{
  "artifact": "s06_bailian_fleet_audit",
  "sprint": "S06-C2",
  "claim": "What a GPU-only cache captures of the reuse in Alibaba Bailian's four public Qwen traces, and what a host tier would add, at the capacity S46 ran; plus what replica scale-out costs and session affinity recovers.",
  "method": "tools/fleet_report.build_report and routing_loss over each file's salted SipHash block ids (16 tokens per block) at 65,536 tokens; session = root of the parent_chat_id chain; controls run first.",
  "prereg": "results/s06_bailian_fleet_audit.prereg.json",
  "capacity_tokens": 65536,
  "block_tokens": 16,
  "positive_control": {
    "trace": "mooncake_timed_600",
    "headroom_pp": 13.004,
    "expected": 13.0039,
    "ok": true
  },
  "negative_control": {
    "file": "qwen_thinking_blksz_16",
    "fresh_ids": true,
    "reuse_fraction": 0.0,
    "report": "OK",
    "refused": true
  },
  "per_file": {
    "qwen_traceA_blksz_16": {
      "report": "OK",
      "reuse_fraction": 0.597975,
      "n_requests": 43058,
      "n_sessions": 23101,
      "n_blocks": 6293976,
      "span_s": 7199.5,
      "max_turn": 37,
      "source_sha256": "07cedc9ed8aff301994ac68ed4aede8123b7603673575eeba9dd677de663db17",
      "reasons": null,
      "cold_miss_floor": 0.402025,
      "max_achievable_hit_rate": 0.597975,
      "gpu_only_lru_hit_rate": 0.152303,
      "offload_headroom_pp": 44.5672,
      "gpu_share_of_achievable": 0.254698,
      "working_set_over_capacity": 617.758,
      "n_block_refs": 6293976,
      "n_distinct_blocks": 2530337,
      "passes_s50_gate_12pp": true,
      "scale_out": {
        "single_replica": 0.152303,
        "round_robin_8": 0.090695,
        "session_affinity_8": 0.101443,
        "loss_pp_round_robin_8": 6.161,
        "recovery_pp_session_affinity_8": 1.075,
        "by_replicas": {
          "2": {
            "round_robin": 0.136809,
            "session_affinity": 0.144439,
            "tenant_affinity": 0.143807,
            "loss_pp_round_robin": 1.5494,
            "recovery_pp_session_affinity": 0.763,
            "recovery_pp_tenant_affinity": 0.6998
          },
          "4": {
            "round_robin": 0.119171,
            "session_affinity": 0.128427,
            "tenant_affinity": 0.128639,
            "loss_pp_round_robin": 3.3132,
            "recovery_pp_session_affinity": 0.9256,
            "recovery_pp_tenant_affinity": 0.9468
          },
          "8": {
            "round_robin": 0.090695,
            "session_affinity": 0.101443,
            "tenant_affinity": 0.1006,
            "loss_pp_round_robin": 6.1609,
            "recovery_pp_session_affinity": 1.0748,
            "recovery_pp_tenant_affinity": 0.9906
          }
        }
      },
      "fleet_report_exit": 0
    },
    "qwen_traceB_blksz_16": {
      "report": "OK",
      "reuse_fraction": 0.615273,
      "n_requests": 172800,
      "n_sessions": 172800,
      "n_blocks": 9956137,
      "span_s": 7200.0,
      "max_turn": 1,
      "source_sha256": "68e3f98e2d601d60d0abf4b89bc8a3654372abab7b1cde6373a13d0054379d59",
      "reasons": null,
      "cold_miss_floor": 0.384727,
      "max_achievable_hit_rate": 0.615273,
      "gpu_only_lru_hit_rate": 0.305622,
      "offload_headroom_pp": 30.9651,
      "gpu_share_of_achievable": 0.496726,
      "working_set_over_capacity": 935.154,
      "n_block_refs": 9956137,
      "n_distinct_blocks": 3830390,
      "passes_s50_gate_12pp": true,
      "scale_out": {
        "single_replica": 0.305622,
        "round_robin_8": 0.153802,
        "session_affinity_8": 0.154451,
        "loss_pp_round_robin_8": 15.182,
        "recovery_pp_session_affinity_8": 0.065,
        "by_replicas": {
          "2": {
            "round_robin": 0.242458,
            "session_affinity": 0.244069,
            "tenant_affinity": 0.247837,
            "loss_pp_round_robin": 6.3163,
            "recovery_pp_session_affinity": 0.1611,
            "recovery_pp_tenant_affinity": 0.5378
          },
          "4": {
            "round_robin": 0.193829,
            "session_affinity": 0.19556,
            "tenant_affinity": 0.202973,
            "loss_pp_round_robin": 11.1793,
            "recovery_pp_session_affinity": 0.1731,
            "recovery_pp_tenant_affinity": 0.9145
          },
          "8": {
            "round_robin": 0.153802,
            "session_affinity": 0.154451,
            "tenant_affinity": 0.163208,
            "loss_pp_round_robin": 15.1819,
            "recovery_pp_session_affinity": 0.0649,
            "recovery_pp_tenant_affinity": 0.9406
          }
        }
      },
      "fleet_report_exit": 0
    },
    "qwen_thinking_blksz_16": {
      "report": "OK",
      "reuse_fraction": 0.461842,
      "n_requests": 10812,
      "n_sessions": 9612,
      "n_blocks": 3223990,
      "span_s": 7191.7,
      "max_turn": 16,
      "source_sha256": "41ac36d9d1b54d084eeb6b05e9d142ba9d02f8bf79cfc9fff418d8ab7cdee906",
      "reasons": null,
      "cold_miss_floor": 0.538158,
      "max_achievable_hit_rate": 0.461842,
      "gpu_only_lru_hit_rate": 0.25042,
      "offload_headroom_pp": 21.1422,
      "gpu_share_of_achievable": 0.54222,
      "working_set_over_capacity": 423.588,
      "n_block_refs": 3223990,
      "n_distinct_blocks": 1735015,
      "passes_s50_gate_12pp": true,
      "scale_out": {
        "single_replica": 0.25042,
        "round_robin_8": 0.100471,
        "session_affinity_8": 0.119674,
        "loss_pp_round_robin_8": 14.995,
        "recovery_pp_session_affinity_8": 1.92,
        "by_replicas": {
          "2": {
            "round_robin": 0.205043,
            "session_affinity": 0.207173,
            "tenant_affinity": 0.191337,
            "loss_pp_round_robin": 4.5378,
            "recovery_pp_session_affinity": 0.2131,
            "recovery_pp_tenant_affinity": -1.3705
          },
          "4": {
            "round_robin": 0.156526,
            "session_affinity": 0.154118,
            "tenant_affinity": 0.139512,
            "loss_pp_round_robin": 9.3895,
            "recovery_pp_session_affinity": -0.2408,
            "recovery_pp_tenant_affinity": -1.7014
          },
          "8": {
            "round_robin": 0.100471,
            "session_affinity": 0.119674,
            "tenant_affinity": 0.115362,
            "loss_pp_round_robin": 14.9949,
            "recovery_pp_session_affinity": 1.9203,
            "recovery_pp_tenant_affinity": 1.489
          }
        }
      },
      "fleet_report_exit": 0
    },
    "qwen_coder_blksz_16": {
      "report": "OK",
      "reuse_fraction": 0.663723,
      "n_requests": 43011,
      "n_sessions": 26406,
      "n_blocks": 15470907,
      "span_s": 7198.9,
      "max_turn": 54,
      "source_sha256": "3d74974cebb7bccb69b1ca6ea210a44d9a3033443c0271ed015327284dea1aff",
      "reasons": null,
      "cold_miss_floor": 0.336277,
      "max_achievable_hit_rate": 0.663723,
      "gpu_only_lru_hit_rate": 0.147633,
      "offload_headroom_pp": 51.609,
      "gpu_share_of_achievable": 0.222432,
      "working_set_over_capacity": 1270.146,
      "n_block_refs": 15470907,
      "n_distinct_blocks": 5202517,
      "passes_s50_gate_12pp": true,
      "scale_out": {
        "single_replica": 0.147633,
        "round_robin_8": 0.025134,
        "session_affinity_8": 0.006353,
        "loss_pp_round_robin_8": 12.25,
        "recovery_pp_session_affinity_8": -1.878,
        "by_replicas": {
          "2": {
            "round_robin": 0.083684,
            "session_affinity": 0.085123,
            "tenant_affinity": 0.06654,
            "loss_pp_round_robin": 6.3949,
            "recovery_pp_session_affinity": 0.1439,
            "recovery_pp_tenant_affinity": -1.7144
          },
          "4": {
            "round_robin": 0.045276,
            "session_affinity": 0.041853,
            "tenant_affinity": 0.029119,
            "loss_pp_round_robin": 10.2357,
            "recovery_pp_session_affinity": -0.3423,
            "recovery_pp_tenant_affinity": -1.6157
          },
          "8": {
            "round_robin": 0.025134,
            "session_affinity": 0.006353,
            "tenant_affinity": 0.004378,
            "loss_pp_round_robin": 12.2499,
            "recovery_pp_session_affinity": -1.8781,
            "recovery_pp_tenant_affinity": -2.0756
          }
        }
      },
      "fleet_report_exit": 0
    }
  },
  "files_passing_gate": [
    "qwen_traceA_blksz_16",
    "qwen_traceB_blksz_16",
    "qwen_thinking_blksz_16",
    "qwen_coder_blksz_16"
  ],
  "verdict": "BAILIAN, FOUR PUBLIC QWEN TRACES AT 65,536 TOKENS: qwen_traceA_blksz_16 floor 0.402025 / GPU-only LRU 0.152303 / ceiling 0.597975 / headroom 44.5672 pp / 8-replica round-robin loses 6.161 pp, session affinity recovers 1.075; qwen_traceB_blksz_16 floor 0.384727 / GPU-only LRU 0.305622 / ceiling 0.615273 / headroom 30.9651 pp / 8-replica round-robin loses 15.182 pp, session affinity recovers 0.065; qwen_thinking_blksz_16 floor 0.538158 / GPU-only LRU 0.25042 / ceiling 0.461842 / headroom 21.1422 pp / 8-replica round-robin loses 14.995 pp, session affinity recovers 1.92; qwen_coder_blksz_16 floor 0.336277 / GPU-only LRU 0.147633 / ceiling 0.663723 / headroom 51.609 pp / 8-replica round-robin loses 12.25 pp, session affinity recovers -1.878. Files clearing the S50 gate (>= 12 pp): ['qwen_traceA_blksz_16', 'qwen_traceB_blksz_16', 'qwen_thinking_blksz_16', 'qwen_coder_blksz_16']. Controls: positive OK, negative REFUSED.",
  "wall_s": 544.2,
  "cost_usd": 0.0,
  "does_not_prove": [
    "Block hits, not GPU-hours: converting needs the bandwidth-bound fraction of GPU time, never measured here.",
    "LRU stands in for the engine's prefix tree; S46b corroborated the stand-in on Mooncake (0.0376 vs 0.0387), not on these traces.",
    "Sessions are reconstructed from parent_chat_id; the traces carry no tenant id, so tenant affinity is not measurable.",
    "Two hours per trace, December 2024, anonymised by the provider; the salting is theirs."
  ],
  "generated_utc": "2026-09-02T23:06:59.356290+00:00",
  "git_rev": "093ddaf",
  "git_dirty": true,
  "run_id": "295b3556aae94757",
  "host_kind": "Darwin-arm64",
  "provenance_schema": "gpu-cert-provenance/1"
}