwen38-context-envelope\\serve-recipes.160k.toml#lyf/Qwen3.8-27B-Huihui-NInfer-NVFP4-MTP3-160K-Baked", "started_at": "2026-09-20T01:14:49Z", "started_at_unix_ns": 1789866889005631600, "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": 139221.26809996553 }