{ "study": "repeated-seed-policy-consolidation", "protocol": "matched-seeds-bootstrap-and-bounded-offline-oracle", "repetitions": 12, "bootstrap_samples": 600, "config": { "model": "Qwen2.5-3B", "accelerator": "L4", "quantization": "int8", "seed": 7, "duration_s": 120.0, "workflow_rate_rps": 0.18, "dynamic_prompt_tokens_mean": 320, "dynamic_prompt_cv": 0.4, "output_tokens_mean": 64, "output_tokens_cv": 0.45, "tool_gap_mean_s": 1.2, "tool_gap_cv": 0.65, "max_steps": 8, "shift_fraction": 0.55, "prefix_cache_budget_fraction": 0.62, "host_bandwidth_gbps": 32.0, "transfer_base_ms": 0.15, "prefetch_policy": "decayed", "transition_decay": 0.85, "confidence_threshold": 0.55, "prior_strength": 0.35, "forecast_horizon": 3, "prefetch_top_k": 2, "forecast_discount": 0.75, "forecast_min_score": 0.1, "utility_threshold_ms": 0.0, "timeline_points": 240 }, "policies": [ { "label": "No prefetch", "mean_ttft_ms": 6802.071434116555, "median_ttft_ms": 6490.661548131788, "ttft_ci95_low_ms": 5079.289415881574, "ttft_ci95_high_ms": 8648.920669433282, "paired_delta_vs_top1_mean_ms": -123.246831951494, "paired_delta_ci95_low_ms": -347.35450716706316, "paired_delta_ci95_high_ms": 47.018145035094136, "ttft_win_rate": 0.4166666666666667, "pareto_stability": 0.9166666666666666, "median_oracle_regret_ms": 34.71482691034703, "mean_oracle_regret_ms": 71.59505551067814, "oracle_regret_ci95_low_ms": 20.256193997241034, "oracle_regret_ci95_high_ms": 140.55634036325446, "worst_seed_ttft_ms": 14328.174603781865, "mean_unused_prefetch_gb": 0.0, "mean_hbm_gb": 0.05799153067644665, "mean_completion_rate": 1.0, "robust_rank": 2 }, { "label": "Top-1 decayed", "mean_ttft_ms": 6925.318266068049, "median_ttft_ms": 6474.91717792896, "ttft_ci95_low_ms": 5105.737657279251, "ttft_ci95_high_ms": 8965.982492493727, "paired_delta_vs_top1_mean_ms": 0.0, "paired_delta_ci95_low_ms": 0.0, "paired_delta_ci95_high_ms": 0.0, "ttft_win_rate": 0.4166666666666667, "pareto_stability": 0.75, "median_oracle_regret_ms": 59.272217524663574, "mean_oracle_regret_ms": 194.84188746217214, "oracle_regret_ci95_low_ms": 54.639857512317526, "oracle_regret_ci95_high_ms": 381.3867898898039, "worst_seed_ttft_ms": 14688.079385783494, "mean_unused_prefetch_gb": 0.0933888, "mean_hbm_gb": 0.05756119811086612, "mean_completion_rate": 1.0, "robust_rank": 1 }, { "label": "Multi-step top-k", "mean_ttft_ms": 7053.099997696324, "median_ttft_ms": 6640.303485776007, "ttft_ci95_low_ms": 5341.689146280877, "ttft_ci95_high_ms": 9082.241581969562, "paired_delta_vs_top1_mean_ms": 127.78173162827515, "paired_delta_ci95_low_ms": -78.61311912594543, "paired_delta_ci95_high_ms": 285.7402664741835, "ttft_win_rate": 0.08333333333333333, "pareto_stability": 0.8333333333333334, "median_oracle_regret_ms": 215.01034621539475, "mean_oracle_regret_ms": 322.6236190904473, "oracle_regret_ci95_low_ms": 163.1324022990532, "oracle_regret_ci95_high_ms": 482.28187467234943, "worst_seed_ttft_ms": 14999.960971925384, "mean_unused_prefetch_gb": 2.776301568, "mean_hbm_gb": 0.05447990052348106, "mean_completion_rate": 1.0, "robust_rank": 4 }, { "label": "Utility-aware multi-step", "mean_ttft_ms": 7008.096711452864, "median_ttft_ms": 6507.529283460886, "ttft_ci95_low_ms": 5199.467094878383, "ttft_ci95_high_ms": 9162.994321123462, "paired_delta_vs_top1_mean_ms": 82.77844538481479, "paired_delta_ci95_low_ms": -23.44354791792356, "paired_delta_ci95_high_ms": 214.12739203275032, "ttft_win_rate": 0.08333333333333333, "pareto_stability": 0.5833333333333334, "median_oracle_regret_ms": 133.08541840114094, "mean_oracle_regret_ms": 277.6203328469869, "oracle_regret_ci95_low_ms": 100.84880035727063, "oracle_regret_ci95_high_ms": 481.4559675209379, "worst_seed_ttft_ms": 14634.729400052454, "mean_unused_prefetch_gb": 1.69377792, "mean_hbm_gb": 0.057408777204922434, "mean_completion_rate": 1.0, "robust_rank": 3 } ], "seed_runs": [ { "rep": 1, "seed": 7, "rows": [ { "label": "No prefetch", "p95_ttft_ms": 2508.9158110308517, "p95_workflow_e2e_ms": 26656.556855177485, "workflow_throughput_rps": 0.1577017775042944, "completion_rate": 1.0, "prefix_hit_rate": 0.42857142857142855, "prefetch_utilization": 0.0, "future_role_recall_at_k": 0.0, "mean_hbm_gb": 0.0599599629511816, "unused_prefetch_gb": 0.0, "pressure_evictions": 49, "saved_prefill_tokens": 26176, "pareto": true, "ttft_winner": false, "oracle_regret_ms": 49.90402867636931, "oracle_regret_pct": 2.0294343050518706 }, { "label": "Top-1 decayed", "p95_ttft_ms": 2459.0117823544824, "p95_workflow_e2e_ms": 26667.23774828302, "workflow_throughput_rps": 0.15768970423702405, "completion_rate": 1.0, "prefix_hit_rate": 0.4945054945054945, "prefetch_utilization": 0.8125, "future_role_recall_at_k": 0.13526570048309178, "mean_hbm_gb": 0.05817502692720098, "unused_prefetch_gb": 0.09909043199999998, "pressure_evictions": 59, "saved_prefill_tokens": 32320, "pareto": true, "ttft_winner": true, "oracle_regret_ms": 0.0, "oracle_regret_pct": 0.0 }, { "label": "Multi-step top-k", "p95_ttft_ms": 2599.2291767138786, "p95_workflow_e2e_ms": 27096.713211323044, "workflow_throughput_rps": 0.1572831336707985, "completion_rate": 1.0, "prefix_hit_rate": 0.37362637362637363, "prefetch_utilization": 0.1640625, "future_role_recall_at_k": 0.5700483091787439, "mean_hbm_gb": 0.054251348984577404, "unused_prefetch_gb": 2.96091648, "pressure_evictions": 183, "saved_prefill_tokens": 24320, "pareto": true, "ttft_winner": false, "oracle_regret_ms": 140.21739435939617, "oracle_regret_pct": 5.702184729881173 }, { "label": "Utility-aware multi-step", "p95_ttft_ms": 2593.2876340481835, "p95_workflow_e2e_ms": 26697.511231196324, "workflow_throughput_rps": 0.15773330264095717, "completion_rate": 1.0, "prefix_hit_rate": 0.4835164835164835, "prefetch_utilization": 0.36619718309859156, "future_role_recall_at_k": 0.5096618357487923, "mean_hbm_gb": 0.0583545951025611, "unused_prefetch_gb": 1.2386304, "pressure_evictions": 115, "saved_prefill_tokens": 31744, "pareto": false, "ttft_winner": false, "oracle_regret_ms": 134.27585169370104, "oracle_regret_pct": 5.460561541723606 } ], "oracle": { "label": "Top-1 decayed", "kind": "deployable", "config": { "policy": "decayed", "horizon": 3, "top_k": 2 }, "metrics": { "label": "Top-1 decayed", "p95_ttft_ms": 2459.0117823544824, "p95_workflow_e2e_ms": 26667.23774828302, "workflow_throughput_rps": 0.15768970423702405, "completion_rate": 1.0, "prefix_hit_rate": 0.4945054945054945, "prefetch_utilization": 0.8125, "future_role_recall_at_k": 0.13526570048309178, "mean_hbm_gb": 0.05817502692720098, "unused_prefetch_gb": 0.09909043199999998, "pressure_evictions": 59, "saved_prefill_tokens": 32320 }, "candidate_count": 19, "definition": "bounded-full-trace-serving-oracle", "note": "Exhaustive upper bound over the declared candidate family, including clairvoyant future-set plans. It uses the complete trace for policy selection and future-role actions, but is not a proof of global optimality." } }, { "rep": 2, "seed": 1016, "rows": [ { "label": "No prefetch", "p95_ttft_ms": 14328.174603781865, "p95_workflow_e2e_ms": 100660.43087786615, "workflow_throughput_rps": 0.1552783155805788, "completion_rate": 1.0, "prefix_hit_rate": 0.5517241379310345, "prefetch_utilization": 0.0, "future_role_recall_at_k": 0.0, "mean_hbm_gb": 0.056923401277366226, "unused_prefetch_gb": 0.0, "pressure_evictions": 63, "saved_prefill_tokens": 55296, "pareto": true, "ttft_winner": true, "oracle_regret_ms": 0.0, "oracle_regret_pct": 0.0 }, { "label": "Top-1 decayed", "p95_ttft_ms": 14688.079385783494, "p95_workflow_e2e_ms": 100710.97712955547, "workflow_throughput_rps": 0.15479050575342704, "completion_rate": 1.0, "prefix_hit_rate": 0.5379310344827586, "prefetch_utilization": 0.4166666666666667, "future_role_recall_at_k": 0.08764367816091954, "mean_hbm_gb": 0.056442561619372084, "unused_prefetch_gb": 0.202899456, "pressure_evictions": 77, "saved_prefill_tokens": 54272, "pareto": true, "ttft_winner": false, "oracle_regret_ms": 359.9047820016294, "oracle_regret_pct": 2.5118676450706707 }, { "label": "Multi-step top-k", "p95_ttft_ms": 14999.960971925384, "p95_workflow_e2e_ms": 101549.98705723105, "workflow_throughput_rps": 0.15434055828231827, "completion_rate": 1.0, "prefix_hit_rate": 0.45517241379310347, "prefetch_utilization": 0.1787709497206704, "future_role_recall_at_k": 0.5689655172413793, "mean_hbm_gb": 0.05619447488989339, "unused_prefetch_gb": 4.05798912, "pressure_evictions": 256, "saved_prefill_tokens": 48256, "pareto": true, "ttft_winner": false, "oracle_regret_ms": 671.7863681435192, "oracle_regret_pct": 4.688569107513555 }, { "label": "Utility-aware multi-step", "p95_ttft_ms": 14634.729400052454, "p95_workflow_e2e_ms": 100642.13358865703, "workflow_throughput_rps": 0.15493195402903986, "completion_rate": 1.0, "prefix_hit_rate": 0.5172413793103449, "prefetch_utilization": 0.24603174603174602, "future_role_recall_at_k": 0.5545977011494253, "mean_hbm_gb": 0.05871891330180658, "unused_prefetch_gb": 2.5952256, "pressure_evictions": 194, "saved_prefill_tokens": 54976, "pareto": false, "ttft_winner": false, "oracle_regret_ms": 306.5547962705896, "oracle_regret_pct": 2.139524431742168 } ], "oracle": { "label": "No prefetch", "kind": "deployable", "config": { "policy": "none", "horizon": 3, "top_k": 2 }, "metrics": { "label": "No prefetch", "p95_ttft_ms": 14328.174603781865, "p95_workflow_e2e_ms": 100660.43087786615, "workflow_throughput_rps": 0.1552783155805788, "completion_rate": 1.0, "prefix_hit_rate": 0.5517241379310345, "prefetch_utilization": 0.0, "future_role_recall_at_k": 0.0, "mean_hbm_gb": 0.056923401277366226, "unused_prefetch_gb": 0.0, "pressure_evictions": 63, "saved_prefill_tokens": 55296 }, "candidate_count": 19, "definition": "bounded-full-trace-serving-oracle", "note": "Exhaustive upper bound over the declared candidate family, including clairvoyant future-set plans. It uses the complete trace for policy selection and future-role actions, but is not a proof of global optimality." } }, { "rep": 3, "seed": 2025, "rows": [ { "label": "No prefetch", "p95_ttft_ms": 5155.081577515879, "p95_workflow_e2e_ms": 39206.99976041411, "workflow_throughput_rps": 0.14003550604961185, "completion_rate": 1.0, "prefix_hit_rate": 0.5686274509803921, "prefetch_utilization": 0.0, "future_role_recall_at_k": 0.0, "mean_hbm_gb": 0.05860227184049213, "unused_prefetch_gb": 0.0, "pressure_evictions": 41, "saved_prefill_tokens": 41664, "pareto": true, "ttft_winner": false, "oracle_regret_ms": 67.74273330644246, "oracle_regret_pct": 1.3315946781006989 }, { "label": "Top-1 decayed", "p95_ttft_ms": 5087.3388442094365, "p95_workflow_e2e_ms": 38880.38800539655, "workflow_throughput_rps": 0.14024477205584812, "completion_rate": 1.0, "prefix_hit_rate": 0.5980392156862745, "prefetch_utilization": 0.7142857142857143, "future_role_recall_at_k": 0.07028112449799197, "mean_hbm_gb": 0.05831155256626185, "unused_prefetch_gb": 0.066060288, "pressure_evictions": 45, "saved_prefill_tokens": 44288, "pareto": true, "ttft_winner": true, "oracle_regret_ms": 0.0, "oracle_regret_pct": 0.0 }, { "label": "Multi-step top-k", "p95_ttft_ms": 5308.289648386624, "p95_workflow_e2e_ms": 39362.44388383438, "workflow_throughput_rps": 0.13921395060797437, "completion_rate": 1.0, "prefix_hit_rate": 0.5686274509803921, "prefetch_utilization": 0.24427480916030533, "future_role_recall_at_k": 0.5341365461847389, "mean_hbm_gb": 0.054135117019982595, "unused_prefetch_gb": 2.753298432, "pressure_evictions": 172, "saved_prefill_tokens": 43072, "pareto": true, "ttft_winner": false, "oracle_regret_ms": 220.95080417718782, "oracle_regret_pct": 4.3431509271036814 }, { "label": "Utility-aware multi-step", "p95_ttft_ms": 5219.233829318017, "p95_workflow_e2e_ms": 39436.99299180437, "workflow_throughput_rps": 0.13981017212606242, "completion_rate": 1.0, "prefix_hit_rate": 0.5588235294117647, "prefetch_utilization": 0.25274725274725274, "future_role_recall_at_k": 0.5281124497991968, "mean_hbm_gb": 0.05662322676367652, "unused_prefetch_gb": 1.882718208, "pressure_evictions": 133, "saved_prefill_tokens": 42624, "pareto": true, "ttft_winner": false, "oracle_regret_ms": 131.89498510858084, "oracle_regret_pct": 2.5926125455297266 } ], "oracle": { "label": "Top-1 decayed", "kind": "deployable", "config": { "policy": "decayed", "horizon": 3, "top_k": 2 }, "metrics": { "label": "Top-1 decayed", "p95_ttft_ms": 5087.3388442094365, "p95_workflow_e2e_ms": 38880.38800539655, "workflow_throughput_rps": 0.14024477205584812, "completion_rate": 1.0, "prefix_hit_rate": 0.5980392156862745, "prefetch_utilization": 0.7142857142857143, "future_role_recall_at_k": 0.07028112449799197, "mean_hbm_gb": 0.05831155256626185, "unused_prefetch_gb": 0.066060288, "pressure_evictions": 45, "saved_prefill_tokens": 44288 }, "candidate_count": 19, "definition": "bounded-full-trace-serving-oracle", "note": "Exhaustive upper bound over the declared candidate family, including clairvoyant future-set plans. It uses the complete trace for policy selection and future-role actions, but is not a proof of global optimality." } }, { "rep": 4, "seed": 3034, "rows": [ { "label": "No prefetch", "p95_ttft_ms": 8081.134170861211, "p95_workflow_e2e_ms": 48408.234787678535, "workflow_throughput_rps": 0.19158305063580688, "completion_rate": 1.0, "prefix_hit_rate": 0.6384615384615384, "prefetch_utilization": 0.0, "future_role_recall_at_k": 0.0, "mean_hbm_gb": 0.058352465405090974, "unused_prefetch_gb": 0.0, "pressure_evictions": 44, "saved_prefill_tokens": 56064, "pareto": true, "ttft_winner": true, "oracle_regret_ms": 0.0, "oracle_regret_pct": 0.0 }, { "label": "Top-1 decayed", "p95_ttft_ms": 8345.445612762906, "p95_workflow_e2e_ms": 52337.020757564795, "workflow_throughput_rps": 0.19128533155691, "completion_rate": 1.0, "prefix_hit_rate": 0.6153846153846154, "prefetch_utilization": 1.0, "future_role_recall_at_k": 0.06397306397306397, "mean_hbm_gb": 0.058647667236896704, "unused_prefetch_gb": 0.0, "pressure_evictions": 51, "saved_prefill_tokens": 54016, "pareto": false, "ttft_winner": false, "oracle_regret_ms": 264.3114419016947, "oracle_regret_pct": 3.270722107977659 }, { "label": "Multi-step top-k", "p95_ttft_ms": 8856.407713802682, "p95_workflow_e2e_ms": 55194.71833526091, "workflow_throughput_rps": 0.1895469793366219, "completion_rate": 1.0, "prefix_hit_rate": 0.4307692307692308, "prefetch_utilization": 0.15853658536585366, "future_role_recall_at_k": 0.5168350168350169, "mean_hbm_gb": 0.0523639442797909, "unused_prefetch_gb": 3.803185152, "pressure_evictions": 236, "saved_prefill_tokens": 41984, "pareto": true, "ttft_winner": false, "oracle_regret_ms": 775.2735429414715, "oracle_regret_pct": 9.593622956254546 }, { "label": "Utility-aware multi-step", "p95_ttft_ms": 9065.113180839384, "p95_workflow_e2e_ms": 54196.43843042602, "workflow_throughput_rps": 0.1905534529191346, "completion_rate": 1.0, "prefix_hit_rate": 0.5076923076923077, "prefetch_utilization": 0.26666666666666666, "future_role_recall_at_k": 0.4764309764309764, "mean_hbm_gb": 0.05500864112250217, "unused_prefetch_gb": 2.538602496, "pressure_evictions": 181, "saved_prefill_tokens": 48640, "pareto": true, "ttft_winner": false, "oracle_regret_ms": 983.979009978173, "oracle_regret_pct": 12.17624889246592 } ], "oracle": { "label": "No prefetch", "kind": "deployable", "config": { "policy": "none", "horizon": 3, "top_k": 2 }, "metrics": { "label": "No prefetch", "p95_ttft_ms": 8081.134170861211, "p95_workflow_e2e_ms": 48408.234787678535, "workflow_throughput_rps": 0.19158305063580688, "completion_rate": 1.0, "prefix_hit_rate": 0.6384615384615384, "prefetch_utilization": 0.0, "future_role_recall_at_k": 0.0, "mean_hbm_gb": 0.058352465405090974, "unused_prefetch_gb": 0.0, "pressure_evictions": 44, "saved_prefill_tokens": 56064 }, "candidate_count": 19, "definition": "bounded-full-trace-serving-oracle", "note": "Exhaustive upper bound over the declared candidate family, including clairvoyant future-set plans. It uses the complete trace for policy selection and future-role actions, but is not a proof of global optimality." } }, { "rep": 5, "seed": 4043, "rows": [ { "label": "No prefetch", "p95_ttft_ms": 3974.694612251191, "p95_workflow_e2e_ms": 18764.681149810734, "workflow_throughput_rps": 0.14120209678649592, "completion_rate": 1.0, "prefix_hit_rate": 0.48484848484848486, "prefetch_utilization": 0.0, "future_role_recall_at_k": 0.0, "mean_hbm_gb": 0.053247840133508445, "unused_prefetch_gb": 0.0, "pressure_evictions": 32, "saved_prefill_tokens": 21760, "pareto": true, "ttft_winner": false, "oracle_regret_ms": 85.79909391552974, "oracle_regret_pct": 2.2062586539287987 }, { "label": "Top-1 decayed", "p95_ttft_ms": 3888.895518335661, "p95_workflow_e2e_ms": 18671.951222411175, "workflow_throughput_rps": 0.1412992144994211, "completion_rate": 1.0, "prefix_hit_rate": 0.5, "prefetch_utilization": 0.7142857142857143, "future_role_recall_at_k": 0.05555555555555555, "mean_hbm_gb": 0.05256473368734914, "unused_prefetch_gb": 0.061341696, "pressure_evictions": 38, "saved_prefill_tokens": 22720, "pareto": true, "ttft_winner": true, "oracle_regret_ms": 0.0, "oracle_regret_pct": 0.0 }, { "label": "Multi-step top-k", "p95_ttft_ms": 3990.111065086865, "p95_workflow_e2e_ms": 18883.9144056687, "workflow_throughput_rps": 0.1409739259838125, "completion_rate": 1.0, "prefix_hit_rate": 0.3939393939393939, "prefetch_utilization": 0.16091954022988506, "future_role_recall_at_k": 0.5104166666666666, "mean_hbm_gb": 0.045406284458674596, "unused_prefetch_gb": 2.01719808, "pressure_evictions": 125, "saved_prefill_tokens": 18496, "pareto": true, "ttft_winner": false, "oracle_regret_ms": 101.21554675120387, "oracle_regret_pct": 2.6026810510589726 }, { "label": "Utility-aware multi-step", "p95_ttft_ms": 3961.805766823279, "p95_workflow_e2e_ms": 19016.7853637802, "workflow_throughput_rps": 0.1409052485639348, "completion_rate": 1.0, "prefix_hit_rate": 0.36363636363636365, "prefetch_utilization": 0.16923076923076924, "future_role_recall_at_k": 0.4895833333333333, "mean_hbm_gb": 0.04901078264991126, "unused_prefetch_gb": 1.519386624, "pressure_evictions": 105, "saved_prefill_tokens": 16960, "pareto": true, "ttft_winner": false, "oracle_regret_ms": 72.91024848761799, "oracle_regret_pct": 1.8748317650565616 } ], "oracle": { "label": "Top-1 decayed", "kind": "deployable", "config": { "policy": "decayed", "horizon": 3, "top_k": 2 }, "metrics": { "label": "Top-1 decayed", "p95_ttft_ms": 3888.895518335661, "p95_workflow_e2e_ms": 18671.951222411175, "workflow_throughput_rps": 0.1412992144994211, "completion_rate": 1.0, "prefix_hit_rate": 0.5, "prefetch_utilization": 0.7142857142857143, "future_role_recall_at_k": 0.05555555555555555, "mean_hbm_gb": 0.05256473368734914, "unused_prefetch_gb": 0.061341696, "pressure_evictions": 38, "saved_prefill_tokens": 22720 }, "candidate_count": 19, "definition": "bounded-full-trace-serving-oracle", "note": "Exhaustive upper bound over the declared candidate family, including clairvoyant future-set plans. It uses the complete trace for policy selection and future-role actions, but is not a proof of global optimality." } }, { "rep": 6, "seed": 5052, "rows": [ { "label": "No prefetch", "p95_ttft_ms": 7309.107177665744, "p95_workflow_e2e_ms": 47619.448084895426, "workflow_throughput_rps": 0.1484392590003076, "completion_rate": 1.0, "prefix_hit_rate": 0.48484848484848486, "prefetch_utilization": 0.0, "future_role_recall_at_k": 0.0, "mean_hbm_gb": 0.05471706918379192, "unused_prefetch_gb": 0.0, "pressure_evictions": 49, "saved_prefill_tokens": 33280, "pareto": true, "ttft_winner": true, "oracle_regret_ms": 0.0, "oracle_regret_pct": 0.0 }, { "label": "Top-1 decayed", "p95_ttft_ms": 7309.87009396774, "p95_workflow_e2e_ms": 47381.367578903475, "workflow_throughput_rps": 0.14866123026478553, "completion_rate": 1.0, "prefix_hit_rate": 0.5454545454545454, "prefetch_utilization": 0.75, "future_role_recall_at_k": 0.11842105263157894, "mean_hbm_gb": 0.054990227626166545, "unused_prefetch_gb": 0.089653248, "pressure_evictions": 54, "saved_prefill_tokens": 38592, "pareto": false, "ttft_winner": false, "oracle_regret_ms": 0.7629163019964835, "oracle_regret_pct": 0.0104378863717816 }, { "label": "Multi-step top-k", "p95_ttft_ms": 7517.707112366715, "p95_workflow_e2e_ms": 47497.83258699973, "workflow_throughput_rps": 0.14871579121181555, "completion_rate": 1.0, "prefix_hit_rate": 0.5454545454545454, "prefetch_utilization": 0.33620689655172414, "future_role_recall_at_k": 0.5394736842105263, "mean_hbm_gb": 0.05581068796513829, "unused_prefetch_gb": 2.137522176, "pressure_evictions": 159, "saved_prefill_tokens": 40064, "pareto": false, "ttft_winner": false, "oracle_regret_ms": 208.59993470097106, "oracle_regret_pct": 2.8539728537349225 }, { "label": "Utility-aware multi-step", "p95_ttft_ms": 7506.956342852778, "p95_workflow_e2e_ms": 47305.033350410624, "workflow_throughput_rps": 0.14888303988588122, "completion_rate": 1.0, "prefix_hit_rate": 0.5656565656565656, "prefetch_utilization": 0.34177215189873417, "future_role_recall_at_k": 0.5197368421052632, "mean_hbm_gb": 0.05676471264317459, "unused_prefetch_gb": 1.42737408, "pressure_evictions": 119, "saved_prefill_tokens": 41856, "pareto": false, "ttft_winner": false, "oracle_regret_ms": 197.8491651870345, "oracle_regret_pct": 2.7068855385182644 } ], "oracle": { "label": "No prefetch", "kind": "deployable", "config": { "policy": "none", "horizon": 3, "top_k": 2 }, "metrics": { "label": "No prefetch", "p95_ttft_ms": 7309.107177665744, "p95_workflow_e2e_ms": 47619.448084895426, "workflow_throughput_rps": 0.1484392590003076, "completion_rate": 1.0, "prefix_hit_rate": 0.48484848484848486, "prefetch_utilization": 0.0, "future_role_recall_at_k": 0.0, "mean_hbm_gb": 0.05471706918379192, "unused_prefetch_gb": 0.0, "pressure_evictions": 49, "saved_prefill_tokens": 33280 }, "candidate_count": 19, "definition": "bounded-full-trace-serving-oracle", "note": "Exhaustive upper bound over the declared candidate family, including clairvoyant future-set plans. It uses the complete trace for policy selection and future-role actions, but is not a proof of global optimality." } }, { "rep": 7, "seed": 6061, "rows": [ { "label": "No prefetch", "p95_ttft_ms": 5672.215918597831, "p95_workflow_e2e_ms": 39980.01576300871, "workflow_throughput_rps": 0.1526413288015884, "completion_rate": 1.0, "prefix_hit_rate": 0.4574468085106383, "prefetch_utilization": 0.0, "future_role_recall_at_k": 0.0, "mean_hbm_gb": 0.05576787962082442, "unused_prefetch_gb": 0.0, "pressure_evictions": 49, "saved_prefill_tokens": 30272, "pareto": true, "ttft_winner": false, "oracle_regret_ms": 222.25583063586964, "oracle_regret_pct": 4.0781185008454495 }, { "label": "Top-1 decayed", "p95_ttft_ms": 5639.96426189018, "p95_workflow_e2e_ms": 42335.09842778248, "workflow_throughput_rps": 0.15362763283307423, "completion_rate": 1.0, "prefix_hit_rate": 0.4574468085106383, "prefetch_utilization": 0.375, "future_role_recall_at_k": 0.07432432432432433, "mean_hbm_gb": 0.056659199341786655, "unused_prefetch_gb": 0.155713536, "pressure_evictions": 57, "saved_prefill_tokens": 31424, "pareto": true, "ttft_winner": false, "oracle_regret_ms": 190.00417392821873, "oracle_regret_pct": 3.4863406495013747 }, { "label": "Multi-step top-k", "p95_ttft_ms": 5449.960087961961, "p95_workflow_e2e_ms": 37358.71643947743, "workflow_throughput_rps": 0.15301958531801996, "completion_rate": 1.0, "prefix_hit_rate": 0.5319148936170213, "prefetch_utilization": 0.30392156862745096, "future_role_recall_at_k": 0.5112612612612613, "mean_hbm_gb": 0.054008815280052504, "unused_prefetch_gb": 1.880358912, "pressure_evictions": 144, "saved_prefill_tokens": 37120, "pareto": true, "ttft_winner": true, "oracle_regret_ms": 0.0, "oracle_regret_pct": 0.0 }, { "label": "Utility-aware multi-step", "p95_ttft_ms": 5471.268696264992, "p95_workflow_e2e_ms": 37317.903368275336, "workflow_throughput_rps": 0.15319363127947774, "completion_rate": 1.0, "prefix_hit_rate": 0.5531914893617021, "prefetch_utilization": 0.36363636363636365, "future_role_recall_at_k": 0.45495495495495497, "mean_hbm_gb": 0.05709563959469793, "unused_prefetch_gb": 1.078198272, "pressure_evictions": 106, "saved_prefill_tokens": 38208, "pareto": true, "ttft_winner": false, "oracle_regret_ms": 21.30860830303118, "oracle_regret_pct": 0.39098650190298245 } ], "oracle": { "label": "Multi-step top-k", "kind": "deployable", "config": { "policy": "multistep", "horizon": 3, "top_k": 2 }, "metrics": { "label": "Multi-step top-k", "p95_ttft_ms": 5449.960087961961, "p95_workflow_e2e_ms": 37358.71643947743, "workflow_throughput_rps": 0.15301958531801996, "completion_rate": 1.0, "prefix_hit_rate": 0.5319148936170213, "prefetch_utilization": 0.30392156862745096, "future_role_recall_at_k": 0.5112612612612613, "mean_hbm_gb": 0.054008815280052504, "unused_prefetch_gb": 1.880358912, "pressure_evictions": 144, "saved_prefill_tokens": 37120 }, "candidate_count": 19, "definition": "bounded-full-trace-serving-oracle", "note": "Exhaustive upper bound over the declared candidate family, including clairvoyant future-set plans. It uses the complete trace for policy selection and future-role actions, but is not a proof of global optimality." } }, { "rep": 8, "seed": 7070, "rows": [ { "label": "No prefetch", "p95_ttft_ms": 7955.834288389353, "p95_workflow_e2e_ms": 63458.63274772634, "workflow_throughput_rps": 0.14314957761896735, "completion_rate": 1.0, "prefix_hit_rate": 0.5392156862745098, "prefetch_utilization": 0.0, "future_role_recall_at_k": 0.0, "mean_hbm_gb": 0.05757162554782224, "unused_prefetch_gb": 0.0, "pressure_evictions": 44, "saved_prefill_tokens": 36608, "pareto": false, "ttft_winner": false, "oracle_regret_ms": 360.1196655210979, "oracle_regret_pct": 4.741089988253289 }, { "label": "Top-1 decayed", "p95_ttft_ms": 7663.490235151211, "p95_workflow_e2e_ms": 63010.589234624604, "workflow_throughput_rps": 0.14356812595083052, "completion_rate": 1.0, "prefix_hit_rate": 0.5882352941176471, "prefetch_utilization": 1.0, "future_role_recall_at_k": 0.08125, "mean_hbm_gb": 0.05697446543608208, "unused_prefetch_gb": 0.0, "pressure_evictions": 43, "saved_prefill_tokens": 40448, "pareto": true, "ttft_winner": true, "oracle_regret_ms": 67.77561228295599, "oracle_regret_pct": 0.8922875022042747 }, { "label": "Multi-step top-k", "p95_ttft_ms": 7804.784511121857, "p95_workflow_e2e_ms": 63128.95114673241, "workflow_throughput_rps": 0.14353084410317019, "completion_rate": 1.0, "prefix_hit_rate": 0.5980392156862745, "prefetch_utilization": 0.33766233766233766, "future_role_recall_at_k": 0.5333333333333333, "mean_hbm_gb": 0.05700914860577035, "unused_prefetch_gb": 1.363673088, "pressure_evictions": 115, "saved_prefill_tokens": 42816, "pareto": false, "ttft_winner": false, "oracle_regret_ms": 209.06988825360168, "oracle_regret_pct": 2.7524716058204644 }, { "label": "Utility-aware multi-step", "p95_ttft_ms": 7821.45396304061, "p95_workflow_e2e_ms": 63158.31108861863, "workflow_throughput_rps": 0.14349908334078196, "completion_rate": 1.0, "prefix_hit_rate": 0.5980392156862745, "prefetch_utilization": 0.4230769230769231, "future_role_recall_at_k": 0.5333333333333333, "mean_hbm_gb": 0.06056314337054253, "unused_prefetch_gb": 0.7667712, "pressure_evictions": 90, "saved_prefill_tokens": 41472, "pareto": false, "ttft_winner": false, "oracle_regret_ms": 225.7393401723548, "oracle_regret_pct": 2.97193024462407 } ], "oracle": { "label": "clairvoyant H1/K1", "kind": "clairvoyant", "config": { "policy": "oracle_horizon", "horizon": 1, "top_k": 1 }, "metrics": { "label": "clairvoyant H1/K1", "p95_ttft_ms": 7595.714622868255, "p95_workflow_e2e_ms": 62644.984909819585, "workflow_throughput_rps": 0.1439811704946119, "completion_rate": 1.0, "prefix_hit_rate": 0.6568627450980392, "prefetch_utilization": 0.5111111111111111, "future_role_recall_at_k": 1.0, "mean_hbm_gb": 0.0593557108369556, "unused_prefetch_gb": 0.606339072, "pressure_evictions": 77, "saved_prefill_tokens": 46272 }, "candidate_count": 19, "definition": "bounded-full-trace-serving-oracle", "note": "Exhaustive upper bound over the declared candidate family, including clairvoyant future-set plans. It uses the complete trace for policy selection and future-role actions, but is not a proof of global optimality." } }, { "rep": 9, "seed": 8079, "rows": [ { "label": "No prefetch", "p95_ttft_ms": 2321.8881795970965, "p95_workflow_e2e_ms": 23073.36518580538, "workflow_throughput_rps": 0.14600754420187276, "completion_rate": 1.0, "prefix_hit_rate": 0.5, "prefetch_utilization": 0.0, "future_role_recall_at_k": 0.0, "mean_hbm_gb": 0.05803852146421546, "unused_prefetch_gb": 0.0, "pressure_evictions": 38, "saved_prefill_tokens": 28544, "pareto": true, "ttft_winner": false, "oracle_regret_ms": 23.945841894531895, "oracle_regret_pct": 1.0420558210556516 }, { "label": "Top-1 decayed", "p95_ttft_ms": 2297.9423377025646, "p95_workflow_e2e_ms": 22972.359046059566, "workflow_throughput_rps": 0.14610416511124308, "completion_rate": 1.0, "prefix_hit_rate": 0.524390243902439, "prefetch_utilization": 0.625, "future_role_recall_at_k": 0.044270833333333336, "mean_hbm_gb": 0.05741847323154553, "unused_prefetch_gb": 0.084934656, "pressure_evictions": 44, "saved_prefill_tokens": 30208, "pareto": true, "ttft_winner": true, "oracle_regret_ms": 0.0, "oracle_regret_pct": 0.0 }, { "label": "Multi-step top-k", "p95_ttft_ms": 2358.7664926649986, "p95_workflow_e2e_ms": 23190.068065034015, "workflow_throughput_rps": 0.14635545368977013, "completion_rate": 1.0, "prefix_hit_rate": 0.4878048780487805, "prefetch_utilization": 0.23232323232323232, "future_role_recall_at_k": 0.5130208333333334, "mean_hbm_gb": 0.052585587672706685, "unused_prefetch_gb": 2.102132736, "pressure_evictions": 138, "saved_prefill_tokens": 28608, "pareto": true, "ttft_winner": false, "oracle_regret_ms": 60.82415496243402, "oracle_regret_pct": 2.646896484932897 }, { "label": "Utility-aware multi-step", "p95_ttft_ms": 2354.8786376844973, "p95_workflow_e2e_ms": 22931.69235351764, "workflow_throughput_rps": 0.1463801438395915, "completion_rate": 1.0, "prefix_hit_rate": 0.573170731707317, "prefetch_utilization": 0.3220338983050847, "future_role_recall_at_k": 0.5208333333333334, "mean_hbm_gb": 0.0555423079515505, "unused_prefetch_gb": 1.14425856, "pressure_evictions": 91, "saved_prefill_tokens": 34176, "pareto": true, "ttft_winner": false, "oracle_regret_ms": 56.936299981932734, "oracle_regret_pct": 2.4777079497502306 } ], "oracle": { "label": "Top-1 decayed", "kind": "deployable", "config": { "policy": "decayed", "horizon": 3, "top_k": 2 }, "metrics": { "label": "Top-1 decayed", "p95_ttft_ms": 2297.9423377025646, "p95_workflow_e2e_ms": 22972.359046059566, "workflow_throughput_rps": 0.14610416511124308, "completion_rate": 1.0, "prefix_hit_rate": 0.524390243902439, "prefetch_utilization": 0.625, "future_role_recall_at_k": 0.044270833333333336, "mean_hbm_gb": 0.05741847323154553, "unused_prefetch_gb": 0.084934656, "pressure_evictions": 44, "saved_prefill_tokens": 30208 }, "candidate_count": 19, "definition": "bounded-full-trace-serving-oracle", "note": "Exhaustive upper bound over the declared candidate family, including clairvoyant future-set plans. It uses the complete trace for policy selection and future-role actions, but is not a proof of global optimality." } }, { "rep": 10, "seed": 9088, "rows": [ { "label": "No prefetch", "p95_ttft_ms": 8393.380878206393, "p95_workflow_e2e_ms": 57992.77157876873, "workflow_throughput_rps": 0.18631227548846913, "completion_rate": 1.0, "prefix_hit_rate": 0.5902777777777778, "prefetch_utilization": 0.0, "future_role_recall_at_k": 0.0, "mean_hbm_gb": 0.0608883540146992, "unused_prefetch_gb": 0.0, "pressure_evictions": 57, "saved_prefill_tokens": 59328, "pareto": true, "ttft_winner": true, "oracle_regret_ms": 45.48381192616216, "oracle_regret_pct": 0.5448535309555447 }, { "label": "Top-1 decayed", "p95_ttft_ms": 8603.374943396353, "p95_workflow_e2e_ms": 57847.66200555299, "workflow_throughput_rps": 0.18647490382989684, "completion_rate": 1.0, "prefix_hit_rate": 0.5902777777777778, "prefetch_utilization": 0.6111111111111112, "future_role_recall_at_k": 0.09734513274336283, "mean_hbm_gb": 0.06097675059220976, "unused_prefetch_gb": 0.207618048, "pressure_evictions": 75, "saved_prefill_tokens": 60544, "pareto": false, "ttft_winner": false, "oracle_regret_ms": 255.47787711612182, "oracle_regret_pct": 3.0603860479794 }, { "label": "Multi-step top-k", "p95_ttft_ms": 9184.166173240812, "p95_workflow_e2e_ms": 61245.254849217556, "workflow_throughput_rps": 0.1857623669079808, "completion_rate": 1.0, "prefix_hit_rate": 0.4861111111111111, "prefetch_utilization": 0.2154696132596685, "future_role_recall_at_k": 0.6002949852507374, "mean_hbm_gb": 0.05400505762073378, "unused_prefetch_gb": 3.925868544, "pressure_evictions": 253, "saved_prefill_tokens": 51264, "pareto": true, "ttft_winner": false, "oracle_regret_ms": 836.2691069605808, "oracle_regret_pct": 10.017721832466448 }, { "label": "Utility-aware multi-step", "p95_ttft_ms": 8457.279531340011, "p95_workflow_e2e_ms": 60583.73120156277, "workflow_throughput_rps": 0.18650546463814055, "completion_rate": 1.0, "prefix_hit_rate": 0.4722222222222222, "prefetch_utilization": 0.23809523809523808, "future_role_recall_at_k": 0.5265486725663717, "mean_hbm_gb": 0.05733022563941532, "unused_prefetch_gb": 2.691956736, "pressure_evictions": 200, "saved_prefill_tokens": 50432, "pareto": true, "ttft_winner": false, "oracle_regret_ms": 109.38246505977986, "oracle_regret_pct": 1.3102996382359562 } ], "oracle": { "label": "clairvoyant H3/K3", "kind": "clairvoyant", "config": { "policy": "oracle_horizon", "horizon": 3, "top_k": 3 }, "metrics": { "label": "clairvoyant H3/K3", "p95_ttft_ms": 8347.897066280231, "p95_workflow_e2e_ms": 58517.02815386715, "workflow_throughput_rps": 0.18581440744158234, "completion_rate": 1.0, "prefix_hit_rate": 0.5416666666666666, "prefetch_utilization": 0.24, "future_role_recall_at_k": 1.0, "mean_hbm_gb": 0.05724896307500904, "unused_prefetch_gb": 3.104833536, "pressure_evictions": 214, "saved_prefill_tokens": 55872 }, "candidate_count": 19, "definition": "bounded-full-trace-serving-oracle", "note": "Exhaustive upper bound over the declared candidate family, including clairvoyant future-set plans. It uses the complete trace for policy selection and future-role actions, but is not a proof of global optimality." } }, { "rep": 11, "seed": 10097, "rows": [ { "label": "No prefetch", "p95_ttft_ms": 5511.991884321129, "p95_workflow_e2e_ms": 39379.92626385874, "workflow_throughput_rps": 0.18375402111103414, "completion_rate": 1.0, "prefix_hit_rate": 0.5405405405405406, "prefetch_utilization": 0.0, "future_role_recall_at_k": 0.0, "mean_hbm_gb": 0.06035724347004058, "unused_prefetch_gb": 0.0, "pressure_evictions": 48, "saved_prefill_tokens": 39744, "pareto": true, "ttft_winner": false, "oracle_regret_ms": 3.8896602521344903, "oracle_regret_pct": 0.07061706725662556 }, { "label": "Top-1 decayed", "p95_ttft_ms": 5558.871046835366, "p95_workflow_e2e_ms": 39579.196475851335, "workflow_throughput_rps": 0.18346078167175595, "completion_rate": 1.0, "prefix_hit_rate": 0.5135135135135135, "prefetch_utilization": 0.5, "future_role_recall_at_k": 0.047619047619047616, "mean_hbm_gb": 0.059572909587982954, "unused_prefetch_gb": 0.101449728, "pressure_evictions": 59, "saved_prefill_tokens": 37696, "pareto": true, "ttft_winner": false, "oracle_regret_ms": 50.76882276637116, "oracle_regret_pct": 0.921711702163486 }, { "label": "Multi-step top-k", "p95_ttft_ms": 5762.899859185298, "p95_workflow_e2e_ms": 40501.25080020354, "workflow_throughput_rps": 0.18338521635563007, "completion_rate": 1.0, "prefix_hit_rate": 0.4864864864864865, "prefetch_utilization": 0.23529411764705882, "future_role_recall_at_k": 0.5575396825396826, "mean_hbm_gb": 0.05918820422391816, "unused_prefetch_gb": 2.8311552, "pressure_evictions": 190, "saved_prefill_tokens": 38016, "pareto": true, "ttft_winner": false, "oracle_regret_ms": 254.79763511630335, "oracle_regret_pct": 4.625869759695146 }, { "label": "Utility-aware multi-step", "p95_ttft_ms": 5508.1022240689945, "p95_workflow_e2e_ms": 39558.94633951845, "workflow_throughput_rps": 0.18374015675838837, "completion_rate": 1.0, "prefix_hit_rate": 0.5135135135135135, "prefetch_utilization": 0.3375, "future_role_recall_at_k": 0.5555555555555556, "mean_hbm_gb": 0.06210299331912212, "unused_prefetch_gb": 1.382547456, "pressure_evictions": 131, "saved_prefill_tokens": 39616, "pareto": true, "ttft_winner": true, "oracle_regret_ms": 0.0, "oracle_regret_pct": 0.0 } ], "oracle": { "label": "Utility-aware multi-step", "kind": "deployable", "config": { "policy": "utility", "horizon": 3, "top_k": 2 }, "metrics": { "label": "Utility-aware multi-step", "p95_ttft_ms": 5508.1022240689945, "p95_workflow_e2e_ms": 39558.94633951845, "workflow_throughput_rps": 0.18374015675838837, "completion_rate": 1.0, "prefix_hit_rate": 0.5135135135135135, "prefetch_utilization": 0.3375, "future_role_recall_at_k": 0.5555555555555556, "mean_hbm_gb": 0.06210299331912212, "unused_prefetch_gb": 1.382547456, "pressure_evictions": 131, "saved_prefill_tokens": 39616 }, "candidate_count": 19, "definition": "bounded-full-trace-serving-oracle", "note": "Exhaustive upper bound over the declared candidate family, including clairvoyant future-set plans. It uses the complete trace for policy selection and future-role actions, but is not a proof of global optimality." } }, { "rep": 12, "seed": 11106, "rows": [ { "label": "No prefetch", "p95_ttft_ms": 10412.438107180116, "p95_workflow_e2e_ms": 72277.2430475709, "workflow_throughput_rps": 0.17028600915598113, "completion_rate": 1.0, "prefix_hit_rate": 0.46715328467153283, "prefetch_utilization": 0.0, "future_role_recall_at_k": 0.0, "mean_hbm_gb": 0.061471733208326614, "unused_prefetch_gb": 0.0, "pressure_evictions": 71, "saved_prefill_tokens": 44800, "pareto": true, "ttft_winner": true, "oracle_regret_ms": 0.0, "oracle_regret_pct": 0.0 }, { "label": "Top-1 decayed", "p95_ttft_ms": 11561.535130427194, "p95_workflow_e2e_ms": 71765.84830537342, "workflow_throughput_rps": 0.17047301779333604, "completion_rate": 1.0, "prefix_hit_rate": 0.4744525547445255, "prefetch_utilization": 0.8571428571428571, "future_role_recall_at_k": 0.05, "mean_hbm_gb": 0.06000080947753916, "unused_prefetch_gb": 0.051904512, "pressure_evictions": 84, "saved_prefill_tokens": 46080, "pareto": true, "ttft_winner": false, "oracle_regret_ms": 1149.0970232470772, "oracle_regret_pct": 11.035811319297958 }, { "label": "Multi-step top-k", "p95_ttft_ms": 10804.917159898814, "p95_workflow_e2e_ms": 72858.27325261911, "workflow_throughput_rps": 0.1698397999622364, "completion_rate": 1.0, "prefix_hit_rate": 0.46715328467153283, "prefetch_utilization": 0.24096385542168675, "future_role_recall_at_k": 0.5909090909090909, "mean_hbm_gb": 0.05880013528053406, "unused_prefetch_gb": 3.482320896, "pressure_evictions": 236, "saved_prefill_tokens": 46592, "pareto": true, "ttft_winner": false, "oracle_regret_ms": 392.47905271869786, "oracle_regret_pct": 3.769329033975776 }, { "label": "Utility-aware multi-step", "p95_ttft_ms": 11503.051331101164, "p95_workflow_e2e_ms": 70618.34835590329, "workflow_throughput_rps": 0.1704137373168054, "completion_rate": 1.0, "prefix_hit_rate": 0.45985401459854014, "prefetch_utilization": 0.3090909090909091, "future_role_recall_at_k": 0.5409090909090909, "mean_hbm_gb": 0.06179014500010861, "unused_prefetch_gb": 2.059665408, "pressure_evictions": 181, "saved_prefill_tokens": 45696, "pareto": false, "ttft_winner": false, "oracle_regret_ms": 1090.6132239210474, "oracle_regret_pct": 10.474138839480755 } ], "oracle": { "label": "No prefetch", "kind": "deployable", "config": { "policy": "none", "horizon": 3, "top_k": 2 }, "metrics": { "label": "No prefetch", "p95_ttft_ms": 10412.438107180116, "p95_workflow_e2e_ms": 72277.2430475709, "workflow_throughput_rps": 0.17028600915598113, "completion_rate": 1.0, "prefix_hit_rate": 0.46715328467153283, "prefetch_utilization": 0.0, "future_role_recall_at_k": 0.0, "mean_hbm_gb": 0.061471733208326614, "unused_prefetch_gb": 0.0, "pressure_evictions": 71, "saved_prefill_tokens": 44800 }, "candidate_count": 19, "definition": "bounded-full-trace-serving-oracle", "note": "Exhaustive upper bound over the declared candidate family, including clairvoyant future-set plans. It uses the complete trace for policy selection and future-role actions, but is not a proof of global optimality." } } ], "nominal_winner_first_seed": "Top-1 decayed", "robust_winner": "Top-1 decayed", "oracle": { "definition": "bounded-full-trace-serving-oracle", "candidate_count_per_seed": 19, "median_ttft_ms": 6408.604700867369, "mean_ttft_ms": 6730.476378605877, "selected_plan_frequency": { "Top-1 decayed": 4, "No prefetch": 4, "Multi-step top-k": 1, "clairvoyant H1/K1": 1, "clairvoyant H3/K3": 1, "Utility-aware multi-step": 1 }, "note": "The oracle exhaustively selects among the evaluated deployable policies plus clairvoyant future-set plans over H=1..5 and K=1..3 on each complete trace, under the same cache/bandwidth constraints. It is a bounded information upper bound, not a proof of globally optimal action scheduling." }, "pareto_objectives": [ "p95_step_ttft_ms", "unused_prefetch_gb", "mean_hbm_gb" ], "note": "Robust rank prioritizes how often a policy wins TTFT across matched seeds, then median TTFT and regret to the bounded offline oracle. Pareto stability is reported separately rather than silently overriding latency. Bootstrap intervals quantify seed uncertainty, not real-hardware error." }