{ "artifact": "s06_weka_residency", "sprint": "T3", "cost_usd": 0.0, "claim": "KV token-seconds an agent fleet holds while sessions wait, by wait class, and prefill tokens recomputed under LRU-by-session, TTL and oracle-park policies, on a real agentic trace.", "method": "bench/residency.simulate over traces/agentic_real_cc.jsonl.gz; see module doc", "trace": { "name": "cc_traces_weka_042026", "class": "DERIVED_FROM_REAL", "calls": 59204, "programs": 739, "peak_live_tokens": 134745718 }, "hold_share_by_class_unbounded_1h": { "tool_lt_1s": 0.0, "tool_1_6s": 0.0276, "tool_gt_6s": 0.7695, "human": 0.2029, "unknown": 0.0 }, "runs": [ { "policy": "lru", "capacity_tokens": Infinity, "max_hold_s": 300.0, "park_after_s": null, "ttl_s": null, "calls": 59204, "hits": 55484, "misses": 2981, "first_touches": 739, "hit_rate_excluding_first": 0.949012, "prefill_tokens_recomputed": 485226403, "prefill_tokens_first_touch": 68337940, "restored_tokens": 0, "released_by_ttl": 0, "hbm_token_seconds_by_class": { "tool_lt_1s": 0, "tool_1_6s": 4782898424, "tool_gt_6s": 111177579769, "human": 20057466092, "unknown": 0 }, "host_token_seconds_by_class": { "tool_lt_1s": 0, "tool_1_6s": 0, "tool_gt_6s": 0, "human": 0, "unknown": 0 }, "hbm_token_seconds": 136017944285, "peak_live_tokens": 134745718, "capacity_frac_of_peak": null }, { "policy": "lru", "capacity_tokens": 67372859.0, "max_hold_s": 300.0, "park_after_s": null, "ttl_s": null, "calls": 59204, "hits": 55260, "misses": 3205, "first_touches": 739, "hit_rate_excluding_first": 0.945181, "prefill_tokens_recomputed": 514312379, "prefill_tokens_first_touch": 68337940, "restored_tokens": 0, "released_by_ttl": 0, "hbm_token_seconds_by_class": { "tool_lt_1s": 0, "tool_1_6s": 4782131617, "tool_gt_6s": 111109378520, "human": 19953558432, "unknown": 0 }, "host_token_seconds_by_class": { "tool_lt_1s": 0, "tool_1_6s": 0, "tool_gt_6s": 0, "human": 0, "unknown": 0 }, "hbm_token_seconds": 135845068569, "peak_live_tokens": 67372858, "capacity_frac_of_peak": 0.5 }, { "policy": "lru", "capacity_tokens": 33686429.5, "max_hold_s": 300.0, "park_after_s": null, "ttl_s": null, "calls": 59204, "hits": 52355, "misses": 6110, "first_touches": 739, "hit_rate_excluding_first": 0.895493, "prefill_tokens_recomputed": 864201895, "prefill_tokens_first_touch": 68337940, "restored_tokens": 0, "released_by_ttl": 0, "hbm_token_seconds_by_class": { "tool_lt_1s": 0, "tool_1_6s": 4686333019, "tool_gt_6s": 104948261897, "human": 17468760411, "unknown": 0 }, "host_token_seconds_by_class": { "tool_lt_1s": 0, "tool_1_6s": 0, "tool_gt_6s": 0, "human": 0, "unknown": 0 }, "hbm_token_seconds": 127103355327, "peak_live_tokens": 33686424, "capacity_frac_of_peak": 0.25 }, { "policy": "lru", "capacity_tokens": 16843214.75, "max_hold_s": 300.0, "park_after_s": null, "ttl_s": null, "calls": 59204, "hits": 43691, "misses": 14774, "first_touches": 739, "hit_rate_excluding_first": 0.747302, "prefill_tokens_recomputed": 1842949223, "prefill_tokens_first_touch": 68337940, "restored_tokens": 0, "released_by_ttl": 0, "hbm_token_seconds_by_class": { "tool_lt_1s": 0, "tool_1_6s": 4139797924, "tool_gt_6s": 88247412242, "human": 12086168439, "unknown": 0 }, "host_token_seconds_by_class": { "tool_lt_1s": 0, "tool_1_6s": 0, "tool_gt_6s": 0, "human": 0, "unknown": 0 }, "hbm_token_seconds": 104473378605, "peak_live_tokens": 16843207, "capacity_frac_of_peak": 0.125 }, { "policy": "lru", "capacity_tokens": 8421607.375, "max_hold_s": 300.0, "park_after_s": null, "ttl_s": null, "calls": 59204, "hits": 31223, "misses": 27242, "first_touches": 739, "hit_rate_excluding_first": 0.534046, "prefill_tokens_recomputed": 3312076700, "prefill_tokens_first_touch": 68337940, "restored_tokens": 0, "released_by_ttl": 0, "hbm_token_seconds_by_class": { "tool_lt_1s": 0, "tool_1_6s": 3137923704, "tool_gt_6s": 68582438215, "human": 7626013087, "unknown": 0 }, "host_token_seconds_by_class": { "tool_lt_1s": 0, "tool_1_6s": 0, "tool_gt_6s": 0, "human": 0, "unknown": 0 }, "hbm_token_seconds": 79346375006, "peak_live_tokens": 8421605, "capacity_frac_of_peak": 0.0625 }, { "policy": "lru", "capacity_tokens": Infinity, "max_hold_s": 3600.0, "park_after_s": null, "ttl_s": null, "calls": 59204, "hits": 58123, "misses": 342, "first_touches": 739, "hit_rate_excluding_first": 0.99415, "prefill_tokens_recomputed": 69685755, "prefill_tokens_first_touch": 68337940, "restored_tokens": 0, "released_by_ttl": 0, "hbm_token_seconds_by_class": { "tool_lt_1s": 0, "tool_1_6s": 4871580724, "tool_gt_6s": 135743204278, "human": 35791385210, "unknown": 0 }, "host_token_seconds_by_class": { "tool_lt_1s": 0, "tool_1_6s": 0, "tool_gt_6s": 0, "human": 0, "unknown": 0 }, "hbm_token_seconds": 176406170212, "peak_live_tokens": 134745718, "capacity_frac_of_peak": null }, { "policy": "lru", "capacity_tokens": 67372859.0, "max_hold_s": 3600.0, "park_after_s": null, "ttl_s": null, "calls": 59204, "hits": 57538, "misses": 927, "first_touches": 739, "hit_rate_excluding_first": 0.984144, "prefill_tokens_recomputed": 146259831, "prefill_tokens_first_touch": 68337940, "restored_tokens": 0, "released_by_ttl": 0, "hbm_token_seconds_by_class": { "tool_lt_1s": 0, "tool_1_6s": 4861452236, "tool_gt_6s": 135303903183, "human": 34332496618, "unknown": 0 }, "host_token_seconds_by_class": { "tool_lt_1s": 0, "tool_1_6s": 0, "tool_gt_6s": 0, "human": 0, "unknown": 0 }, "hbm_token_seconds": 174497852037, "peak_live_tokens": 67372858, "capacity_frac_of_peak": 0.5 }, { "policy": "lru", "capacity_tokens": 33686429.5, "max_hold_s": 3600.0, "park_after_s": null, "ttl_s": null, "calls": 59204, "hits": 53921, "misses": 4544, "first_touches": 739, "hit_rate_excluding_first": 0.922278, "prefill_tokens_recomputed": 589578606, "prefill_tokens_first_touch": 68337940, "restored_tokens": 0, "released_by_ttl": 0, "hbm_token_seconds_by_class": { "tool_lt_1s": 0, "tool_1_6s": 4737221504, "tool_gt_6s": 122936223099, "human": 26264599926, "unknown": 0 }, "host_token_seconds_by_class": { "tool_lt_1s": 0, "tool_1_6s": 0, "tool_gt_6s": 0, "human": 0, "unknown": 0 }, "hbm_token_seconds": 153938044529, "peak_live_tokens": 33686424, "capacity_frac_of_peak": 0.25 }, { "policy": "lru", "capacity_tokens": 16843214.75, "max_hold_s": 3600.0, "park_after_s": null, "ttl_s": null, "calls": 59204, "hits": 44774, "misses": 13691, "first_touches": 739, "hit_rate_excluding_first": 0.765826, "prefill_tokens_recomputed": 1638416111, "prefill_tokens_first_touch": 68337940, "restored_tokens": 0, "released_by_ttl": 0, "hbm_token_seconds_by_class": { "tool_lt_1s": 0, "tool_1_6s": 4167554950, "tool_gt_6s": 102189616145, "human": 17312415592, "unknown": 0 }, "host_token_seconds_by_class": { "tool_lt_1s": 0, "tool_1_6s": 0, "tool_gt_6s": 0, "human": 0, "unknown": 0 }, "hbm_token_seconds": 123669586687, "peak_live_tokens": 16843207, "capacity_frac_of_peak": 0.125 }, { "policy": "lru", "capacity_tokens": 8421607.375, "max_hold_s": 3600.0, "park_after_s": null, "ttl_s": null, "calls": 59204, "hits": 32114, "misses": 26351, "first_touches": 739, "hit_rate_excluding_first": 0.549286, "prefill_tokens_recomputed": 3135350390, "prefill_tokens_first_touch": 68337940, "restored_tokens": 0, "released_by_ttl": 0, "hbm_token_seconds_by_class": { "tool_lt_1s": 0, "tool_1_6s": 3160535482, "tool_gt_6s": 80005510578, "human": 11717150911, "unknown": 0 }, "host_token_seconds_by_class": { "tool_lt_1s": 0, "tool_1_6s": 0, "tool_gt_6s": 0, "human": 0, "unknown": 0 }, "hbm_token_seconds": 94883196971, "peak_live_tokens": 8421605, "capacity_frac_of_peak": 0.0625 }, { "policy": "ttl", "capacity_tokens": Infinity, "max_hold_s": 3600.0, "park_after_s": null, "ttl_s": 300.0, "calls": 59204, "hits": 55484, "misses": 2981, "first_touches": 739, "hit_rate_excluding_first": 0.949012, "prefill_tokens_recomputed": 485226403, "prefill_tokens_first_touch": 68337940, "restored_tokens": 0, "released_by_ttl": 3719, "hbm_token_seconds_by_class": { "tool_lt_1s": 0, "tool_1_6s": 4782898424, "tool_gt_6s": 111177579769, "human": 20057466092, "unknown": 0 }, "host_token_seconds_by_class": { "tool_lt_1s": 0, "tool_1_6s": 0, "tool_gt_6s": 0, "human": 0, "unknown": 0 }, "hbm_token_seconds": 136017944285, "peak_live_tokens": 83875657, "capacity_frac_of_peak": null }, { "policy": "ttl", "capacity_tokens": Infinity, "max_hold_s": 3600.0, "park_after_s": null, "ttl_s": 3600.0, "calls": 59204, "hits": 58123, "misses": 342, "first_touches": 739, "hit_rate_excluding_first": 0.99415, "prefill_tokens_recomputed": 69685755, "prefill_tokens_first_touch": 68337940, "restored_tokens": 0, "released_by_ttl": 1080, "hbm_token_seconds_by_class": { "tool_lt_1s": 0, "tool_1_6s": 4871580724, "tool_gt_6s": 135743204278, "human": 35791385210, "unknown": 0 }, "host_token_seconds_by_class": { "tool_lt_1s": 0, "tool_1_6s": 0, "tool_gt_6s": 0, "human": 0, "unknown": 0 }, "hbm_token_seconds": 176406170212, "peak_live_tokens": 112700843, "capacity_frac_of_peak": null }, { "policy": "oracle_park", "capacity_tokens": 33686429.5, "max_hold_s": 3600.0, "park_after_s": 6.0, "ttl_s": null, "calls": 59204, "hits": 58465, "misses": 0, "first_touches": 739, "hit_rate_excluding_first": 1.0, "prefill_tokens_recomputed": 0, "prefill_tokens_first_touch": 68337940, "restored_tokens": 8118824114, "released_by_ttl": 0, "hbm_token_seconds_by_class": { "tool_lt_1s": 0, "tool_1_6s": 203046330, "tool_gt_6s": 113496207, "human": 74813268, "unknown": 0 }, "host_token_seconds_by_class": { "tool_lt_1s": 0, "tool_1_6s": 4673398457, "tool_gt_6s": 135878579789, "human": 45187073937, "unknown": 0 }, "hbm_token_seconds": 391355805, "peak_live_tokens": 33686407, "capacity_frac_of_peak": 0.25 }, { "policy": "oracle_park", "capacity_tokens": 33686429.5, "max_hold_s": 3600.0, "park_after_s": 60.0, "ttl_s": null, "calls": 59204, "hits": 57782, "misses": 683, "first_touches": 739, "hit_rate_excluding_first": 0.988318, "prefill_tokens_recomputed": 89755368, "prefill_tokens_first_touch": 68337940, "restored_tokens": 1556376561, "released_by_ttl": 0, "hbm_token_seconds_by_class": { "tool_lt_1s": 0, "tool_1_6s": 4392052382, "tool_gt_6s": 75539699764, "human": 8590238513, "unknown": 0 }, "host_token_seconds_by_class": { "tool_lt_1s": 0, "tool_1_6s": 472072484, "tool_gt_6s": 59715464294, "human": 36367217223, "unknown": 0 }, "hbm_token_seconds": 88521990659, "peak_live_tokens": 33686418, "capacity_frac_of_peak": 0.25 } ], "priced": { "Qwen/Qwen3-8B": { "lru_quarter_peak_1h": { "bytes_per_token": 147456, "hbm_gb_hours": 6305.302, "gpu_hours_memory_equivalent": 78.8163, "gpu_hours_recompute": 8.1886, "usd_memory": 359.4022, "usd_recompute": 37.34, "usd_total": 396.7422, "assumptions": { "usd_per_gpu_hour": 4.56, "hbm_gb_per_gpu": 80.0, "prefill_tokens_per_gpu_second": 20000.0 } }, "oracle_park_6s_quarter_peak": { "bytes_per_token": 147456, "hbm_gb_hours": 16.03, "gpu_hours_memory_equivalent": 0.2004, "gpu_hours_recompute": 0.0, "usd_memory": 0.9137, "usd_recompute": 0.0, "usd_total": 0.9137, "assumptions": { "usd_per_gpu_hour": 4.56, "hbm_gb_per_gpu": 80.0, "prefill_tokens_per_gpu_second": 20000.0 } }, "unbounded_1h": { "bytes_per_token": 147456, "hbm_gb_hours": 7225.597, "gpu_hours_memory_equivalent": 90.32, "gpu_hours_recompute": 0.9679, "usd_memory": 411.859, "usd_recompute": 4.4134, "usd_total": 416.2724, "assumptions": { "usd_per_gpu_hour": 4.56, "hbm_gb_per_gpu": 80.0, "prefill_tokens_per_gpu_second": 20000.0 } } }, "Qwen/Qwen3.8-27B": { "lru_quarter_peak_1h": { "bytes_per_token": 65536, "hbm_gb_hours": 2802.357, "gpu_hours_memory_equivalent": 35.0295, "gpu_hours_recompute": 8.1886, "usd_memory": 159.7343, "usd_recompute": 37.34, "usd_total": 197.0743, "assumptions": { "usd_per_gpu_hour": 4.56, "hbm_gb_per_gpu": 80.0, "prefill_tokens_per_gpu_second": 20000.0 } }, "oracle_park_6s_quarter_peak": { "bytes_per_token": 65536, "hbm_gb_hours": 7.124, "gpu_hours_memory_equivalent": 0.0891, "gpu_hours_recompute": 0.0, "usd_memory": 0.4061, "usd_recompute": 0.0, "usd_total": 0.4061, "assumptions": { "usd_per_gpu_hour": 4.56, "hbm_gb_per_gpu": 80.0, "prefill_tokens_per_gpu_second": 20000.0 } }, "unbounded_1h": { "bytes_per_token": 65536, "hbm_gb_hours": 3211.376, "gpu_hours_memory_equivalent": 40.1422, "gpu_hours_recompute": 0.9679, "usd_memory": 183.0485, "usd_recompute": 4.4134, "usd_total": 187.4619, "assumptions": { "usd_per_gpu_hour": 4.56, "hbm_gb_per_gpu": 80.0, "prefill_tokens_per_gpu_second": 20000.0 } } }, "Qwen/Qwen3.8-Flash-Next": { "lru_quarter_peak_1h": { "bytes_per_token": 24576, "hbm_gb_hours": 1050.884, "gpu_hours_memory_equivalent": 13.136, "gpu_hours_recompute": 8.1886, "usd_memory": 59.9004, "usd_recompute": 37.34, "usd_total": 97.2404, "assumptions": { "usd_per_gpu_hour": 4.56, "hbm_gb_per_gpu": 80.0, "prefill_tokens_per_gpu_second": 20000.0 } }, "oracle_park_6s_quarter_peak": { "bytes_per_token": 24576, "hbm_gb_hours": 2.672, "gpu_hours_memory_equivalent": 0.0334, "gpu_hours_recompute": 0.0, "usd_memory": 0.1523, "usd_recompute": 0.0, "usd_total": 0.1523, "assumptions": { "usd_per_gpu_hour": 4.56, "hbm_gb_per_gpu": 80.0, "prefill_tokens_per_gpu_second": 20000.0 } }, "unbounded_1h": { "bytes_per_token": 24576, "hbm_gb_hours": 1204.266, "gpu_hours_memory_equivalent": 15.0533, "gpu_hours_recompute": 0.9679, "usd_memory": 68.6432, "usd_recompute": 4.4134, "usd_total": 73.0566, "assumptions": { "usd_per_gpu_hour": 4.56, "hbm_gb_per_gpu": 80.0, "prefill_tokens_per_gpu_second": 20000.0 } } } }, "pricing_assumptions": { "usd_per_gpu_hour": 4.56, "hbm_gb_per_gpu": 80.0, "prefill_tokens_per_gpu_second": 20000.0 }, "verdict": "On 59,204 real agent calls the HBM held while sessions wait is 176,406,170,212 token-seconds (unbounded pool, 1 h hold); 20% of it is human-paced waits and 77% tool waits over 6 s -- 97% of the hold cost sits in waits a gateway can see coming from finish_reason and the tool name. At a quarter of peak capacity LRU-by-session recomputes 589,578,606 prefill tokens (hit rate 0.922); oracle parking after 6 s at the same capacity recomputes 0 and restores 8,118,824,114 from host (hit rate 1.000).", "does_not_prove": [ "One operator's traffic (research-heavy, high effort). TraceLab is the comparison.", "Sessions, not blocks: intra-session prefix sharing across sibling subagents is counted as separate KV (an upper bound on hold cost, lower bound on sharing).", "The prefill rate and GPU price are ASSUMPTIONS stated in `pricing_assumptions`; the memory side is measured bytes/token times measured token-seconds, the recompute side is not a measurement (C11).", "Oracle park knows the coming gap. A real gateway predicts it from finish_reason and the tool name; R4 measures how much of the oracle a proxy can see.", "KV for output tokens is not added to the session's live size; prompts dominate (median prompt 447k tokens vs median output 595)." ], "generated_utc": "2026-09-02T22:52:32.810804+00:00", "git_rev": "575776c", "git_dirty": true, "run_id": "9e715ff79ee94f19", "host_kind": "Darwin-arm64", "provenance_schema": "gpu-cert-provenance/1" }