{ "schema": "anvil-serving.benchmark-decision-summary/v1", "campaign_id": "2026-10-21-flare-glm53-trial", "complete": false, "scout_execution_complete": true, "identity": { "source_revision": "80ad7e5c5de49221ba1e75fa133ef358af3e72e3", "dirty_state": "bounded preflight sampling, inspection-image and seccomp changes; source hashes retained", "hardware": "2x RTX PRO 6000 Blackwell Max-Q 96 GB; native Linux", "flare": { "engine_revision": "47581ccac82660953a5b59cb04b2750d64edc97d", "model": "local-inference-lab/GLM-5.3-Flash-NVFP4-MXFP8-CSF-QAD", "revision": "dec48abd33efa73c3bb7c95b74eee10cad34f9be", "image": "sha256:8908da0c34a446d7aa6a8a907e7cedc100a49a85a78848c624ba47ef64e21f37", "topology": "TP2/C4/MTP3; prefill chunk4096; native runtime" }, "incumbent": { "model": "nvidia/GLM-5.3-Flash-FP4-Local", "revision": "76348ddb2be2e35e39ecfc00e75dc07703f4f3d5", "image": "sha256:745feb916ee816f5409fd58386114be4f22fbede7f9c0e0a7b8c4fe8dac0fb29", "topology": "vLLM TP2/DCP2/C4/MTP3; batch4096; FP8 KV" } }, "workload": { "temperature": 1, "top_p": 0.95, "max_tokens": 2048, "response_words": 0, "controlled_output_policy": "observe", "context_seed": 0, "prompt_cache_mode": "unique within run; same namespace across cells", "request_canaries": true, "order": "10240/C1,10240/C4,32768/C1,32768/C4 per arm; same preflight before natural cells" }, "measurement": { "path": "online direct", "warm_cold_state": "C1 near-zero cache hits; C4 100 percent hits after preceding C1", "sample_count": 64, "statistics": [ "p50", "p95", "native distributions; p99 descriptive only at n8" ] }, "gates": [ { "name": "Flare initial and warm protocol preflight", "status": "passed 7/7 each" }, { "name": "natural scout responses and canaries", "status": "passed 32/32 per arm" }, { "name": "strict 128-word controlled output", "status": "failed 0/4 eligible in C1 and C4 for each arm" }, { "name": "scoped restoration", "status": "verified exact identity, direct7/7, routed7/7 and readmission" } ], "headline_results": [ { "arm": "flare", "context_budget": 10240, "concurrency": 1, "cache_state": "near-zero reported hit ratio", "artifact": "raw/flare-natural-ctx10240-c1.json", "requests": 8, "completed": 8, "performance_eligible": true, "metrics": { "completion_tokens_max": 478.0, "completion_tokens_mean": 272.0, "completion_tokens_mean_ci95_high": 351.43442578630504, "completion_tokens_mean_ci95_low": 192.565574213695, "completion_tokens_min": 95.0, "completion_tokens_p25": 171.0, "completion_tokens_p50": 248.0, "completion_tokens_p75": 312.0, "completion_tokens_p90": 478.0, "completion_tokens_p95": 478.0, "completion_tokens_p99": 478.0, "completion_tokens_samples": 8, "completion_tokens_stddev": 114.6298390472568, "content_chunks": 485, "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": 205.09424020698404, "decode_tok_s_mean": 193.50178044806765, "decode_tok_s_mean_ci95_high": 200.2292147664627, "decode_tok_s_mean_ci95_low": 186.7743461296726, "decode_tok_s_min": 176.94381571847478, "decode_tok_s_p25": 184.9362836698744, "decode_tok_s_p50": 192.46979497685334, "decode_tok_s_p75": 202.20997592784684, "decode_tok_s_p90": 205.09424020698404, "decode_tok_s_p95": 205.09424020698404, "decode_tok_s_p99": 205.09424020698404, "decode_tok_s_samples": 8, "decode_tok_s_stddev": 9.708192707192325, "e2e_max_ms": 3089.8574750171974, "e2e_mean_ci95_high_ms": 2549.6443755091454, "e2e_mean_ci95_low_ms": 1778.0630632496618, "e2e_mean_ms": 2163.8537193794036, "e2e_min_ms": 1271.2004219647497, "e2e_p25_ms": 1656.2796760117635, "e2e_p50_ms": 2107.206608983688, "e2e_p75_ms": 2453.5648570163175, "e2e_p90_ms": 3089.8574750171974, "e2e_p95_ms": 3089.8574750171974, "e2e_p99_ms": 3089.8574750171974, "e2e_samples": 8, "e2e_stddev_ms": 556.7248756484652, "effective_prefill_tok_s_max": 10053.69239721204, "effective_prefill_tok_s_mean": 9976.683170995868, "effective_prefill_tok_s_mean_ci95_high": 10012.246104920354, "effective_prefill_tok_s_mean_ci95_low": 9941.120237071382, "effective_prefill_tok_s_min": 9890.870390237693, "effective_prefill_tok_s_p25": 9937.59313145853, "effective_prefill_tok_s_p50": 9961.381324297408, "effective_prefill_tok_s_p75": 10000.706715824008, "effective_prefill_tok_s_p90": 10053.69239721204, "effective_prefill_tok_s_p95": 10053.69239721204, "effective_prefill_tok_s_p99": 10053.69239721204, "effective_prefill_tok_s_samples": 8, "effective_prefill_tok_s_stddev": 51.31998313651622, "generation_max_ms": 2325.760097010061, "generation_mean_ci95_high_ms": 1782.03582743944, "generation_mean_ci95_low_ms": 1005.8710235688616, "generation_mean_ms": 1393.9534255041508, "generation_min_ms": 494.8278989177197, "generation_p25_ms": 883.2554740365595, "generation_p50_ms": 1335.5951309204102, "generation_p75_ms": 1685.5191360227764, "generation_p90_ms": 2325.760097010061, "generation_p95_ms": 2325.760097010061, "generation_p99_ms": 2325.760097010061, "generation_samples": 8, "generation_stddev_ms": 560.0320368726661, "mean_inter_token_latency_ms_max": 5.651511446950161, "mean_inter_token_latency_ms_mean": 5.179540692442038, "mean_inter_token_latency_ms_mean_ci95_high": 5.363314866723622, "mean_inter_token_latency_ms_mean_ci95_low": 4.995766518160455, "mean_inter_token_latency_ms_min": 4.87580733125799, "mean_inter_token_latency_ms_p25": 4.928418526382387, "mean_inter_token_latency_ms_p50": 5.168219049897124, "mean_inter_token_latency_ms_p75": 5.264126584231061, "mean_inter_token_latency_ms_p90": 5.651511446950161, "mean_inter_token_latency_ms_p95": 5.651511446950161, "mean_inter_token_latency_ms_p99": 5.651511446950161, "mean_inter_token_latency_ms_samples": 8, "mean_inter_token_latency_ms_stddev": 0.2651999282478897, "mean_time_per_output_token_max_ms": 5.651511446950161, "mean_time_per_output_token_mean_ci95_high_ms": 5.363314866723622, "mean_time_per_output_token_mean_ci95_low_ms": 4.995766518160455, "mean_time_per_output_token_mean_ms": 5.179540692442038, "mean_time_per_output_token_min_ms": 4.87580733125799, "mean_time_per_output_token_p25_ms": 4.928418526382387, "mean_time_per_output_token_p50_ms": 5.168219049897124, "mean_time_per_output_token_p75_ms": 5.264126584231061, "mean_time_per_output_token_p90_ms": 5.651511446950161, "mean_time_per_output_token_p95_ms": 5.651511446950161, "mean_time_per_output_token_p99_ms": 5.651511446950161, "mean_time_per_output_token_samples": 8, "mean_time_per_output_token_stddev_ms": 0.2651999282478897, "output_token_sources": { "usage": 8 }, "output_tokens": 2176, "prefix_cache_hit_avg": 0.001171741522840524, "prompt_token_samples": 8, "prompt_tokens": 61447, "prompt_tokens_max": 7682.0, "prompt_tokens_mean": 7680.875, "prompt_tokens_mean_ci95_high": 7681.655272388336, "prompt_tokens_mean_ci95_low": 7680.094727611664, "prompt_tokens_min": 7679.0, "prompt_tokens_p25": 7680.0, "prompt_tokens_p50": 7681.0, "prompt_tokens_p75": 7682.0, "prompt_tokens_p90": 7682.0, "prompt_tokens_p95": 7682.0, "prompt_tokens_p99": 7682.0, "prompt_tokens_samples": 8, "prompt_tokens_stddev": 1.1259916264596033, "reasoning_chunks": 1674, "request_canary_passed": 8, "request_canary_samples": 8, "throughput_performance_eligible": true, "throughput_tok_s": 125.55796552003589, "time_to_first_output_max_ms": 776.37252304703, "time_to_first_output_mean_ci95_high_ms": 772.6219260880754, "time_to_first_output_mean_ci95_low_ms": 767.1786616624303, "time_to_first_output_mean_ms": 769.9002938752528, "time_to_first_output_min_ms": 764.0973780071363, "time_to_first_output_p25_ms": 766.3153519388288, "time_to_first_output_p50_ms": 768.5575089417398, "time_to_first_output_p75_ms": 771.6114780632779, "time_to_first_output_p90_ms": 776.37252304703, "time_to_first_output_p95_ms": 776.37252304703, "time_to_first_output_p99_ms": 776.37252304703, "time_to_first_output_samples": 8, "time_to_first_output_stddev_ms": 3.9275195787400192, "tpot_max_ms": 5.651511446950161, "tpot_mean_ci95_high_ms": 5.363314866723622, "tpot_mean_ci95_low_ms": 4.995766518160455, "tpot_mean_ms": 5.179540692442038, "tpot_min_ms": 4.87580733125799, "tpot_p25_ms": 4.928418526382387, "tpot_p50_ms": 5.168219049897124, "tpot_p75_ms": 5.264126584231061, "tpot_p90_ms": 5.651511446950161, "tpot_p95_ms": 5.651511446950161, "tpot_p99_ms": 5.651511446950161, "tpot_samples": 8, "tpot_stddev_ms": 0.2651999282478897, "ttft_max_ms": 2800.771370064467, "ttft_mean_ci95_high_ms": 2255.4769949359084, "ttft_mean_ci95_low_ms": 1470.785441079623, "ttft_mean_ms": 1863.1312180077657, "ttft_min_ms": 948.4395879553631, "ttft_p25_ms": 1317.667520022951, "ttft_p50_ms": 1855.9025480644777, "ttft_p75_ms": 2159.4590159365907, "ttft_p90_ms": 2800.771370064467, "ttft_p95_ms": 2800.771370064467, "ttft_p99_ms": 2800.771370064467, "ttft_samples": 8, "ttft_stddev_ms": 566.1844070118249 } }, { "arm": "incumbent", "context_budget": 10240, "concurrency": 1, "cache_state": "near-zero reported hit ratio", "artifact": "raw/incumbent-natural-ctx10240-c1.json", "requests": 8, "completed": 8, "performance_eligible": true, "metrics": { "completion_tokens_max": 485.0, "completion_tokens_mean": 289.5, "completion_tokens_mean_ci95_high": 363.38724653145493, "completion_tokens_mean_ci95_low": 215.61275346854507, "completion_tokens_min": 161.0, "completion_tokens_p25": 176.0, "completion_tokens_p50": 261.0, "completion_tokens_p75": 350.0, "completion_tokens_p90": 485.0, "completion_tokens_p95": 485.0, "completion_tokens_p99": 485.0, "completion_tokens_samples": 8, "completion_tokens_stddev": 106.62484299100817, "content_chunks": 157, "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": 213.45600019777473, "decode_tok_s_mean": 200.6991554216142, "decode_tok_s_mean_ci95_high": 205.17360336137733, "decode_tok_s_mean_ci95_low": 196.22470748185106, "decode_tok_s_min": 194.10751109000236, "decode_tok_s_p25": 194.81667071789502, "decode_tok_s_p50": 198.616648358572, "decode_tok_s_p75": 201.87090312337935, "decode_tok_s_p90": 213.45600019777473, "decode_tok_s_p95": 213.45600019777473, "decode_tok_s_p99": 213.45600019777473, "decode_tok_s_samples": 8, "decode_tok_s_stddev": 6.456964245454492, "e2e_max_ms": 3348.137261928059, "e2e_mean_ci95_high_ms": 2682.7081975540536, "e2e_mean_ci95_low_ms": 1914.459653616471, "e2e_mean_ms": 2298.5839255852625, "e2e_min_ms": 1671.6018830193207, "e2e_p25_ms": 1759.116111905314, "e2e_p50_ms": 2111.882359953597, "e2e_p75_ms": 2615.115041960962, "e2e_p90_ms": 3348.137261928059, "e2e_p95_ms": 3348.137261928059, "e2e_p99_ms": 3348.137261928059, "e2e_samples": 8, "e2e_stddev_ms": 554.3201582193429, "effective_prefill_tok_s_max": 9027.308911214079, "effective_prefill_tok_s_mean": 8969.325329417454, "effective_prefill_tok_s_mean_ci95_high": 8994.231725631687, "effective_prefill_tok_s_mean_ci95_low": 8944.41893320322, "effective_prefill_tok_s_min": 8922.72478452453, "effective_prefill_tok_s_p25": 8942.642319759992, "effective_prefill_tok_s_p50": 8953.784502289207, "effective_prefill_tok_s_p75": 8985.885165607184, "effective_prefill_tok_s_p90": 9027.308911214079, "effective_prefill_tok_s_p95": 9027.308911214079, "effective_prefill_tok_s_p99": 9027.308911214079, "effective_prefill_tok_s_samples": 8, "effective_prefill_tok_s_stddev": 35.941799302048416, "generation_max_ms": 2493.4635310200974, "generation_mean_ci95_high_ms": 1827.4815602311367, "generation_mean_ci95_low_ms": 1056.963523736739, "generation_mean_ms": 1442.2225419839378, "generation_min_ms": 812.5716629438102, "generation_p25_ms": 898.2804159168154, "generation_p50_ms": 1261.1304830061272, "generation_p75_ms": 1757.1538080228493, "generation_p90_ms": 2493.4635310200974, "generation_p95_ms": 2493.4635310200974, "generation_p99_ms": 2493.4635310200974, "generation_samples": 8, "generation_stddev_ms": 555.9576822772779, "mean_inter_token_latency_ms_max": 5.151784155000201, "mean_inter_token_latency_ms_mean": 4.986975919232905, "mean_inter_token_latency_ms_mean_ci95_high": 5.095183575182536, "mean_inter_token_latency_ms_mean_ci95_low": 4.878768263283274, "mean_inter_token_latency_ms_min": 4.684806232073419, "mean_inter_token_latency_ms_p25": 4.850501857715874, "mean_inter_token_latency_ms_p50": 5.008625701023005, "mean_inter_token_latency_ms_p75": 5.078572893398814, "mean_inter_token_latency_ms_p90": 5.151784155000201, "mean_inter_token_latency_ms_p95": 5.151784155000201, "mean_inter_token_latency_ms_p99": 5.151784155000201, "mean_inter_token_latency_ms_samples": 8, "mean_inter_token_latency_ms_stddev": 0.15615176999649993, "mean_time_per_output_token_max_ms": 5.151784155000201, "mean_time_per_output_token_mean_ci95_high_ms": 5.095183575182536, "mean_time_per_output_token_mean_ci95_low_ms": 4.878768263283274, "mean_time_per_output_token_mean_ms": 4.986975919232905, "mean_time_per_output_token_min_ms": 4.684806232073419, "mean_time_per_output_token_p25_ms": 4.850501857715874, "mean_time_per_output_token_p50_ms": 5.008625701023005, "mean_time_per_output_token_p75_ms": 5.078572893398814, "mean_time_per_output_token_p90_ms": 5.151784155000201, "mean_time_per_output_token_p95_ms": 5.151784155000201, "mean_time_per_output_token_p99_ms": 5.151784155000201, "mean_time_per_output_token_samples": 8, "mean_time_per_output_token_stddev_ms": 0.15615176999649993, "output_token_sources": { "usage": 8 }, "output_tokens": 2316, "prefix_cache_hit_avg": 0.0, "prompt_token_samples": 8, "prompt_tokens": 61447, "prompt_tokens_max": 7682.0, "prompt_tokens_mean": 7680.875, "prompt_tokens_mean_ci95_high": 7681.655272388336, "prompt_tokens_mean_ci95_low": 7680.094727611664, "prompt_tokens_min": 7679.0, "prompt_tokens_p25": 7680.0, "prompt_tokens_p50": 7681.0, "prompt_tokens_p75": 7682.0, "prompt_tokens_p90": 7682.0, "prompt_tokens_p95": 7682.0, "prompt_tokens_p99": 7682.0, "prompt_tokens_samples": 8, "prompt_tokens_stddev": 1.1259916264596033, "reasoning_chunks": 617, "request_canary_passed": 8, "request_canary_samples": 8, "throughput_performance_eligible": true, "throughput_tok_s": 125.81587140068487, "time_to_first_output_max_ms": 860.8356959884986, "time_to_first_output_mean_ci95_high_ms": 858.7612455807147, "time_to_first_output_mean_ci95_low_ms": 853.9615216219346, "time_to_first_output_mean_ms": 856.3613836013246, "time_to_first_output_min_ms": 850.75187694747, "time_to_first_output_p25_ms": 852.3556960280985, "time_to_first_output_p50_ms": 857.0858809398487, "time_to_first_output_p75_ms": 858.1967339850962, "time_to_first_output_p90_ms": 860.8356959884986, "time_to_first_output_p95_ms": 860.8356959884986, "time_to_first_output_p99_ms": 860.8356959884986, "time_to_first_output_samples": 8, "time_to_first_output_stddev_ms": 3.4631809786498824, "tpot_max_ms": 5.151784155000201, "tpot_mean_ci95_high_ms": 5.095183575182536, "tpot_mean_ci95_low_ms": 4.878768263283274, "tpot_mean_ms": 4.986975919232905, "tpot_min_ms": 4.684806232073419, "tpot_p25_ms": 4.850501857715874, "tpot_p50_ms": 5.008625701023005, "tpot_p75_ms": 5.078572893398814, "tpot_p90_ms": 5.151784155000201, "tpot_p95_ms": 5.151784155000201, "tpot_p99_ms": 5.151784155000201, "tpot_samples": 8, "tpot_stddev_ms": 0.15615176999649993, "ttft_max_ms": 3077.7389080030844, "ttft_mean_ci95_high_ms": 2397.3253977065046, "ttft_mean_ci95_low_ms": 1600.4591662740263, "ttft_mean_ms": 1998.8922819902655, "ttft_min_ms": 1307.9173170262948, "ttft_p25_ms": 1442.5693129887804, "ttft_p50_ms": 1824.5955089805648, "ttft_p75_ms": 2299.055816954933, "ttft_p90_ms": 3077.7389080030844, "ttft_p95_ms": 3077.7389080030844, "ttft_p99_ms": 3077.7389080030844, "ttft_samples": 8, "ttft_stddev_ms": 574.9688938208922 } }, { "arm": "flare", "context_budget": 10240, "concurrency": 4, "cache_state": "warm across cells after C1", "artifact": "raw/flare-natural-ctx10240-c4.json", "requests": 8, "completed": 8, "performance_eligible": true, "metrics": { "completion_tokens_max": 387.0, "completion_tokens_mean": 265.875, "completion_tokens_mean_ci95_high": 328.5889157205161, "completion_tokens_mean_ci95_low": 203.16108427948387, "completion_tokens_min": 161.0, "completion_tokens_p25": 181.0, "completion_tokens_p50": 225.0, "completion_tokens_p75": 316.0, "completion_tokens_p90": 387.0, "completion_tokens_p95": 387.0, "completion_tokens_p99": 387.0, "completion_tokens_samples": 8, "completion_tokens_stddev": 90.50088791987466, "content_chunks": 485, "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": 87.31333346575815, "decode_tok_s_mean": 73.9529831083895, "decode_tok_s_mean_ci95_high": 78.0075243049015, "decode_tok_s_mean_ci95_low": 69.89844191187751, "decode_tok_s_min": 69.54799636945198, "decode_tok_s_p25": 70.7737066428023, "decode_tok_s_p50": 71.00887070674433, "decode_tok_s_p75": 74.14889223341659, "decode_tok_s_p90": 87.31333346575815, "decode_tok_s_p95": 87.31333346575815, "decode_tok_s_p99": 87.31333346575815, "decode_tok_s_samples": 8, "decode_tok_s_stddev": 5.851007295211942, "e2e_max_ms": 5482.929053949192, "e2e_mean_ci95_high_ms": 4496.777693094385, "e2e_mean_ci95_low_ms": 2757.280486904153, "e2e_mean_ms": 3627.0290899992688, "e2e_min_ms": 2202.5491460226476, "e2e_p25_ms": 2328.7220239872113, "e2e_p50_ms": 3187.5118321040645, "e2e_p75_ms": 4277.376459911466, "e2e_p90_ms": 5482.929053949192, "e2e_p95_ms": 5482.929053949192, "e2e_p99_ms": 5482.929053949192, "e2e_samples": 8, "e2e_stddev_ms": 1255.1125207675173, "effective_prefill_tok_s_max": 527652.9423398477, "effective_prefill_tok_s_mean": 323538.22090762435, "effective_prefill_tok_s_mean_ci95_high": 407647.65310614184, "effective_prefill_tok_s_mean_ci95_low": 239428.78870910686, "effective_prefill_tok_s_min": 242763.3198846212, "effective_prefill_tok_s_p25": 248626.660947824, "effective_prefill_tok_s_p50": 263376.7504350506, "effective_prefill_tok_s_p75": 272835.324100177, "effective_prefill_tok_s_p90": 527652.9423398477, "effective_prefill_tok_s_p95": 527652.9423398477, "effective_prefill_tok_s_p99": 527652.9423398477, "effective_prefill_tok_s_samples": 8, "effective_prefill_tok_s_stddev": 121376.22422310588, "generation_max_ms": 5454.0028820047155, "generation_mean_ci95_high_ms": 4470.809662593117, "generation_mean_ci95_low_ms": 2731.1677539597194, "generation_mean_ms": 3600.988708276418, "generation_min_ms": 2187.5238571083173, "generation_p25_ms": 2300.5695110186934, "generation_p50_ms": 3156.622144044377, "generation_p75_ms": 4248.2091169804335, "generation_p90_ms": 5454.0028820047155, "generation_p95_ms": 5454.0028820047155, "generation_p99_ms": 5454.0028820047155, "generation_samples": 8, "generation_stddev_ms": 1255.2169290112074, "mean_inter_token_latency_ms_max": 14.378559443866834, "mean_inter_token_latency_ms_mean": 13.58848320469619, "mean_inter_token_latency_ms_mean_ci95_high": 14.254113463533791, "mean_inter_token_latency_ms_mean_ci95_low": 12.92285294585859, "mean_inter_token_latency_ms_min": 11.4530044874781, "mean_inter_token_latency_ms_p25": 13.05857091693333, "mean_inter_token_latency_ms_p50": 14.027000801631045, "mean_inter_token_latency_ms_p75": 14.092063143055254, "mean_inter_token_latency_ms_p90": 14.378559443866834, "mean_inter_token_latency_ms_p95": 14.378559443866834, "mean_inter_token_latency_ms_p99": 14.378559443866834, "mean_inter_token_latency_ms_samples": 8, "mean_inter_token_latency_ms_stddev": 0.9605544281367837, "mean_time_per_output_token_max_ms": 14.378559443866834, "mean_time_per_output_token_mean_ci95_high_ms": 14.254113463533791, "mean_time_per_output_token_mean_ci95_low_ms": 12.92285294585859, "mean_time_per_output_token_mean_ms": 13.58848320469619, "mean_time_per_output_token_min_ms": 11.4530044874781, "mean_time_per_output_token_p25_ms": 13.05857091693333, "mean_time_per_output_token_p50_ms": 14.027000801631045, "mean_time_per_output_token_p75_ms": 14.092063143055254, "mean_time_per_output_token_p90_ms": 14.378559443866834, "mean_time_per_output_token_p95_ms": 14.378559443866834, "mean_time_per_output_token_p99_ms": 14.378559443866834, "mean_time_per_output_token_samples": 8, "mean_time_per_output_token_stddev_ms": 0.9605544281367837, "output_token_sources": { "usage": 8 }, "output_tokens": 2127, "prefix_cache_hit_avg": 1.0, "prompt_token_samples": 8, "prompt_tokens": 61447, "prompt_tokens_max": 7682.0, "prompt_tokens_mean": 7680.875, "prompt_tokens_mean_ci95_high": 7681.655272388336, "prompt_tokens_mean_ci95_low": 7680.094727611664, "prompt_tokens_min": 7679.0, "prompt_tokens_p25": 7680.0, "prompt_tokens_p50": 7681.0, "prompt_tokens_p75": 7682.0, "prompt_tokens_p90": 7682.0, "prompt_tokens_p95": 7682.0, "prompt_tokens_p99": 7682.0, "prompt_tokens_samples": 8, "prompt_tokens_stddev": 1.1259916264596033, "reasoning_chunks": 1626, "request_canary_passed": 8, "request_canary_samples": 8, "throughput_performance_eligible": true, "throughput_tok_s": 276.13009709914655, "time_to_first_output_max_ms": 31.631632009521127, "time_to_first_output_mean_ci95_high_ms": 30.91200450309939, "time_to_first_output_mean_ci95_low_ms": 21.16875894260206, "time_to_first_output_mean_ms": 26.040381722850725, "time_to_first_output_min_ms": 14.558812021277845, "time_to_first_output_p25_ms": 15.025288914330304, "time_to_first_output_p50_ms": 28.926171944476664, "time_to_first_output_p75_ms": 29.971604933962226, "time_to_first_output_p90_ms": 31.631632009521127, "time_to_first_output_p95_ms": 31.631632009521127, "time_to_first_output_p99_ms": 31.631632009521127, "time_to_first_output_samples": 8, "time_to_first_output_stddev_ms": 7.030117353666718, "tpot_max_ms": 14.378559443866834, "tpot_mean_ci95_high_ms": 14.254113463533791, "tpot_mean_ci95_low_ms": 12.92285294585859, "tpot_mean_ms": 13.58848320469619, "tpot_min_ms": 11.4530044874781, "tpot_p25_ms": 13.05857091693333, "tpot_p50_ms": 14.027000801631045, "tpot_p75_ms": 14.092063143055254, "tpot_p90_ms": 14.378559443866834, "tpot_p95_ms": 14.378559443866834, "tpot_p99_ms": 14.378559443866834, "tpot_samples": 8, "tpot_stddev_ms": 0.9605544281367837, "ttft_max_ms": 4627.0872809691355, "ttft_mean_ci95_high_ms": 3758.1727189572093, "ttft_mean_ci95_low_ms": 1953.2574272852826, "ttft_mean_ms": 2855.715073121246, "ttft_min_ms": 1394.9398359982297, "ttft_p25_ms": 1455.7648989139125, "ttft_p50_ms": 2368.828277103603, "ttft_p75_ms": 3541.1222679540515, "ttft_p90_ms": 4627.0872809691355, "ttft_p95_ms": 4627.0872809691355, "ttft_p99_ms": 4627.0872809691355, "ttft_samples": 8, "ttft_stddev_ms": 1302.3141247025662 } }, { "arm": "incumbent", "context_budget": 10240, "concurrency": 4, "cache_state": "warm across cells after C1", "artifact": "raw/incumbent-natural-ctx10240-c4.json", "requests": 8, "completed": 8, "performance_eligible": true, "metrics": { "completion_tokens_max": 890.0, "completion_tokens_mean": 349.75, "completion_tokens_mean_ci95_high": 535.0347109180894, "completion_tokens_mean_ci95_low": 164.46528908191053, "completion_tokens_min": 150.0, "completion_tokens_p25": 168.0, "completion_tokens_p50": 187.0, "completion_tokens_p75": 363.0, "completion_tokens_p90": 890.0, "completion_tokens_p95": 890.0, "completion_tokens_p99": 890.0, "completion_tokens_samples": 8, "completion_tokens_stddev": 267.3797460007556, "content_chunks": 179, "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": 128.95026564736233, "decode_tok_s_mean": 102.67066493807818, "decode_tok_s_mean_ci95_high": 112.1620806037178, "decode_tok_s_mean_ci95_low": 93.17924927243855, "decode_tok_s_min": 86.42699650539578, "decode_tok_s_p25": 93.65869993795428, "decode_tok_s_p50": 97.65725287105504, "decode_tok_s_p75": 105.46171011999456, "decode_tok_s_p90": 128.95026564736233, "decode_tok_s_p95": 128.95026564736233, "decode_tok_s_p99": 128.95026564736233, "decode_tok_s_samples": 8, "decode_tok_s_stddev": 13.696825265783692, "e2e_max_ms": 7744.674894027412, "e2e_mean_ci95_high_ms": 5010.794126294284, "e2e_mean_ci95_low_ms": 1874.6240382204194, "e2e_mean_ms": 3442.709082257352, "e2e_min_ms": 1552.1803179290146, "e2e_p25_ms": 1697.6592480205, "e2e_p50_ms": 2118.2844400173053, "e2e_p75_ms": 3809.936085017398, "e2e_p90_ms": 7744.674894027412, "e2e_p95_ms": 7744.674894027412, "e2e_p99_ms": 7744.674894027412, "e2e_samples": 8, "e2e_stddev_ms": 2262.8644247259613, "effective_prefill_tok_s_max": 118350.27486934699, "effective_prefill_tok_s_mean": 67548.41413856778, "effective_prefill_tok_s_mean_ci95_high": 84190.45227695463, "effective_prefill_tok_s_mean_ci95_low": 50906.37600018093, "effective_prefill_tok_s_min": 42637.993502953104, "effective_prefill_tok_s_p25": 44030.47276961466, "effective_prefill_tok_s_p50": 65715.79333295872, "effective_prefill_tok_s_p75": 71944.59801219677, "effective_prefill_tok_s_p90": 118350.27486934699, "effective_prefill_tok_s_p95": 118350.27486934699, "effective_prefill_tok_s_p99": 118350.27486934699, "effective_prefill_tok_s_samples": 8, "effective_prefill_tok_s_stddev": 24015.710245752023, "generation_max_ms": 7679.782776045613, "generation_mean_ci95_high_ms": 4901.461498694055, "generation_mean_ci95_low_ms": 1733.8819355368348, "generation_mean_ms": 3317.671717115445, "generation_min_ms": 1442.4165709642693, "generation_p25_ms": 1590.882641961798, "generation_p50_ms": 1943.8370739808306, "generation_p75_ms": 3706.8419329589233, "generation_p90_ms": 7679.782776045613, "generation_p95_ms": 7679.782776045613, "generation_p99_ms": 7679.782776045613, "generation_samples": 8, "generation_stddev_ms": 2285.527539853462, "mean_inter_token_latency_ms_max": 11.570458773695421, "mean_inter_token_latency_ms_mean": 9.880378003074437, "mean_inter_token_latency_ms_mean_ci95_high": 10.720587655091807, "mean_inter_token_latency_ms_mean_ci95_low": 9.040168351057067, "mean_inter_token_latency_ms_min": 7.754927800883168, "mean_inter_token_latency_ms_p25": 8.638675788577743, "mean_inter_token_latency_ms_p50": 10.217192505925002, "mean_inter_token_latency_ms_p75": 10.462695203615564, "mean_inter_token_latency_ms_p90": 11.570458773695421, "mean_inter_token_latency_ms_p95": 11.570458773695421, "mean_inter_token_latency_ms_p99": 11.570458773695421, "mean_inter_token_latency_ms_samples": 8, "mean_inter_token_latency_ms_stddev": 1.2124855970609627, "mean_time_per_output_token_max_ms": 11.570458773695421, "mean_time_per_output_token_mean_ci95_high_ms": 10.720587655091807, "mean_time_per_output_token_mean_ci95_low_ms": 9.040168351057067, "mean_time_per_output_token_mean_ms": 9.880378003074437, "mean_time_per_output_token_min_ms": 7.754927800883168, "mean_time_per_output_token_p25_ms": 8.638675788577743, "mean_time_per_output_token_p50_ms": 10.217192505925002, "mean_time_per_output_token_p75_ms": 10.462695203615564, "mean_time_per_output_token_p90_ms": 11.570458773695421, "mean_time_per_output_token_p95_ms": 11.570458773695421, "mean_time_per_output_token_p99_ms": 11.570458773695421, "mean_time_per_output_token_samples": 8, "mean_time_per_output_token_stddev_ms": 1.2124855970609627, "output_token_sources": { "usage": 8 }, "output_tokens": 2798, "prefix_cache_hit_avg": 1.0, "prompt_token_samples": 8, "prompt_tokens": 61447, "prompt_tokens_max": 7682.0, "prompt_tokens_mean": 7680.875, "prompt_tokens_mean_ci95_high": 7681.655272388336, "prompt_tokens_mean_ci95_low": 7680.094727611664, "prompt_tokens_min": 7679.0, "prompt_tokens_p25": 7680.0, "prompt_tokens_p50": 7681.0, "prompt_tokens_p75": 7682.0, "prompt_tokens_p90": 7682.0, "prompt_tokens_p95": 7682.0, "prompt_tokens_p99": 7682.0, "prompt_tokens_samples": 8, "prompt_tokens_stddev": 1.1259916264596033, "reasoning_chunks": 741, "request_canary_passed": 8, "request_canary_samples": 8, "throughput_performance_eligible": true, "throughput_tok_s": 360.2057017593626, "time_to_first_output_max_ms": 180.09759299457073, "time_to_first_output_mean_ci95_high_ms": 151.97630227665312, "time_to_first_output_mean_ci95_low_ms": 98.09842800716045, "time_to_first_output_mean_ms": 125.03736514190678, "time_to_first_output_min_ms": 64.89211798179895, "time_to_first_output_p25_ms": 103.09415205847472, "time_to_first_output_p50_ms": 109.76374696474522, "time_to_first_output_p75_ms": 144.3604549858719, "time_to_first_output_p90_ms": 180.09759299457073, "time_to_first_output_p95_ms": 180.09759299457073, "time_to_first_output_p99_ms": 180.09759299457073, "time_to_first_output_samples": 8, "time_to_first_output_stddev_ms": 38.874908420280065, "tpot_max_ms": 11.570458773695421, "tpot_mean_ci95_high_ms": 10.720587655091807, "tpot_mean_ci95_low_ms": 9.040168351057067, "tpot_mean_ms": 9.880378003074437, "tpot_min_ms": 7.754927800883168, "tpot_p25_ms": 8.638675788577743, "tpot_p50_ms": 10.217192505925002, "tpot_p75_ms": 10.462695203615564, "tpot_p90_ms": 11.570458773695421, "tpot_p95_ms": 11.570458773695421, "tpot_p99_ms": 11.570458773695421, "tpot_samples": 8, "tpot_stddev_ms": 1.2124855970609627, "ttft_max_ms": 7441.858428996056, "ttft_mean_ci95_high_ms": 4484.017374939396, "ttft_mean_ci95_low_ms": 1198.4772130822585, "ttft_mean_ms": 2841.2472940108273, "ttft_min_ms": 1063.058627070859, "ttft_p25_ms": 1084.5879840198904, "ttft_p50_ms": 1383.3696600049734, "ttft_p75_ms": 3303.18836402148, "ttft_p90_ms": 7441.858428996056, "ttft_p95_ms": 7441.858428996056, "ttft_p99_ms": 7441.858428996056, "ttft_samples": 8, "ttft_stddev_ms": 2370.6405390917644 } }, { "arm": "flare", "context_budget": 32768, "concurrency": 1, "cache_state": "near-zero reported hit ratio", "artifact": "raw/flare-natural-ctx32768-c1.json", "requests": 8, "completed": 8, "performance_eligible": true, "metrics": { "completion_tokens_max": 352.0, "completion_tokens_mean": 224.75, "completion_tokens_mean_ci95_high": 281.5377777342977, "completion_tokens_mean_ci95_low": 167.96222226570228, "completion_tokens_min": 134.0, "completion_tokens_p25": 144.0, "completion_tokens_p50": 177.0, "completion_tokens_p75": 278.0, "completion_tokens_p90": 352.0, "completion_tokens_p95": 352.0, "completion_tokens_p99": 352.0, "completion_tokens_samples": 8, "completion_tokens_stddev": 81.94902596823746, "content_chunks": 462, "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": 199.529454980295, "decode_tok_s_mean": 189.04768860391374, "decode_tok_s_mean_ci95_high": 196.56480438638116, "decode_tok_s_mean_ci95_low": 181.53057282144633, "decode_tok_s_min": 172.83327391917567, "decode_tok_s_p25": 174.99983267413782, "decode_tok_s_p50": 192.39176299909897, "decode_tok_s_p75": 196.80954334738811, "decode_tok_s_p90": 199.529454980295, "decode_tok_s_p95": 199.529454980295, "decode_tok_s_p99": 199.529454980295, "decode_tok_s_samples": 8, "decode_tok_s_stddev": 10.847762336218649, "e2e_max_ms": 4076.1594999348745, "e2e_mean_ci95_high_ms": 3772.904666564058, "e2e_mean_ci95_low_ms": 3285.2610764516485, "e2e_mean_ms": 3529.0828715078533, "e2e_min_ms": 3050.2038310514763, "e2e_p25_ms": 3214.1551630338654, "e2e_p50_ms": 3406.731904949993, "e2e_p75_ms": 3746.698296046816, "e2e_p90_ms": 4076.1594999348745, "e2e_p95_ms": 4076.1594999348745, "e2e_p99_ms": 4076.1594999348745, "e2e_samples": 8, "e2e_stddev_ms": 351.85315241901816, "effective_prefill_tok_s_max": 9951.259229928351, "effective_prefill_tok_s_mean": 9722.091830235728, "effective_prefill_tok_s_mean_ci95_high": 9836.138388834152, "effective_prefill_tok_s_mean_ci95_low": 9608.045271637304, "effective_prefill_tok_s_min": 9441.423647534995, "effective_prefill_tok_s_p25": 9597.181835510393, "effective_prefill_tok_s_p50": 9716.792263891199, "effective_prefill_tok_s_p75": 9788.352060848687, "effective_prefill_tok_s_p90": 9951.259229928351, "effective_prefill_tok_s_p95": 9951.259229928351, "effective_prefill_tok_s_p99": 9951.259229928351, "effective_prefill_tok_s_samples": 8, "effective_prefill_tok_s_stddev": 164.57774480803073, "generation_max_ms": 1759.1387699358165, "generation_mean_ci95_high_ms": 1433.8179602306795, "generation_mean_ci95_low_ms": 907.792595219054, "generation_mean_ms": 1170.8052777248668, "generation_min_ms": 691.2977870088071, "generation_p25_ms": 786.55516798608, "generation_p50_ms": 1018.3224329957739, "generation_p75_ms": 1397.2913989564404, "generation_p90_ms": 1759.1387699358165, "generation_p95_ms": 1759.1387699358165, "generation_p99_ms": 1759.1387699358165, "generation_samples": 8, "generation_stddev_ms": 379.5470435467851, "mean_inter_token_latency_ms_max": 5.785922914748715, "mean_inter_token_latency_ms_mean": 5.3054427130290005, "mean_inter_token_latency_ms_mean_ci95_high": 5.523218859690406, "mean_inter_token_latency_ms_mean_ci95_low": 5.087666566367595, "mean_inter_token_latency_ms_min": 5.011791367338509, "mean_inter_token_latency_ms_p25": 5.044373281431192, "mean_inter_token_latency_ms_p50": 5.107995031364992, "mean_inter_token_latency_ms_p75": 5.500385790112448, "mean_inter_token_latency_ms_p90": 5.785922914748715, "mean_inter_token_latency_ms_p95": 5.785922914748715, "mean_inter_token_latency_ms_p99": 5.785922914748715, "mean_inter_token_latency_ms_samples": 8, "mean_inter_token_latency_ms_stddev": 0.3142673267039916, "mean_time_per_output_token_max_ms": 5.785922914748715, "mean_time_per_output_token_mean_ci95_high_ms": 5.523218859690406, "mean_time_per_output_token_mean_ci95_low_ms": 5.087666566367595, "mean_time_per_output_token_mean_ms": 5.3054427130290005, "mean_time_per_output_token_min_ms": 5.011791367338509, "mean_time_per_output_token_p25_ms": 5.044373281431192, "mean_time_per_output_token_p50_ms": 5.107995031364992, "mean_time_per_output_token_p75_ms": 5.500385790112448, "mean_time_per_output_token_p90_ms": 5.785922914748715, "mean_time_per_output_token_p95_ms": 5.785922914748715, "mean_time_per_output_token_p99_ms": 5.785922914748715, "mean_time_per_output_token_samples": 8, "mean_time_per_output_token_stddev_ms": 0.3142673267039916, "output_token_sources": { "usage": 8 }, "output_tokens": 1798, "prefix_cache_hit_avg": 0.00039264231994294026, "prompt_token_samples": 8, "prompt_tokens": 183373, "prompt_tokens_max": 22923.0, "prompt_tokens_mean": 22921.625, "prompt_tokens_mean_ci95_high": 22922.36, "prompt_tokens_mean_ci95_low": 22920.89, "prompt_tokens_min": 22920.0, "prompt_tokens_p25": 22921.0, "prompt_tokens_p50": 22921.0, "prompt_tokens_p75": 22922.0, "prompt_tokens_p90": 22923.0, "prompt_tokens_p95": 22923.0, "prompt_tokens_p99": 22923.0, "prompt_tokens_samples": 8, "prompt_tokens_stddev": 1.0606601717798212, "reasoning_chunks": 1320, "request_canary_passed": 8, "request_canary_samples": 8, "throughput_performance_eligible": true, "throughput_tok_s": 63.605123578605166, "time_to_first_output_max_ms": 2427.5999950477853, "time_to_first_output_mean_ci95_high_ms": 2386.0292139124494, "time_to_first_output_mean_ci95_low_ms": 2330.5259736535236, "time_to_first_output_mean_ms": 2358.2775937829865, "time_to_first_output_min_ms": 2303.4270809730515, "time_to_first_output_p25_ms": 2317.020729999058, "time_to_first_output_p50_ms": 2349.4068970903754, "time_to_first_output_p75_ms": 2379.5854391064495, "time_to_first_output_p90_ms": 2427.5999950477853, "time_to_first_output_p95_ms": 2427.5999950477853, "time_to_first_output_p99_ms": 2427.5999950477853, "time_to_first_output_samples": 8, "time_to_first_output_stddev_ms": 40.04767098460469, "tpot_max_ms": 5.785922914748715, "tpot_mean_ci95_high_ms": 5.523218859690406, "tpot_mean_ci95_low_ms": 5.087666566367595, "tpot_mean_ms": 5.3054427130290005, "tpot_min_ms": 5.011791367338509, "tpot_p25_ms": 5.044373281431192, "tpot_p50_ms": 5.107995031364992, "tpot_p75_ms": 5.500385790112448, "tpot_p90_ms": 5.785922914748715, "tpot_p95_ms": 5.785922914748715, "tpot_p99_ms": 5.785922914748715, "tpot_samples": 8, "tpot_stddev_ms": 0.3142673267039916, "ttft_max_ms": 3758.2150369416922, "ttft_mean_ci95_high_ms": 3480.721853078112, "ttft_mean_ci95_low_ms": 2987.9638504777085, "ttft_mean_ms": 3234.3428517779103, "ttft_min_ms": 2808.9349310612306, "ttft_p25_ms": 2883.779038093053, "ttft_p50_ms": 3052.4571960559115, "ttft_p75_ms": 3493.3041071053594, "ttft_p90_ms": 3758.2150369416922, "ttft_p95_ms": 3758.2150369416922, "ttft_p99_ms": 3758.2150369416922, "ttft_samples": 8, "ttft_stddev_ms": 355.5433929823302 } }, { "arm": "incumbent", "context_budget": 32768, "concurrency": 1, "cache_state": "near-zero reported hit ratio", "artifact": "raw/incumbent-natural-ctx32768-c1.json", "requests": 8, "completed": 8, "performance_eligible": true, "metrics": { "completion_tokens_max": 451.0, "completion_tokens_mean": 285.25, "completion_tokens_mean_ci95_high": 361.8887480325716, "completion_tokens_mean_ci95_low": 208.6112519674284, "completion_tokens_min": 141.0, "completion_tokens_p25": 155.0, "completion_tokens_p50": 271.0, "completion_tokens_p75": 311.0, "completion_tokens_p90": 451.0, "completion_tokens_p95": 451.0, "completion_tokens_p99": 451.0, "completion_tokens_samples": 8, "completion_tokens_stddev": 110.59546619485423, "content_chunks": 157, "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": 208.98728017478112, "decode_tok_s_mean": 199.88318486802692, "decode_tok_s_mean_ci95_high": 204.46343676610505, "decode_tok_s_mean_ci95_low": 195.30293296994878, "decode_tok_s_min": 191.3738713502919, "decode_tok_s_p25": 192.00432047075415, "decode_tok_s_p50": 197.80995221265655, "decode_tok_s_p75": 204.6028775550115, "decode_tok_s_p90": 208.98728017478112, "decode_tok_s_p95": 208.98728017478112, "decode_tok_s_p99": 208.98728017478112, "decode_tok_s_samples": 8, "decode_tok_s_stddev": 6.6096472993338775, "e2e_max_ms": 4890.272064018063, "e2e_mean_ci95_high_ms": 4374.101438704288, "e2e_mean_ci95_low_ms": 3556.9536780248245, "e2e_mean_ms": 3965.527558364556, "e2e_min_ms": 3221.342335920781, "e2e_p25_ms": 3268.378110951744, "e2e_p50_ms": 3828.3064720453694, "e2e_p75_ms": 4086.1453220713884, "e2e_p90_ms": 4890.272064018063, "e2e_p95_ms": 4890.272064018063, "e2e_p99_ms": 4890.272064018063, "e2e_samples": 8, "e2e_stddev_ms": 589.6027783753577, "effective_prefill_tok_s_max": 9252.554229201067, "effective_prefill_tok_s_mean": 9046.670291941677, "effective_prefill_tok_s_mean_ci95_high": 9113.523040875636, "effective_prefill_tok_s_mean_ci95_low": 8979.817543007717, "effective_prefill_tok_s_min": 8951.319478228159, "effective_prefill_tok_s_p25": 8957.935404239368, "effective_prefill_tok_s_p50": 9021.867565924147, "effective_prefill_tok_s_p75": 9050.186139151434, "effective_prefill_tok_s_p90": 9252.554229201067, "effective_prefill_tok_s_p95": 9252.554229201067, "effective_prefill_tok_s_p99": 9252.554229201067, "effective_prefill_tok_s_samples": 8, "effective_prefill_tok_s_stddev": 96.47353492278373, "generation_max_ms": 2343.6972610652447, "generation_mean_ci95_high_ms": 1837.10243579654, "generation_mean_ci95_low_ms": 1026.0388511761373, "generation_mean_ms": 1431.5706434863387, "generation_min_ms": 707.7500319574028, "generation_p25_ms": 743.864286923781, "generation_p50_ms": 1291.9446569867432, "generation_p75_ms": 1545.4296850366518, "generation_p90_ms": 2343.6972610652447, "generation_p95_ms": 2343.6972610652447, "generation_p99_ms": 2343.6972610652447, "generation_samples": 8, "generation_stddev_ms": 585.2128169985267, "mean_inter_token_latency_ms_max": 5.2253737301974414, "mean_inter_token_latency_ms_mean": 5.0077102444289245, "mean_inter_token_latency_ms_mean_ci95_high": 5.122420697531688, "mean_inter_token_latency_ms_mean_ci95_low": 4.892999791326161, "mean_inter_token_latency_ms_min": 4.784980211062012, "mean_inter_token_latency_ms_p25": 4.83028757742715, "mean_inter_token_latency_ms_p50": 4.9852570485053285, "mean_inter_token_latency_ms_p75": 5.0846930881518695, "mean_inter_token_latency_ms_p90": 5.2253737301974414, "mean_inter_token_latency_ms_p95": 5.2253737301974414, "mean_inter_token_latency_ms_p99": 5.2253737301974414, "mean_inter_token_latency_ms_samples": 8, "mean_inter_token_latency_ms_stddev": 0.16553579441213362, "mean_time_per_output_token_max_ms": 5.2253737301974414, "mean_time_per_output_token_mean_ci95_high_ms": 5.122420697531688, "mean_time_per_output_token_mean_ci95_low_ms": 4.892999791326161, "mean_time_per_output_token_mean_ms": 5.0077102444289245, "mean_time_per_output_token_min_ms": 4.784980211062012, "mean_time_per_output_token_p25_ms": 4.83028757742715, "mean_time_per_output_token_p50_ms": 4.9852570485053285, "mean_time_per_output_token_p75_ms": 5.0846930881518695, "mean_time_per_output_token_p90_ms": 5.2253737301974414, "mean_time_per_output_token_p95_ms": 5.2253737301974414, "mean_time_per_output_token_p99_ms": 5.2253737301974414, "mean_time_per_output_token_samples": 8, "mean_time_per_output_token_stddev_ms": 0.16553579441213362, "output_token_sources": { "usage": 8 }, "output_tokens": 2282, "prefix_cache_hit_avg": 0.0, "prompt_token_samples": 8, "prompt_tokens": 183373, "prompt_tokens_max": 22923.0, "prompt_tokens_mean": 22921.625, "prompt_tokens_mean_ci95_high": 22922.36, "prompt_tokens_mean_ci95_low": 22920.89, "prompt_tokens_min": 22920.0, "prompt_tokens_p25": 22921.0, "prompt_tokens_p50": 22921.0, "prompt_tokens_p75": 22922.0, "prompt_tokens_p90": 22923.0, "prompt_tokens_p95": 22923.0, "prompt_tokens_p99": 22923.0, "prompt_tokens_samples": 8, "prompt_tokens_stddev": 1.0606601717798212, "reasoning_chunks": 609, "request_canary_passed": 8, "request_canary_samples": 8, "throughput_performance_eligible": true, "throughput_tok_s": 71.87331390922793, "time_to_first_output_max_ms": 2560.628078994341, "time_to_first_output_mean_ci95_high_ms": 2552.4098976914706, "time_to_first_output_mean_ci95_low_ms": 2515.5039320649644, "time_to_first_output_mean_ms": 2533.9569148782175, "time_to_first_output_min_ms": 2477.478048997, "time_to_first_output_p25_ms": 2518.3955730171874, "time_to_first_output_p50_ms": 2536.3618150586262, "time_to_first_output_p75_ms": 2546.5748029528186, "time_to_first_output_p90_ms": 2560.628078994341, "time_to_first_output_p95_ms": 2560.628078994341, "time_to_first_output_p99_ms": 2560.628078994341, "time_to_first_output_samples": 8, "time_to_first_output_stddev_ms": 26.629039347694125, "tpot_max_ms": 5.2253737301974414, "tpot_mean_ci95_high_ms": 5.122420697531688, "tpot_mean_ci95_low_ms": 4.892999791326161, "tpot_mean_ms": 5.0077102444289245, "tpot_min_ms": 4.784980211062012, "tpot_p25_ms": 4.83028757742715, "tpot_p50_ms": 4.9852570485053285, "tpot_p75_ms": 5.0846930881518695, "tpot_p90_ms": 5.2253737301974414, "tpot_p95_ms": 5.2253737301974414, "tpot_p99_ms": 5.2253737301974414, "tpot_samples": 8, "tpot_stddev_ms": 0.16553579441213362, "ttft_max_ms": 4572.150354972109, "ttft_mean_ci95_high_ms": 4072.4067943403106, "ttft_mean_ci95_low_ms": 3255.296162137568, "ttft_mean_ms": 3663.8514782389393, "ttft_min_ms": 2934.1588499955833, "ttft_p25_ms": 2952.3767579812557, "ttft_p50_ms": 3509.0971150202677, "ttft_p75_ms": 3814.9695590836927, "ttft_p90_ms": 4572.150354972109, "ttft_p95_ms": 4572.150354972109, "ttft_p99_ms": 4572.150354972109, "ttft_samples": 8, "ttft_stddev_ms": 589.5759887859047 } }, { "arm": "flare", "context_budget": 32768, "concurrency": 4, "cache_state": "warm across cells after C1", "artifact": "raw/flare-natural-ctx32768-c4.json", "requests": 8, "completed": 8, "performance_eligible": true, "metrics": { "completion_tokens_max": 457.0, "completion_tokens_mean": 265.625, "completion_tokens_mean_ci95_high": 347.26547663383644, "completion_tokens_mean_ci95_low": 183.98452336616353, "completion_tokens_min": 110.0, "completion_tokens_p25": 157.0, "completion_tokens_p50": 260.0, "completion_tokens_p75": 346.0, "completion_tokens_p90": 457.0, "completion_tokens_p95": 457.0, "completion_tokens_p99": 457.0, "completion_tokens_samples": 8, "completion_tokens_stddev": 117.81333601446461, "content_chunks": 493, "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": 94.75523288618872, "decode_tok_s_mean": 74.68984303008176, "decode_tok_s_mean_ci95_high": 81.3534410631126, "decode_tok_s_mean_ci95_low": 68.02624499705092, "decode_tok_s_min": 68.43877027283686, "decode_tok_s_p25": 68.83328192224222, "decode_tok_s_p50": 69.57331136787838, "decode_tok_s_p75": 71.19990140622666, "decode_tok_s_p90": 94.75523288618872, "decode_tok_s_p95": 94.75523288618872, "decode_tok_s_p99": 94.75523288618872, "decode_tok_s_samples": 8, "decode_tok_s_stddev": 9.61607215562744, "e2e_max_ms": 5454.599478980526, "e2e_mean_ci95_high_ms": 4642.858450165285, "e2e_mean_ci95_low_ms": 2530.8482088341507, "e2e_mean_ms": 3586.853329499718, "e2e_min_ms": 1581.3772779656574, "e2e_p25_ms": 2248.110329033807, "e2e_p50_ms": 2777.716259006411, "e2e_p75_ms": 5049.771215999499, "e2e_p90_ms": 5454.599478980526, "e2e_p95_ms": 5454.599478980526, "e2e_p99_ms": 5454.599478980526, "e2e_samples": 8, "e2e_stddev_ms": 1523.894656714981, "effective_prefill_tok_s_max": 684587.7677774223, "effective_prefill_tok_s_mean": 547736.6778909026, "effective_prefill_tok_s_mean_ci95_high": 605512.6314789429, "effective_prefill_tok_s_mean_ci95_low": 489960.7243028623, "effective_prefill_tok_s_min": 401428.7565089348, "effective_prefill_tok_s_p25": 502197.5577641423, "effective_prefill_tok_s_p50": 524383.5134052385, "effective_prefill_tok_s_p75": 577511.9808166852, "effective_prefill_tok_s_p90": 684587.7677774223, "effective_prefill_tok_s_p95": 684587.7677774223, "effective_prefill_tok_s_p99": 684587.7677774223, "effective_prefill_tok_s_samples": 8, "effective_prefill_tok_s_stddev": 83375.03790127051, "generation_max_ms": 5410.889104008675, "generation_mean_ci95_high_ms": 4601.763092098111, "generation_mean_ci95_low_ms": 2486.415259627108, "generation_mean_ms": 3544.0891758626094, "generation_min_ms": 1535.7338859466836, "generation_p25_ms": 2191.014269948937, "generation_p50_ms": 2733.3582759601995, "generation_p75_ms": 5012.110281037167, "generation_p90_ms": 5410.889104008675, "generation_p95_ms": 5410.889104008675, "generation_p99_ms": 5410.889104008675, "generation_samples": 8, "generation_stddev_ms": 1526.302853988277, "mean_inter_token_latency_ms_max": 14.61160093925441, "mean_inter_token_latency_ms_mean": 13.557709802147496, "mean_inter_token_latency_ms_mean_ci95_high": 14.599647191975583, "mean_inter_token_latency_ms_mean_ci95_low": 12.515772412319409, "mean_inter_token_latency_ms_min": 10.553506856989188, "mean_inter_token_latency_ms_p25": 11.865984877212005, "mean_inter_token_latency_ms_p50": 14.089301705932877, "mean_inter_token_latency_ms_p75": 14.395137231622357, "mean_inter_token_latency_ms_p90": 14.61160093925441, "mean_inter_token_latency_ms_p95": 14.61160093925441, "mean_inter_token_latency_ms_p99": 14.61160093925441, "mean_inter_token_latency_ms_samples": 8, "mean_inter_token_latency_ms_stddev": 1.503593865141331, "mean_time_per_output_token_max_ms": 14.61160093925441, "mean_time_per_output_token_mean_ci95_high_ms": 14.599647191975583, "mean_time_per_output_token_mean_ci95_low_ms": 12.515772412319409, "mean_time_per_output_token_mean_ms": 13.557709802147496, "mean_time_per_output_token_min_ms": 10.553506856989188, "mean_time_per_output_token_p25_ms": 11.865984877212005, "mean_time_per_output_token_p50_ms": 14.089301705932877, "mean_time_per_output_token_p75_ms": 14.395137231622357, "mean_time_per_output_token_p90_ms": 14.61160093925441, "mean_time_per_output_token_p95_ms": 14.61160093925441, "mean_time_per_output_token_p99_ms": 14.61160093925441, "mean_time_per_output_token_samples": 8, "mean_time_per_output_token_stddev_ms": 1.503593865141331, "output_token_sources": { "usage": 8 }, "output_tokens": 2125, "prefix_cache_hit_avg": 1.0, "prompt_token_samples": 8, "prompt_tokens": 183373, "prompt_tokens_max": 22923.0, "prompt_tokens_mean": 22921.625, "prompt_tokens_mean_ci95_high": 22922.36, "prompt_tokens_mean_ci95_low": 22920.89, "prompt_tokens_min": 22920.0, "prompt_tokens_p25": 22921.0, "prompt_tokens_p50": 22921.0, "prompt_tokens_p75": 22922.0, "prompt_tokens_p90": 22923.0, "prompt_tokens_p95": 22923.0, "prompt_tokens_p99": 22923.0, "prompt_tokens_samples": 8, "prompt_tokens_stddev": 1.0606601717798212, "reasoning_chunks": 1616, "request_canary_passed": 8, "request_canary_samples": 8, "throughput_performance_eligible": true, "throughput_tok_s": 270.47051528774455, "time_to_first_output_max_ms": 57.09605908486992, "time_to_first_output_mean_ci95_high_ms": 47.62222525956946, "time_to_first_output_mean_ci95_low_ms": 37.906082014647325, "time_to_first_output_mean_ms": 42.76415363710839, "time_to_first_output_min_ms": 33.48146297503263, "time_to_first_output_p25_ms": 37.66093496233225, "time_to_first_output_p50_ms": 40.47033900860697, "time_to_first_output_p75_ms": 44.3579830462113, "time_to_first_output_p90_ms": 57.09605908486992, "time_to_first_output_p95_ms": 57.09605908486992, "time_to_first_output_p99_ms": 57.09605908486992, "time_to_first_output_samples": 8, "time_to_first_output_stddev_ms": 7.010562015779907, "tpot_max_ms": 14.61160093925441, "tpot_mean_ci95_high_ms": 14.599647191975583, "tpot_mean_ci95_low_ms": 12.515772412319409, "tpot_mean_ms": 13.557709802147496, "tpot_min_ms": 10.553506856989188, "tpot_p25_ms": 11.865984877212005, "tpot_p50_ms": 14.089301705932877, "tpot_p75_ms": 14.395137231622357, "tpot_p90_ms": 14.61160093925441, "tpot_p95_ms": 14.61160093925441, "tpot_p99_ms": 14.61160093925441, "tpot_samples": 8, "tpot_stddev_ms": 1.503593865141331, "ttft_max_ms": 5056.205847999081, "ttft_mean_ci95_high_ms": 3915.7741655775153, "ttft_mean_ci95_low_ms": 1679.7216811561684, "ttft_mean_ms": 2797.747923366842, "ttft_min_ms": 573.4389149583876, "ttft_p25_ms": 1300.9663440752774, "ttft_p50_ms": 2301.5108529943973, "ttft_p75_ms": 4121.775856940076, "ttft_p90_ms": 5056.205847999081, "ttft_p95_ms": 5056.205847999081, "ttft_p99_ms": 5056.205847999081, "ttft_samples": 8, "ttft_stddev_ms": 1613.3957906360831 } }, { "arm": "incumbent", "context_budget": 32768, "concurrency": 4, "cache_state": "warm across cells after C1", "artifact": "raw/incumbent-natural-ctx32768-c4.json", "requests": 8, "completed": 8, "performance_eligible": true, "metrics": { "completion_tokens_max": 605.0, "completion_tokens_mean": 277.125, "completion_tokens_mean_ci95_high": 389.95484368065036, "completion_tokens_mean_ci95_low": 164.29515631934964, "completion_tokens_min": 128.0, "completion_tokens_p25": 134.0, "completion_tokens_p50": 233.0, "completion_tokens_p75": 322.0, "completion_tokens_p90": 605.0, "completion_tokens_p95": 605.0, "completion_tokens_p99": 605.0, "completion_tokens_samples": 8, "completion_tokens_stddev": 162.8219338506245, "content_chunks": 152, "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": 136.77290117732184, "decode_tok_s_mean": 105.80067730594726, "decode_tok_s_mean_ci95_high": 116.01509784527006, "decode_tok_s_mean_ci95_low": 95.58625676662446, "decode_tok_s_min": 86.34139604441637, "decode_tok_s_p25": 93.64217796384582, "decode_tok_s_p50": 105.11871870733648, "decode_tok_s_p75": 107.9028148098366, "decode_tok_s_p90": 136.77290117732184, "decode_tok_s_p95": 136.77290117732184, "decode_tok_s_p99": 136.77290117732184, "decode_tok_s_samples": 8, "decode_tok_s_stddev": 14.740175569890415, "e2e_max_ms": 5783.462824067101, "e2e_mean_ci95_high_ms": 3679.789739112315, "e2e_mean_ci95_low_ms": 1696.3728636575945, "e2e_mean_ms": 2688.0813013849547, "e2e_min_ms": 1341.1104390397668, "e2e_p25_ms": 1364.0983930090442, "e2e_p50_ms": 2302.195120952092, "e2e_p75_ms": 3024.4138630805537, "e2e_p90_ms": 5783.462824067101, "e2e_p95_ms": 5783.462824067101, "e2e_p99_ms": 5783.462824067101, "e2e_samples": 8, "e2e_stddev_ms": 1431.109716891701, "effective_prefill_tok_s_max": 339120.0369978834, "effective_prefill_tok_s_mean": 209147.4818901875, "effective_prefill_tok_s_mean_ci95_high": 250139.39152959868, "effective_prefill_tok_s_mean_ci95_low": 168155.57225077634, "effective_prefill_tok_s_min": 141506.55941750386, "effective_prefill_tok_s_p25": 159491.40970984523, "effective_prefill_tok_s_p50": 210110.95428258958, "effective_prefill_tok_s_p75": 211665.33945100434, "effective_prefill_tok_s_p90": 339120.0369978834, "effective_prefill_tok_s_p95": 339120.0369978834, "effective_prefill_tok_s_p99": 339120.0369978834, "effective_prefill_tok_s_samples": 8, "effective_prefill_tok_s_stddev": 59154.40261186499, "generation_max_ms": 5639.737215009518, "generation_mean_ci95_high_ms": 3555.285259069128, "generation_mean_ci95_low_ms": 1588.5084399132074, "generation_mean_ms": 2571.896849491168, "generation_min_ms": 1232.5906440382823, "generation_p25_ms": 1255.8095250278711, "generation_p50_ms": 2140.2167639462277, "generation_p75_ms": 2917.244545998983, "generation_p90_ms": 5639.737215009518, "generation_p95_ms": 5639.737215009518, "generation_p99_ms": 5639.737215009518, "generation_samples": 8, "generation_stddev_ms": 1419.1032917404693, "mean_inter_token_latency_ms_max": 11.581929941062949, "mean_inter_token_latency_ms_mean": 9.600446275626046, "mean_inter_token_latency_ms_mean_ci95_high": 10.458821133309595, "mean_inter_token_latency_ms_mean_ci95_low": 8.742071417942498, "mean_inter_token_latency_ms_min": 7.3113898395964485, "mean_inter_token_latency_ms_p25": 9.225072258388913, "mean_inter_token_latency_ms_p50": 9.337313269883309, "mean_inter_token_latency_ms_p75": 9.888263976597411, "mean_inter_token_latency_ms_p90": 11.581929941062949, "mean_inter_token_latency_ms_p95": 11.581929941062949, "mean_inter_token_latency_ms_p99": 11.581929941062949, "mean_inter_token_latency_ms_samples": 8, "mean_inter_token_latency_ms_stddev": 1.2386993523838277, "mean_time_per_output_token_max_ms": 11.581929941062949, "mean_time_per_output_token_mean_ci95_high_ms": 10.458821133309595, "mean_time_per_output_token_mean_ci95_low_ms": 8.742071417942498, "mean_time_per_output_token_mean_ms": 9.600446275626046, "mean_time_per_output_token_min_ms": 7.3113898395964485, "mean_time_per_output_token_p25_ms": 9.225072258388913, "mean_time_per_output_token_p50_ms": 9.337313269883309, "mean_time_per_output_token_p75_ms": 9.888263976597411, "mean_time_per_output_token_p90_ms": 11.581929941062949, "mean_time_per_output_token_p95_ms": 11.581929941062949, "mean_time_per_output_token_p99_ms": 11.581929941062949, "mean_time_per_output_token_samples": 8, "mean_time_per_output_token_stddev_ms": 1.2386993523838277, "output_token_sources": { "usage": 8 }, "output_tokens": 2217, "prefix_cache_hit_avg": 1.0, "prompt_token_samples": 8, "prompt_tokens": 183373, "prompt_tokens_max": 22923.0, "prompt_tokens_mean": 22921.625, "prompt_tokens_mean_ci95_high": 22922.36, "prompt_tokens_mean_ci95_low": 22920.89, "prompt_tokens_min": 22920.0, "prompt_tokens_p25": 22921.0, "prompt_tokens_p50": 22921.0, "prompt_tokens_p75": 22922.0, "prompt_tokens_p90": 22923.0, "prompt_tokens_p95": 22923.0, "prompt_tokens_p99": 22923.0, "prompt_tokens_samples": 8, "prompt_tokens_stddev": 1.0606601717798212, "reasoning_chunks": 609, "request_canary_passed": 8, "request_canary_samples": 8, "throughput_performance_eligible": true, "throughput_tok_s": 331.3804800165646, "time_to_first_output_max_ms": 161.97835700586438, "time_to_first_output_mean_ci95_high_ms": 135.63885194671573, "time_to_first_output_mean_ci95_low_ms": 96.73005184085807, "time_to_first_output_mean_ms": 116.1844518937869, "time_to_first_output_min_ms": 67.59553402662277, "time_to_first_output_p25_ms": 107.16931708157063, "time_to_first_output_p50_ms": 108.51979500148445, "time_to_first_output_p75_ms": 123.10815404634923, "time_to_first_output_p90_ms": 161.97835700586438, "time_to_first_output_p95_ms": 161.97835700586438, "time_to_first_output_p99_ms": 161.97835700586438, "time_to_first_output_samples": 8, "time_to_first_output_stddev_ms": 28.074159594575324, "tpot_max_ms": 11.581929941062949, "tpot_mean_ci95_high_ms": 10.458821133309595, "tpot_mean_ci95_low_ms": 8.742071417942498, "tpot_mean_ms": 9.600446275626046, "tpot_min_ms": 7.3113898395964485, "tpot_p25_ms": 9.225072258388913, "tpot_p50_ms": 9.337313269883309, "tpot_p75_ms": 9.888263976597411, "tpot_p90_ms": 11.581929941062949, "tpot_p95_ms": 11.581929941062949, "tpot_p99_ms": 11.581929941062949, "tpot_samples": 8, "tpot_stddev_ms": 1.2386993523838277, "ttft_max_ms": 5344.398624030873, "ttft_mean_ci95_high_ms": 3203.7807130965693, "ttft_mean_ci95_low_ms": 1140.6835111427763, "ttft_mean_ms": 2172.232112119673, "ttft_min_ms": 821.7533009592444, "ttft_p25_ms": 844.6574690751731, "ttft_p50_ms": 1634.988614008762, "ttft_p75_ms": 2737.8025780199096, "ttft_p90_ms": 5344.398624030873, "ttft_p95_ms": 5344.398624030873, "ttft_p99_ms": 5344.398624030873, "ttft_samples": 8, "ttft_stddev_ms": 1488.602063008693 } } ], "failures": [ "Strict-output cells are performance-ineligible: extra code words, no canary errors.", "r2 io_uring EPERM under default seccomp; r3 intentionally stopped for bounded worker increase.", "Preparation reached 52 GiB cgroup cap; max events7463, no observed OOM or OOM kill." ], "limitations": [ "Different checkpoints and runtime recipes: no isolated engine effect.", "Natural output lengths vary; not controlled decode.", "n8/cell; no stable p99 or endurance claim.", "No cold-C4, MTP0, SWE, representative coding quality, vision, cancellation, full-context or real-client acceptance.", "Initial incumbent strict diagnostic preceded structured scout plan; full frozen serving-comparison envelope unavailable." ], "decision": { "evidence_labels": [ "functional", "capacity" ], "decision_labels": [ "no-promotion" ], "promotion_authorized": false, "next_gate": "Retain incumbent. A future conversion requires a frozen matched finalist and independent quality/client acceptance." }, "evidence": [ "raw/benchmark-source-verification.json", "raw/candidate-end.json", "raw/candidate-export-memory-pressure.json", "raw/candidate-final-absence.json", "raw/candidate-final-unload.json", "raw/candidate-package-identity.json", "raw/candidate-r2-logs.json", "raw/candidate-r3-preparation-summary.json", "raw/candidate-r4-final-metrics.json", "raw/candidate-r4-preparation-logs.json", "raw/candidate-r4-resource-samples.jsonl", "raw/candidate-runtime-samples.jsonl", "raw/candidate-warm-final-logs.json", "raw/candidate-warm-final-metrics.json", "raw/candidate-warm-load-timing.json", "raw/flare-natural-ctx10240-c1.json", "raw/flare-natural-ctx10240-c1.json.journal.jsonl", "raw/flare-natural-ctx10240-c4.json", "raw/flare-natural-ctx10240-c4.json.journal.jsonl", "raw/flare-natural-ctx32768-c1.json", "raw/flare-natural-ctx32768-c1.json.journal.jsonl", "raw/flare-natural-ctx32768-c4.json", "raw/flare-natural-ctx32768-c4.json.journal.jsonl", "raw/flare-preflight.json", "raw/flare-strict-ctx8192-c1.json", "raw/flare-strict-ctx8192-c1.json.journal.jsonl", "raw/flare-strict-ctx8192-c4.json", "raw/flare-strict-ctx8192-c4.json.journal.jsonl", "raw/flare-warm-preflight.json", "raw/incumbent-before-interruption.json", "raw/incumbent-capacity-c1.json", "raw/incumbent-capacity-c1.json.journal.jsonl", "raw/incumbent-capacity-c4.json", "raw/incumbent-capacity-c4.json.journal.jsonl", "raw/incumbent-capacity-summary.json", "raw/incumbent-final-status.json", "raw/incumbent-natural-ctx10240-c1.json", "raw/incumbent-natural-ctx10240-c1.json.journal.jsonl", "raw/incumbent-natural-ctx10240-c4.json", "raw/incumbent-natural-ctx10240-c4.json.journal.jsonl", "raw/incumbent-natural-ctx32768-c1.json", "raw/incumbent-natural-ctx32768-c1.json.journal.jsonl", "raw/incumbent-natural-ctx32768-c4.json", "raw/incumbent-natural-ctx32768-c4.json.journal.jsonl", "raw/incumbent-runtime-samples.jsonl.gz", "raw/process-lesson.json", "raw/protected-manifest-verification.json", "raw/restoration-final.json", "raw/restore-final-apply.json", "raw/restored-incumbent-preflight.json", "raw/restored-routed-preflight.json", "raw/router-restored-readmit.json", "raw/router-restored-status.json", "raw/scout-execution-priority.json" ], "metrics_coverage": { "profile": "serving-comparison/v1", "status": "partial", "frozen_plan": null, "required_cells": [], "cells": {}, "missing_required_metrics": [ "full pre-request content-bound serving comparison envelope", "cold-C4", "MTP0 control", "full-window measured capacity", "quality and client acceptance" ] }, "comparison_evidence": null }