{
  "date": "2026-11-04",
  "method": "全部同节点 claim 对作词项/汉字二元组重合筛查；每节点取分数≥0.18的前2对，再逐项语义裁决。不是全库语义无矛盾证明。",
  "screened_pairs": {
    "agent-system": 751,
    "cluster-infra": 2879,
    "distributed-training": 298,
    "inference-system": 1960,
    "kvcache": 1932,
    "recommendation-system": 54
  },
  "manual_pairs": [
    {
      "domain": "agent-system",
      "a": "agent-system:agent-infra-kv-010",
      "b": "agent-system:agent-infra-kv-011",
      "screen_score": 0.271,
      "verdict": "同源推演",
      "reason": "理想缓存下的时间分解和容量越界解释不同，旧 I/O 比率不能直接当重算增幅。"
    },
    {
      "domain": "agent-system",
      "a": "agent-system:agent-infra-kv-007",
      "b": "agent-system:agent-infra-kv-015",
      "screen_score": 0.19,
      "verdict": "互补/正交",
      "reason": "模型、硬件、控制粒度或指标范围不同；未见同前提下相反结论。"
    },
    {
      "domain": "agent-system",
      "a": "agent-system:agent-memory-002",
      "b": "agent-system:agent-memory-009",
      "screen_score": 0.279,
      "verdict": "互补/正交",
      "reason": "模型、硬件、控制粒度或指标范围不同；未见同前提下相反结论。"
    },
    {
      "domain": "agent-system",
      "a": "agent-system:agent-memory-001",
      "b": "agent-system:agent-memory-009",
      "screen_score": 0.224,
      "verdict": "互补/正交",
      "reason": "模型、硬件、控制粒度或指标范围不同；未见同前提下相反结论。"
    },
    {
      "domain": "agent-system",
      "a": "agent-system:skill-management-005",
      "b": "agent-system:skill-management-007",
      "screen_score": 0.193,
      "verdict": "层级互补",
      "reason": "实现比例与渐进披露接口是不同对象；005 过期，不再支撑当前立场。"
    },
    {
      "domain": "cluster-infra",
      "a": "cluster-infra:compute-storage-coupling-001",
      "b": "cluster-infra:compute-storage-coupling-002",
      "screen_score": 0.186,
      "verdict": "互补/正交",
      "reason": "同源研究的现象、机制或适用边界不同；不重复计为独立验证。"
    },
    {
      "domain": "cluster-infra",
      "a": "cluster-infra:data-movement-006",
      "b": "cluster-infra:data-movement-007",
      "screen_score": 0.241,
      "verdict": "互补/正交",
      "reason": "模型、硬件、控制粒度或指标范围不同；未见同前提下相反结论。"
    },
    {
      "domain": "cluster-infra",
      "a": "cluster-infra:hardware-generations-009",
      "b": "cluster-infra:hardware-generations-011",
      "screen_score": 0.2,
      "verdict": "互补/正交",
      "reason": "模型、硬件、控制粒度或指标范围不同；未见同前提下相反结论。"
    },
    {
      "domain": "cluster-infra",
      "a": "cluster-infra:hardware-generations-001",
      "b": "cluster-infra:hardware-generations-005",
      "screen_score": 0.188,
      "verdict": "互补/正交",
      "reason": "模型、硬件、控制粒度或指标范围不同；未见同前提下相反结论。"
    },
    {
      "domain": "cluster-infra",
      "a": "cluster-infra:heterogeneous-cluster-003",
      "b": "cluster-infra:heterogeneous-cluster-005",
      "screen_score": 0.234,
      "verdict": "互补/正交",
      "reason": "模型、硬件、控制粒度或指标范围不同；未见同前提下相反结论。"
    },
    {
      "domain": "cluster-infra",
      "a": "cluster-infra:heterogeneous-cluster-001",
      "b": "cluster-infra:heterogeneous-cluster-003",
      "screen_score": 0.191,
      "verdict": "互补/正交",
      "reason": "同源研究的现象、机制或适用边界不同；不重复计为独立验证。"
    },
    {
      "domain": "cluster-infra",
      "a": "cluster-infra:memory-fabric-010",
      "b": "cluster-infra:memory-fabric-011",
      "screen_score": 0.229,
      "verdict": "互补/正交",
      "reason": "同源研究的现象、机制或适用边界不同；不重复计为独立验证。"
    },
    {
      "domain": "cluster-infra",
      "a": "cluster-infra:scale-up-interconnect-014",
      "b": "cluster-infra:scale-up-interconnect-020",
      "screen_score": 0.229,
      "verdict": "互补/正交",
      "reason": "同源研究的现象、机制或适用边界不同；不重复计为独立验证。"
    },
    {
      "domain": "cluster-infra",
      "a": "cluster-infra:scale-up-interconnect-014",
      "b": "cluster-infra:scale-up-interconnect-016",
      "screen_score": 0.208,
      "verdict": "互补/正交",
      "reason": "模型、硬件、控制粒度或指标范围不同；未见同前提下相反结论。"
    },
    {
      "domain": "cluster-infra",
      "a": "cluster-infra:supernode-topology-001",
      "b": "cluster-infra:supernode-topology-018",
      "screen_score": 0.207,
      "verdict": "互补/正交",
      "reason": "模型、硬件、控制粒度或指标范围不同；未见同前提下相反结论。"
    },
    {
      "domain": "cluster-infra",
      "a": "cluster-infra:supernode-topology-009",
      "b": "cluster-infra:supernode-topology-016",
      "screen_score": 0.198,
      "verdict": "互补/正交",
      "reason": "模型、硬件、控制粒度或指标范围不同；未见同前提下相反结论。"
    },
    {
      "domain": "distributed-training",
      "a": "distributed-training:async-distributed-001",
      "b": "distributed-training:async-distributed-003",
      "screen_score": 0.191,
      "verdict": "互补/正交",
      "reason": "同源研究的现象、机制或适用边界不同；不重复计为独立验证。"
    },
    {
      "domain": "distributed-training",
      "a": "distributed-training:collective-communication-003",
      "b": "distributed-training:collective-communication-005",
      "screen_score": 0.227,
      "verdict": "互补/正交",
      "reason": "同源研究的现象、机制或适用边界不同；不重复计为独立验证。"
    },
    {
      "domain": "distributed-training",
      "a": "distributed-training:long-context-training-001",
      "b": "distributed-training:long-context-training-002",
      "screen_score": 0.224,
      "verdict": "互补/正交",
      "reason": "同源研究的现象、机制或适用边界不同；不重复计为独立验证。"
    },
    {
      "domain": "inference-system",
      "a": "inference-system:engine-architecture-013",
      "b": "inference-system:engine-architecture-014",
      "screen_score": 0.255,
      "verdict": "互补/正交",
      "reason": "模型、硬件、控制粒度或指标范围不同；未见同前提下相反结论。"
    },
    {
      "domain": "inference-system",
      "a": "inference-system:engine-architecture-014",
      "b": "inference-system:engine-architecture-015",
      "screen_score": 0.191,
      "verdict": "互补/正交",
      "reason": "模型、硬件、控制粒度或指标范围不同；未见同前提下相反结论。"
    },
    {
      "domain": "inference-system",
      "a": "inference-system:hardware-specific-010",
      "b": "inference-system:hardware-specific-018",
      "screen_score": 0.222,
      "verdict": "互补/正交",
      "reason": "模型、硬件、控制粒度或指标范围不同；未见同前提下相反结论。"
    },
    {
      "domain": "inference-system",
      "a": "inference-system:inference-parallelism-003",
      "b": "inference-system:inference-parallelism-005",
      "screen_score": 0.214,
      "verdict": "互补/正交",
      "reason": "模型、硬件、控制粒度或指标范围不同；未见同前提下相反结论。"
    },
    {
      "domain": "inference-system",
      "a": "inference-system:long-context-serving-008",
      "b": "inference-system:long-context-serving-011",
      "screen_score": 0.205,
      "verdict": "互补/正交",
      "reason": "模型、硬件、控制粒度或指标范围不同；未见同前提下相反结论。"
    },
    {
      "domain": "inference-system",
      "a": "inference-system:model-quantization-001",
      "b": "inference-system:model-quantization-002",
      "screen_score": 0.231,
      "verdict": "互补/正交",
      "reason": "模型、硬件、控制粒度或指标范围不同；未见同前提下相反结论。"
    },
    {
      "domain": "inference-system",
      "a": "inference-system:model-quantization-003",
      "b": "inference-system:model-quantization-004",
      "screen_score": 0.219,
      "verdict": "互补/正交",
      "reason": "同源研究的现象、机制或适用边界不同；不重复计为独立验证。"
    },
    {
      "domain": "inference-system",
      "a": "inference-system:moe-serving-006",
      "b": "inference-system:moe-serving-010",
      "screen_score": 0.283,
      "verdict": "条件互补",
      "reason": "训练均衡目标与生产路由分布不是同一证据，K3 官方缺生产负载验证。"
    },
    {
      "domain": "inference-system",
      "a": "inference-system:moe-serving-007",
      "b": "inference-system:moe-serving-010",
      "screen_score": 0.212,
      "verdict": "条件互补",
      "reason": "MiMo 的部署观察不能替 K3 证明训练分布与生产分布一致。"
    },
    {
      "domain": "inference-system",
      "a": "inference-system:pd-disaggregation-002",
      "b": "inference-system:pd-disaggregation-013",
      "screen_score": 0.208,
      "verdict": "口径互补",
      "reason": "raw throughput 与 SLO goodput 要分别比较，无法据此裁决所有 MoE 的 PD 收益。"
    },
    {
      "domain": "inference-system",
      "a": "inference-system:request-scheduling-007",
      "b": "inference-system:request-scheduling-016",
      "screen_score": 0.234,
      "verdict": "互补/正交",
      "reason": "模型、硬件、控制粒度或指标范围不同；未见同前提下相反结论。"
    },
    {
      "domain": "inference-system",
      "a": "inference-system:request-scheduling-001",
      "b": "inference-system:request-scheduling-011",
      "screen_score": 0.2,
      "verdict": "互补/正交",
      "reason": "模型、硬件、控制粒度或指标范围不同；未见同前提下相反结论。"
    },
    {
      "domain": "inference-system",
      "a": "inference-system:serving-infrastructure-001",
      "b": "inference-system:serving-infrastructure-012",
      "screen_score": 0.182,
      "verdict": "互补/正交",
      "reason": "模型、硬件、控制粒度或指标范围不同；未见同前提下相反结论。"
    },
    {
      "domain": "inference-system",
      "a": "inference-system:speculative-decoding-002",
      "b": "inference-system:speculative-decoding-013",
      "screen_score": 0.485,
      "verdict": "条件互补",
      "reason": "TPU 特定验证成本曲线不能当作 H200 所有并发的实证；K-Flat 不泛化。"
    },
    {
      "domain": "inference-system",
      "a": "inference-system:speculative-decoding-006",
      "b": "inference-system:speculative-decoding-008",
      "screen_score": 0.216,
      "verdict": "互补/正交",
      "reason": "同源研究的现象、机制或适用边界不同；不重复计为独立验证。"
    },
    {
      "domain": "kvcache",
      "a": "kvcache:attention-variants-005",
      "b": "kvcache:attention-variants-010",
      "screen_score": 0.372,
      "verdict": "重复合并",
      "reason": "同一五模型 BF16/100K 容量表，005 指向 010，不能重复计算验证来源。"
    },
    {
      "domain": "kvcache",
      "a": "kvcache:attention-variants-007",
      "b": "kvcache:attention-variants-010",
      "screen_score": 0.281,
      "verdict": "互补/正交",
      "reason": "模型、硬件、控制粒度或指标范围不同；未见同前提下相反结论。"
    },
    {
      "domain": "kvcache",
      "a": "kvcache:compression-quantization-007",
      "b": "kvcache:compression-quantization-010",
      "screen_score": 0.382,
      "verdict": "对象纠错",
      "reason": "原 010 把 INT8 模型+FP8 KV 的退化归因到 INT8 KV；保留对照现象、撤回错位因果。"
    },
    {
      "domain": "kvcache",
      "a": "kvcache:compression-quantization-013",
      "b": "kvcache:compression-quantization-014",
      "screen_score": 0.279,
      "verdict": "互补/正交",
      "reason": "模型、硬件、控制粒度或指标范围不同；未见同前提下相反结论。"
    },
    {
      "domain": "kvcache",
      "a": "kvcache:framework-mooncake-006",
      "b": "kvcache:framework-mooncake-011",
      "screen_score": 0.276,
      "verdict": "历史状态替代",
      "reason": "07-12 独立服务落地，早期无代码结论只作快照。"
    },
    {
      "domain": "kvcache",
      "a": "kvcache:framework-mooncake-017",
      "b": "kvcache:framework-mooncake-018",
      "screen_score": 0.212,
      "verdict": "互补/正交",
      "reason": "同源研究的现象、机制或适用边界不同；不重复计为独立验证。"
    },
    {
      "domain": "kvcache",
      "a": "kvcache:offload-007",
      "b": "kvcache:offload-008",
      "screen_score": 0.258,
      "verdict": "互补/正交",
      "reason": "同源研究的现象、机制或适用边界不同；不重复计为独立验证。"
    },
    {
      "domain": "kvcache",
      "a": "kvcache:offload-018",
      "b": "kvcache:offload-019",
      "screen_score": 0.23,
      "verdict": "边界互补",
      "reason": "局部可定位与全局稠密证据任务不同，前者收益不能外推后者。"
    },
    {
      "domain": "kvcache",
      "a": "kvcache:paged-kv-001",
      "b": "kvcache:paged-kv-002",
      "screen_score": 0.182,
      "verdict": "互补/正交",
      "reason": "同源研究的现象、机制或适用边界不同；不重复计为独立验证。"
    },
    {
      "domain": "kvcache",
      "a": "kvcache:prefix-cache-014",
      "b": "kvcache:prefix-cache-016",
      "screen_score": 0.246,
      "verdict": "边界已解",
      "reason": "引擎保持数学等价与模型侧接受近似重放是不同权限边界。"
    },
    {
      "domain": "kvcache",
      "a": "kvcache:prefix-cache-002",
      "b": "kvcache:prefix-cache-003",
      "screen_score": 0.2,
      "verdict": "互补/正交",
      "reason": "同源研究的现象、机制或适用边界不同；不重复计为独立验证。"
    },
    {
      "domain": "kvcache",
      "a": "kvcache:sparsity-002",
      "b": "kvcache:sparsity-007",
      "screen_score": 0.321,
      "verdict": "重复合并",
      "reason": "相同 HiSparse 热 buffer/选择重叠机制归并到 007。"
    },
    {
      "domain": "kvcache",
      "a": "kvcache:sparsity-004",
      "b": "kvcache:sparsity-007",
      "screen_score": 0.309,
      "verdict": "重复合并",
      "reason": "适用边界已由 007 覆盖，保留历史 id。"
    },
    {
      "domain": "kvcache",
      "a": "kvcache:storage-hierarchy-010",
      "b": "kvcache:storage-hierarchy-016",
      "screen_score": 0.232,
      "verdict": "非独立验证",
      "reason": "HiSim 稿引用 Tutti 批评带宽模型，没有新测量，撤出 verified_by。"
    },
    {
      "domain": "kvcache",
      "a": "kvcache:storage-hierarchy-002",
      "b": "kvcache:storage-hierarchy-012",
      "screen_score": 0.185,
      "verdict": "互补/正交",
      "reason": "模型、硬件、控制粒度或指标范围不同；未见同前提下相反结论。"
    },
    {
      "domain": "agent-system",
      "a": "agent-system:agent-infra-kv-024",
      "b": "agent-system:agent-infra-kv-005",
      "verdict": "边界扩展",
      "reason": "固定 TTL/角色提示之外增加调用中的剩余时间信号，不推翻原生命周期机制。"
    },
    {
      "domain": "agent-system",
      "a": "agent-system:agent-infra-kv-025",
      "b": "agent-system:agent-infra-kv-011",
      "verdict": "一致",
      "reason": "不同实验观察容量越界后的命中陡降；百分比不交叉复用。"
    },
    {
      "domain": "agent-system",
      "a": "agent-system:code-agent-009",
      "b": "agent-system:code-agent-008",
      "verdict": "正交",
      "reason": "版本档案的探索规则不同于代码定位；低分祖先仍有搜索价值。"
    },
    {
      "domain": "agent-system",
      "a": "agent-system:code-agent-010",
      "b": "agent-system:harness-architecture-012",
      "verdict": "边界互补",
      "reason": "两种迁移机制不同，但都要求按模型和任务报告迁移收益。"
    },
    {
      "domain": "agent-system",
      "a": "agent-system:harness-architecture-016",
      "b": "agent-system:harness-architecture-015",
      "verdict": "互补",
      "reason": "追加式事实日志与工具状态旁路分别解决恢复和状态可见性。"
    },
    {
      "domain": "cluster-infra",
      "a": "cluster-infra:data-movement-023",
      "b": "cluster-infra:data-movement-001",
      "verdict": "边界扩展",
      "reason": "统一传输 API 不等于共享注册内存的所有权；零拷贝仍有容量成本。"
    },
    {
      "domain": "cluster-infra",
      "a": "cluster-infra:data-movement-024",
      "b": "cluster-infra:data-movement-016",
      "verdict": "一致",
      "reason": "buffer 可达、设备内整理和跨层就绪必须分开判定。"
    },
    {
      "domain": "cluster-infra",
      "a": "cluster-infra:memory-fabric-022",
      "b": "cluster-infra:memory-fabric-006",
      "verdict": "条件不同",
      "reason": "GH200 主机并发供带宽与 CXL warm tier 的延迟窗口不同，不能互相否定或外推。"
    },
    {
      "domain": "cluster-infra",
      "a": "cluster-infra:network-reliability-012",
      "b": "cluster-infra:network-reliability-003",
      "verdict": "互补",
      "reason": "路径恢复与剩余容量分别限制作业吞吐；快收敛不等于无损。"
    },
    {
      "domain": "cluster-infra",
      "a": "cluster-infra:network-reliability-013",
      "b": "cluster-infra:network-reliability-011",
      "verdict": "一致",
      "reason": "静态健康遗漏主机侧负载故障，须使用真实通信路径做主动验收。"
    },
    {
      "domain": "cluster-infra",
      "a": "cluster-infra:rack-scale-scheduling-007",
      "b": "cluster-infra:rack-scale-scheduling-001",
      "verdict": "互补",
      "reason": "拓扑合同为域感知调度提供持续更新的输入，不自带性能保证。"
    },
    {
      "domain": "cluster-infra",
      "a": "cluster-infra:scale-out-network-016",
      "b": "cluster-infra:scale-out-network-015",
      "verdict": "正交",
      "reason": "按 plane 选路与网卡选择 QP 属于不同控制粒度。"
    },
    {
      "domain": "cluster-infra",
      "a": "cluster-infra:scale-out-network-017",
      "b": "cluster-infra:network-reliability-012",
      "verdict": "互补",
      "reason": "分层拥塞控制解释异常平面的隔离，但容量降级仍存在。"
    },
    {
      "domain": "cluster-infra",
      "a": "cluster-infra:scale-out-network-018",
      "b": "cluster-infra:scale-out-network-017",
      "verdict": "条件互补",
      "reason": "PS incast 需应用在途量限制；多平面硬件控制不能替代任意负载的软件预算。"
    },
    {
      "domain": "distributed-training",
      "a": "distributed-training:collective-communication-014",
      "b": "distributed-training:training-parallelism-012",
      "verdict": "同源互补",
      "reason": "ETH 的梯度对齐与 local batch 是同一方案的不同组件，不构成独立互证。"
    },
    {
      "domain": "distributed-training",
      "a": "distributed-training:correctness-verification-005",
      "b": "distributed-training:correctness-verification-003",
      "verdict": "一致",
      "reason": "真实服务 E2E 补足单算子、融合与分布式冒烟的组合盲点。"
    },
    {
      "domain": "distributed-training",
      "a": "distributed-training:long-context-training-004",
      "b": "distributed-training:long-context-training-002",
      "verdict": "正交",
      "reason": "选择重放减重计算，非对称 CP/HP 减通信等待，二者不能直接相乘。"
    },
    {
      "domain": "distributed-training",
      "a": "distributed-training:long-context-training-005",
      "b": "distributed-training:long-context-training-004",
      "verdict": "条件互补",
      "reason": "首次前向仍全候选打分，O(k) 只覆盖保存索引后的选中集合重计算。"
    },
    {
      "domain": "distributed-training",
      "a": "distributed-training:long-context-training-006",
      "b": "distributed-training:long-context-training-004",
      "verdict": "限定",
      "reason": "组件耗时下降与完整模型短程实验、长期收敛分别计量。"
    },
    {
      "domain": "distributed-training",
      "a": "distributed-training:rl-infrastructure-008",
      "b": "distributed-training:rl-infrastructure-007",
      "verdict": "一致",
      "reason": "独立分池样本支持服务边界，但软件多轴变更不能归因 DeepEP 单项。"
    },
    {
      "domain": "distributed-training",
      "a": "distributed-training:training-parallelism-012",
      "b": "distributed-training:training-parallelism-008",
      "verdict": "边界扩展",
      "reason": "在层数与 micro-batch 时序之外，local batch 增加负载调节自由度。"
    },
    {
      "domain": "inference-system",
      "a": "inference-system:engine-architecture-022",
      "b": "inference-system:kernel-optimization-014",
      "verdict": "同源互补",
      "reason": "入口/回退决定局部 kernel 优化是否进入完整请求。"
    },
    {
      "domain": "inference-system",
      "a": "inference-system:hardware-specific-019",
      "b": "inference-system:hardware-specific-008",
      "verdict": "边界扩展",
      "reason": "机密模式进一步改变内存与计时假设，不仅是 GPU 架构和软件版本。"
    },
    {
      "domain": "inference-system",
      "a": "inference-system:kernel-optimization-014",
      "b": "inference-system:kernel-optimization-006",
      "verdict": "条件互补",
      "reason": "模型专用融合路径与通用 primitive 分工，局部时间线不是全请求加速。"
    },
    {
      "domain": "inference-system",
      "a": "inference-system:model-quantization-008",
      "b": "inference-system:model-quantization-001",
      "verdict": "边界扩展",
      "reason": "格式支持须分别验收加载、打包执行和实际服务；不同运行时不互相外推。"
    },
    {
      "domain": "inference-system",
      "a": "inference-system:moe-serving-015",
      "b": "inference-system:moe-serving-001",
      "verdict": "条件不同",
      "reason": "DeepInfer 的重启重配与 Elastic EP 在线拓扑切换是不同实现；001 过期不再支撑当前立场。"
    },
    {
      "domain": "inference-system",
      "a": "inference-system:request-scheduling-021",
      "b": "kvcache:prefix-cache-017",
      "verdict": "互补",
      "reason": "容量边界内看排队/到达模式，越界后看命中与重算；不声称跨 runtime 归因已完整实证。"
    },
    {
      "domain": "kvcache",
      "a": "kvcache:compression-quantization-019",
      "b": "kvcache:compression-quantization-013",
      "verdict": "口径限定",
      "reason": "裸 4-bit payload 减半与含 block scale 的 56.25% footprint 不同；固定并发和容量放大须分列。"
    },
    {
      "domain": "kvcache",
      "a": "kvcache:evaluation-benchmarks-008",
      "b": "kvcache:evaluation-benchmarks-006",
      "verdict": "互补",
      "reason": "实验坐标矩阵补足回放负载，不能把系统指标当任务质量。"
    },
    {
      "domain": "kvcache",
      "a": "kvcache:lifecycle-005",
      "b": "kvcache:lifecycle-004",
      "verdict": "边界扩展",
      "reason": "租户进度是 advisory signal，不能绕过已有状态所有权和提交点。"
    },
    {
      "domain": "kvcache",
      "a": "kvcache:offload-025",
      "b": "kvcache:offload-024",
      "verdict": "条件不同",
      "reason": "GH200 直接供数与缺少异步搬运原语的 PCIe 工作点不同。"
    },
    {
      "domain": "kvcache",
      "a": "kvcache:prefix-cache-018",
      "b": "kvcache:prefix-cache-012",
      "verdict": "条件不同",
      "reason": "API eligible breakpoint 与引擎物理 block 都限制上界，但不能推断二者底层同构。"
    },
    {
      "domain": "kvcache",
      "a": "kvcache:prefix-cache-019",
      "b": "kvcache:prefix-cache-008",
      "verdict": "正交",
      "reason": "供应商 API 计费准入与自部署缓存 TTL 是不同控制面。"
    },
    {
      "domain": "kvcache",
      "a": "kvcache:prefix-cache-020",
      "b": "kvcache:prefix-cache-006",
      "verdict": "限定",
      "reason": "单价示例只对应 10-10 来源快照，整账单降幅依赖读取费用占比；不当作当前价格。"
    },
    {
      "domain": "kvcache",
      "a": "kvcache:prefix-cache-017",
      "b": "agent-system:agent-infra-kv-011",
      "verdict": "一致",
      "reason": "容量阈值机制由 PrefixBench/NVFP4、CONCUR 与独立 workload 分解共同支持，收窄后升 high。"
    },
    {
      "domain": "kvcache",
      "a": "kvcache:storage-hierarchy-025",
      "b": "kvcache:pd-disaggregation-017",
      "verdict": "互补",
      "reason": "持久化范围、完整 handoff 状态和租约方向不是同一个字节量。"
    },
    {
      "domain": "kvcache",
      "a": "kvcache:storage-hierarchy-026",
      "b": "kvcache:storage-hierarchy-022",
      "verdict": "条件不同",
      "reason": "会话生命周期固定放置与有序预取窗口按不同控制自由度优化，不推翻彼此。"
    },
    {
      "domain": "kvcache",
      "a": "kvcache:storage-hierarchy-027",
      "b": "kvcache:storage-hierarchy-020",
      "verdict": "边界扩展",
      "reason": "L2 分配失败会使存储命中变为未交付命中，需分开计量；不能给 Tutti 的 CPU 税数字加一次验证。"
    },
    {
      "domain": "agent-system",
      "a": "agent-system:agentic-rl-002",
      "b": "distributed-training:rl-infrastructure-002",
      "verdict": "重复合并",
      "reason": "同一 RollArt Reward-as-a-Service 原始实验，保留训练域 canonical。"
    },
    {
      "domain": "distributed-training",
      "a": "distributed-training:multimodal-training-001",
      "b": "inference-system:multimodal-inference-001",
      "verdict": "重复合并",
      "reason": "同一 StreamWise serving 成本实验，归入推理域。"
    },
    {
      "domain": "distributed-training",
      "a": "distributed-training:multimodal-training-002",
      "b": "inference-system:multimodal-inference-002",
      "verdict": "重复合并",
      "reason": "同一 StreamWise serving 收益和消融，归入推理域。"
    },
    {
      "domain": "agent-system",
      "a": "agent-system:agent-infra-kv-009",
      "b": "kvcache:evaluation-benchmarks-006",
      "verdict": "历史状态替代",
      "reason": "AgentX 已填补回放负载发生器缺口，联合任务质量评测仍待补。"
    },
    {
      "domain": "distributed-training",
      "a": "distributed-training:large-scale-training-005",
      "b": "cluster-infra:scale-up-interconnect-043",
      "verdict": "口径纠错",
      "reason": "整柜聚合带宽与单端口混比已由同端点同方向约 36 倍口径替代。"
    },
    {
      "domain": "cluster-infra",
      "a": "cluster-infra:memory-fabric-023",
      "b": "cluster-infra:memory-fabric-016",
      "verdict": "条件不同",
      "reason": "CPU消费域与GPU后端消费域不同，不能以G2.5结果验证G1.5。"
    },
    {
      "domain": "cluster-infra",
      "a": "cluster-infra:memory-fabric-024",
      "b": "cluster-infra:memory-fabric-018",
      "verdict": "边界扩展",
      "reason": "单机扩容、静态多主机池、动态隔离和管理软件的成熟度分别验收。"
    },
    {
      "domain": "cluster-infra",
      "a": "cluster-infra:network-reliability-014",
      "b": "cluster-infra:network-reliability-013",
      "verdict": "互补",
      "reason": "主动验收不能替代维修前的设备归属与清场范围证明。"
    },
    {
      "domain": "cluster-infra",
      "a": "cluster-infra:network-reliability-015",
      "b": "cluster-infra:network-reliability-013",
      "verdict": "条件互补",
      "reason": "健康遥测与调度资格分离，故障风暴下还须区分隔离速率和隔离存量。"
    },
    {
      "domain": "cluster-infra",
      "a": "cluster-infra:network-reliability-016",
      "b": "cluster-infra:network-reliability-014",
      "verdict": "同源互补",
      "reason": "局部清场不授予整节点维修或全节点测试权限，规则与证明范围均须匹配。"
    },
    {
      "domain": "distributed-training",
      "a": "distributed-training:collective-communication-015",
      "b": "distributed-training:collective-communication-001",
      "verdict": "版本纠错，已关闭",
      "reason": "NCCL 2.32.3管理接口反驳必须销毁全部旧分布式状态的绝对表述，001降级并由015接替；应用状态仍需恢复。"
    },
    {
      "domain": "inference-system",
      "a": "inference-system:kernel-optimization-015",
      "b": "inference-system:hardware-specific-010",
      "verdict": "边界互补",
      "reason": "静态驻留上界与实测streaming分母不同，不能从occupancy直接推断加速。"
    },
    {
      "domain": "inference-system",
      "a": "inference-system:kernel-optimization-016",
      "b": "inference-system:kernel-optimization-015",
      "verdict": "同源限定",
      "reason": "L2内逻辑有效带宽与HBM物理流量不同，计时协议改变分母与缓存热度。"
    },
    {
      "domain": "inference-system",
      "a": "inference-system:kernel-optimization-017",
      "b": "inference-system:kernel-optimization-009",
      "verdict": "条件不同",
      "reason": "H20形状级cp.async与FastAFD跨step调度重叠不同，不存在通用双缓冲必赢结论。"
    },
    {
      "domain": "inference-system",
      "a": "inference-system:multimodal-inference-008",
      "b": "inference-system:multimodal-inference-007",
      "verdict": "边界互补",
      "reason": "紧凑中间态与模块分段都不自动缩短串行关键路径，动作闭环另测。"
    },
    {
      "domain": "kvcache",
      "a": "kvcache:compression-quantization-020",
      "b": "kvcache:compression-quantization-019",
      "verdict": "互补",
      "reason": "Ascend MTE请求粒度与NVFP4容量/带宽收益是不同硬件维度；未给出端到端量化倍数。"
    },
    {
      "domain": "kvcache",
      "a": "kvcache:storage-hierarchy-024",
      "b": "kvcache:prefix-cache-017",
      "verdict": "一致",
      "reason": "位置替换与减少重算分别度量，约6.7倍复用收益不能解释为CXL媒介速度。"
    },
    {
      "domain": "recommendation-system",
      "a": "recommendation-system:embedding-infrastructure-001",
      "b": "recommendation-system:embedding-infrastructure-002",
      "verdict": "同源互补/条件化",
      "reason": "训练状态完整性、缓存驻留与混合并行是不同要求；字典/缓存按阶段和可重建性并存，容量比较统一表尺寸与dtype。"
    },
    {
      "domain": "recommendation-system",
      "a": "recommendation-system:embedding-infrastructure-001",
      "b": "recommendation-system:embedding-infrastructure-003",
      "verdict": "同源互补/条件化",
      "reason": "训练状态完整性、缓存驻留与混合并行是不同要求；字典/缓存按阶段和可重建性并存，容量比较统一表尺寸与dtype。"
    },
    {
      "domain": "recommendation-system",
      "a": "recommendation-system:embedding-infrastructure-001",
      "b": "recommendation-system:embedding-infrastructure-004",
      "verdict": "同源互补/条件化",
      "reason": "训练状态完整性、缓存驻留与混合并行是不同要求；字典/缓存按阶段和可重建性并存，容量比较统一表尺寸与dtype。"
    },
    {
      "domain": "recommendation-system",
      "a": "recommendation-system:embedding-infrastructure-001",
      "b": "recommendation-system:embedding-infrastructure-005",
      "verdict": "同源互补/条件化",
      "reason": "训练状态完整性、缓存驻留与混合并行是不同要求；字典/缓存按阶段和可重建性并存，容量比较统一表尺寸与dtype。"
    },
    {
      "domain": "recommendation-system",
      "a": "recommendation-system:embedding-infrastructure-001",
      "b": "recommendation-system:embedding-infrastructure-006",
      "verdict": "同源互补/条件化",
      "reason": "训练状态完整性、缓存驻留与混合并行是不同要求；字典/缓存按阶段和可重建性并存，容量比较统一表尺寸与dtype。"
    },
    {
      "domain": "recommendation-system",
      "a": "recommendation-system:embedding-infrastructure-002",
      "b": "recommendation-system:embedding-infrastructure-003",
      "verdict": "同源互补/条件化",
      "reason": "训练状态完整性、缓存驻留与混合并行是不同要求；字典/缓存按阶段和可重建性并存，容量比较统一表尺寸与dtype。"
    },
    {
      "domain": "recommendation-system",
      "a": "recommendation-system:embedding-infrastructure-002",
      "b": "recommendation-system:embedding-infrastructure-004",
      "verdict": "同源互补/条件化",
      "reason": "训练状态完整性、缓存驻留与混合并行是不同要求；字典/缓存按阶段和可重建性并存，容量比较统一表尺寸与dtype。"
    },
    {
      "domain": "recommendation-system",
      "a": "recommendation-system:embedding-infrastructure-002",
      "b": "recommendation-system:embedding-infrastructure-005",
      "verdict": "同源互补/条件化",
      "reason": "训练状态完整性、缓存驻留与混合并行是不同要求；字典/缓存按阶段和可重建性并存，容量比较统一表尺寸与dtype。"
    },
    {
      "domain": "recommendation-system",
      "a": "recommendation-system:embedding-infrastructure-002",
      "b": "recommendation-system:embedding-infrastructure-006",
      "verdict": "同源互补/条件化",
      "reason": "训练状态完整性、缓存驻留与混合并行是不同要求；字典/缓存按阶段和可重建性并存，容量比较统一表尺寸与dtype。"
    },
    {
      "domain": "recommendation-system",
      "a": "recommendation-system:embedding-infrastructure-003",
      "b": "recommendation-system:embedding-infrastructure-004",
      "verdict": "同源互补/条件化",
      "reason": "训练状态完整性、缓存驻留与混合并行是不同要求；字典/缓存按阶段和可重建性并存，容量比较统一表尺寸与dtype。"
    },
    {
      "domain": "recommendation-system",
      "a": "recommendation-system:embedding-infrastructure-003",
      "b": "recommendation-system:embedding-infrastructure-005",
      "verdict": "同源互补/条件化",
      "reason": "训练状态完整性、缓存驻留与混合并行是不同要求；字典/缓存按阶段和可重建性并存，容量比较统一表尺寸与dtype。"
    },
    {
      "domain": "recommendation-system",
      "a": "recommendation-system:embedding-infrastructure-003",
      "b": "recommendation-system:embedding-infrastructure-006",
      "verdict": "同源互补/条件化",
      "reason": "训练状态完整性、缓存驻留与混合并行是不同要求；字典/缓存按阶段和可重建性并存，容量比较统一表尺寸与dtype。"
    },
    {
      "domain": "recommendation-system",
      "a": "recommendation-system:embedding-infrastructure-004",
      "b": "recommendation-system:embedding-infrastructure-005",
      "verdict": "同源互补/条件化",
      "reason": "训练状态完整性、缓存驻留与混合并行是不同要求；字典/缓存按阶段和可重建性并存，容量比较统一表尺寸与dtype。"
    },
    {
      "domain": "recommendation-system",
      "a": "recommendation-system:embedding-infrastructure-004",
      "b": "recommendation-system:embedding-infrastructure-006",
      "verdict": "同源互补/条件化",
      "reason": "训练状态完整性、缓存驻留与混合并行是不同要求；字典/缓存按阶段和可重建性并存，容量比较统一表尺寸与dtype。"
    },
    {
      "domain": "recommendation-system",
      "a": "recommendation-system:embedding-infrastructure-005",
      "b": "recommendation-system:embedding-infrastructure-006",
      "verdict": "同源互补/条件化",
      "reason": "训练状态完整性、缓存驻留与混合并行是不同要求；字典/缓存按阶段和可重建性并存，容量比较统一表尺寸与dtype。"
    },
    {
      "domain": "recommendation-system",
      "a": "recommendation-system:recsys-training-infra-001",
      "b": "recommendation-system:recsys-training-infra-002",
      "verdict": "同源互补/条件化",
      "reason": "训练内梯度一致性、训练→服务发布新鲜度及故障快照分别计量；Monolith、TorchRec与Persia按负载和更新重叠选策略。"
    },
    {
      "domain": "recommendation-system",
      "a": "recommendation-system:recsys-training-infra-001",
      "b": "recommendation-system:recsys-training-infra-003",
      "verdict": "同源互补/条件化",
      "reason": "训练内梯度一致性、训练→服务发布新鲜度及故障快照分别计量；Monolith、TorchRec与Persia按负载和更新重叠选策略。"
    },
    {
      "domain": "recommendation-system",
      "a": "recommendation-system:recsys-training-infra-001",
      "b": "recommendation-system:recsys-training-infra-004",
      "verdict": "同源互补/条件化",
      "reason": "训练内梯度一致性、训练→服务发布新鲜度及故障快照分别计量；Monolith、TorchRec与Persia按负载和更新重叠选策略。"
    },
    {
      "domain": "recommendation-system",
      "a": "recommendation-system:recsys-training-infra-001",
      "b": "recommendation-system:recsys-training-infra-005",
      "verdict": "同源互补/条件化",
      "reason": "训练内梯度一致性、训练→服务发布新鲜度及故障快照分别计量；Monolith、TorchRec与Persia按负载和更新重叠选策略。"
    },
    {
      "domain": "recommendation-system",
      "a": "recommendation-system:recsys-training-infra-001",
      "b": "recommendation-system:recsys-training-infra-006",
      "verdict": "同源互补/条件化",
      "reason": "训练内梯度一致性、训练→服务发布新鲜度及故障快照分别计量；Monolith、TorchRec与Persia按负载和更新重叠选策略。"
    },
    {
      "domain": "recommendation-system",
      "a": "recommendation-system:recsys-training-infra-001",
      "b": "recommendation-system:recsys-training-infra-007",
      "verdict": "表面相反，边界已解",
      "reason": "Monolith的分钟/天级是training PS→serving PS参数发布；Persia同步/异步是训练内梯度更新，不能互相否定。"
    },
    {
      "domain": "recommendation-system",
      "a": "recommendation-system:recsys-training-infra-001",
      "b": "recommendation-system:recsys-training-infra-008",
      "verdict": "同源互补/条件化",
      "reason": "训练内梯度一致性、训练→服务发布新鲜度及故障快照分别计量；Monolith、TorchRec与Persia按负载和更新重叠选策略。"
    },
    {
      "domain": "recommendation-system",
      "a": "recommendation-system:recsys-training-infra-001",
      "b": "recommendation-system:recsys-training-infra-009",
      "verdict": "同源互补/条件化",
      "reason": "训练内梯度一致性、训练→服务发布新鲜度及故障快照分别计量；Monolith、TorchRec与Persia按负载和更新重叠选策略。"
    },
    {
      "domain": "recommendation-system",
      "a": "recommendation-system:recsys-training-infra-002",
      "b": "recommendation-system:recsys-training-infra-003",
      "verdict": "同源互补/条件化",
      "reason": "训练内梯度一致性、训练→服务发布新鲜度及故障快照分别计量；Monolith、TorchRec与Persia按负载和更新重叠选策略。"
    },
    {
      "domain": "recommendation-system",
      "a": "recommendation-system:recsys-training-infra-002",
      "b": "recommendation-system:recsys-training-infra-004",
      "verdict": "同源互补/条件化",
      "reason": "训练内梯度一致性、训练→服务发布新鲜度及故障快照分别计量；Monolith、TorchRec与Persia按负载和更新重叠选策略。"
    },
    {
      "domain": "recommendation-system",
      "a": "recommendation-system:recsys-training-infra-002",
      "b": "recommendation-system:recsys-training-infra-005",
      "verdict": "同源互补/条件化",
      "reason": "训练内梯度一致性、训练→服务发布新鲜度及故障快照分别计量；Monolith、TorchRec与Persia按负载和更新重叠选策略。"
    },
    {
      "domain": "recommendation-system",
      "a": "recommendation-system:recsys-training-infra-002",
      "b": "recommendation-system:recsys-training-infra-006",
      "verdict": "同源互补/条件化",
      "reason": "训练内梯度一致性、训练→服务发布新鲜度及故障快照分别计量；Monolith、TorchRec与Persia按负载和更新重叠选策略。"
    },
    {
      "domain": "recommendation-system",
      "a": "recommendation-system:recsys-training-infra-002",
      "b": "recommendation-system:recsys-training-infra-007",
      "verdict": "同源互补/条件化",
      "reason": "训练内梯度一致性、训练→服务发布新鲜度及故障快照分别计量；Monolith、TorchRec与Persia按负载和更新重叠选策略。"
    },
    {
      "domain": "recommendation-system",
      "a": "recommendation-system:recsys-training-infra-002",
      "b": "recommendation-system:recsys-training-infra-008",
      "verdict": "同源互补/条件化",
      "reason": "训练内梯度一致性、训练→服务发布新鲜度及故障快照分别计量；Monolith、TorchRec与Persia按负载和更新重叠选策略。"
    },
    {
      "domain": "recommendation-system",
      "a": "recommendation-system:recsys-training-infra-002",
      "b": "recommendation-system:recsys-training-infra-009",
      "verdict": "同源互补/条件化",
      "reason": "训练内梯度一致性、训练→服务发布新鲜度及故障快照分别计量；Monolith、TorchRec与Persia按负载和更新重叠选策略。"
    },
    {
      "domain": "recommendation-system",
      "a": "recommendation-system:recsys-training-infra-003",
      "b": "recommendation-system:recsys-training-infra-004",
      "verdict": "同源互补/条件化",
      "reason": "训练内梯度一致性、训练→服务发布新鲜度及故障快照分别计量；Monolith、TorchRec与Persia按负载和更新重叠选策略。"
    },
    {
      "domain": "recommendation-system",
      "a": "recommendation-system:recsys-training-infra-003",
      "b": "recommendation-system:recsys-training-infra-005",
      "verdict": "同源互补/条件化",
      "reason": "训练内梯度一致性、训练→服务发布新鲜度及故障快照分别计量；Monolith、TorchRec与Persia按负载和更新重叠选策略。"
    },
    {
      "domain": "recommendation-system",
      "a": "recommendation-system:recsys-training-infra-003",
      "b": "recommendation-system:recsys-training-infra-006",
      "verdict": "同源互补/条件化",
      "reason": "训练内梯度一致性、训练→服务发布新鲜度及故障快照分别计量；Monolith、TorchRec与Persia按负载和更新重叠选策略。"
    },
    {
      "domain": "recommendation-system",
      "a": "recommendation-system:recsys-training-infra-003",
      "b": "recommendation-system:recsys-training-infra-007",
      "verdict": "同源互补/条件化",
      "reason": "训练内梯度一致性、训练→服务发布新鲜度及故障快照分别计量；Monolith、TorchRec与Persia按负载和更新重叠选策略。"
    },
    {
      "domain": "recommendation-system",
      "a": "recommendation-system:recsys-training-infra-003",
      "b": "recommendation-system:recsys-training-infra-008",
      "verdict": "同源互补/条件化",
      "reason": "训练内梯度一致性、训练→服务发布新鲜度及故障快照分别计量；Monolith、TorchRec与Persia按负载和更新重叠选策略。"
    },
    {
      "domain": "recommendation-system",
      "a": "recommendation-system:recsys-training-infra-003",
      "b": "recommendation-system:recsys-training-infra-009",
      "verdict": "同源互补/条件化",
      "reason": "训练内梯度一致性、训练→服务发布新鲜度及故障快照分别计量；Monolith、TorchRec与Persia按负载和更新重叠选策略。"
    },
    {
      "domain": "recommendation-system",
      "a": "recommendation-system:recsys-training-infra-004",
      "b": "recommendation-system:recsys-training-infra-005",
      "verdict": "同源互补/条件化",
      "reason": "训练内梯度一致性、训练→服务发布新鲜度及故障快照分别计量；Monolith、TorchRec与Persia按负载和更新重叠选策略。"
    },
    {
      "domain": "recommendation-system",
      "a": "recommendation-system:recsys-training-infra-004",
      "b": "recommendation-system:recsys-training-infra-006",
      "verdict": "同源互补/条件化",
      "reason": "训练内梯度一致性、训练→服务发布新鲜度及故障快照分别计量；Monolith、TorchRec与Persia按负载和更新重叠选策略。"
    },
    {
      "domain": "recommendation-system",
      "a": "recommendation-system:recsys-training-infra-004",
      "b": "recommendation-system:recsys-training-infra-007",
      "verdict": "同源互补/条件化",
      "reason": "训练内梯度一致性、训练→服务发布新鲜度及故障快照分别计量；Monolith、TorchRec与Persia按负载和更新重叠选策略。"
    },
    {
      "domain": "recommendation-system",
      "a": "recommendation-system:recsys-training-infra-004",
      "b": "recommendation-system:recsys-training-infra-008",
      "verdict": "同源互补/条件化",
      "reason": "训练内梯度一致性、训练→服务发布新鲜度及故障快照分别计量；Monolith、TorchRec与Persia按负载和更新重叠选策略。"
    },
    {
      "domain": "recommendation-system",
      "a": "recommendation-system:recsys-training-infra-004",
      "b": "recommendation-system:recsys-training-infra-009",
      "verdict": "同源互补/条件化",
      "reason": "训练内梯度一致性、训练→服务发布新鲜度及故障快照分别计量；Monolith、TorchRec与Persia按负载和更新重叠选策略。"
    },
    {
      "domain": "recommendation-system",
      "a": "recommendation-system:recsys-training-infra-005",
      "b": "recommendation-system:recsys-training-infra-006",
      "verdict": "同源互补/条件化",
      "reason": "训练内梯度一致性、训练→服务发布新鲜度及故障快照分别计量；Monolith、TorchRec与Persia按负载和更新重叠选策略。"
    },
    {
      "domain": "recommendation-system",
      "a": "recommendation-system:recsys-training-infra-005",
      "b": "recommendation-system:recsys-training-infra-007",
      "verdict": "同源互补/条件化",
      "reason": "训练内梯度一致性、训练→服务发布新鲜度及故障快照分别计量；Monolith、TorchRec与Persia按负载和更新重叠选策略。"
    },
    {
      "domain": "recommendation-system",
      "a": "recommendation-system:recsys-training-infra-005",
      "b": "recommendation-system:recsys-training-infra-008",
      "verdict": "同源互补/条件化",
      "reason": "训练内梯度一致性、训练→服务发布新鲜度及故障快照分别计量；Monolith、TorchRec与Persia按负载和更新重叠选策略。"
    },
    {
      "domain": "recommendation-system",
      "a": "recommendation-system:recsys-training-infra-005",
      "b": "recommendation-system:recsys-training-infra-009",
      "verdict": "同源互补/条件化",
      "reason": "训练内梯度一致性、训练→服务发布新鲜度及故障快照分别计量；Monolith、TorchRec与Persia按负载和更新重叠选策略。"
    },
    {
      "domain": "recommendation-system",
      "a": "recommendation-system:recsys-training-infra-006",
      "b": "recommendation-system:recsys-training-infra-007",
      "verdict": "同源互补/条件化",
      "reason": "训练内梯度一致性、训练→服务发布新鲜度及故障快照分别计量；Monolith、TorchRec与Persia按负载和更新重叠选策略。"
    },
    {
      "domain": "recommendation-system",
      "a": "recommendation-system:recsys-training-infra-006",
      "b": "recommendation-system:recsys-training-infra-008",
      "verdict": "同源互补/条件化",
      "reason": "训练内梯度一致性、训练→服务发布新鲜度及故障快照分别计量；Monolith、TorchRec与Persia按负载和更新重叠选策略。"
    },
    {
      "domain": "recommendation-system",
      "a": "recommendation-system:recsys-training-infra-006",
      "b": "recommendation-system:recsys-training-infra-009",
      "verdict": "同源互补/条件化",
      "reason": "训练内梯度一致性、训练→服务发布新鲜度及故障快照分别计量；Monolith、TorchRec与Persia按负载和更新重叠选策略。"
    },
    {
      "domain": "recommendation-system",
      "a": "recommendation-system:recsys-training-infra-007",
      "b": "recommendation-system:recsys-training-infra-008",
      "verdict": "同源互补/条件化",
      "reason": "训练内梯度一致性、训练→服务发布新鲜度及故障快照分别计量；Monolith、TorchRec与Persia按负载和更新重叠选策略。"
    },
    {
      "domain": "recommendation-system",
      "a": "recommendation-system:recsys-training-infra-007",
      "b": "recommendation-system:recsys-training-infra-009",
      "verdict": "同源互补/条件化",
      "reason": "训练内梯度一致性、训练→服务发布新鲜度及故障快照分别计量；Monolith、TorchRec与Persia按负载和更新重叠选策略。"
    },
    {
      "domain": "recommendation-system",
      "a": "recommendation-system:recsys-training-infra-008",
      "b": "recommendation-system:recsys-training-infra-009",
      "verdict": "同源互补/条件化",
      "reason": "训练内梯度一致性、训练→服务发布新鲜度及故障快照分别计量；Monolith、TorchRec与Persia按负载和更新重叠选策略。"
    },
    {
      "domain": "recommendation-system",
      "a": "recommendation-system:retrieval-serving-001",
      "b": "recommendation-system:retrieval-serving-002",
      "verdict": "同源互补/条件化",
      "reason": "MIPS任务、损失与完整索引路径的适用边界互补；同一ScaNN论文不是独立验证。"
    },
    {
      "domain": "recommendation-system",
      "a": "recommendation-system:retrieval-serving-001",
      "b": "recommendation-system:retrieval-serving-003",
      "verdict": "同源互补/条件化",
      "reason": "MIPS任务、损失与完整索引路径的适用边界互补；同一ScaNN论文不是独立验证。"
    },
    {
      "domain": "recommendation-system",
      "a": "recommendation-system:retrieval-serving-002",
      "b": "recommendation-system:retrieval-serving-003",
      "verdict": "同源互补/条件化",
      "reason": "MIPS任务、损失与完整索引路径的适用边界互补；同一ScaNN论文不是独立验证。"
    },
    {
      "domain": "recommendation-system",
      "a": "recommendation-system:embedding-infrastructure-001",
      "b": "kvcache:lifecycle-001",
      "verdict": "历史分歧条件化，已关闭",
      "reason": "字典保持训练参数身份，缓存接受可从后备状态恢复的驻留淘汰；分阶段落位，非同一前提下相反结论。HKV源码快照已stale，不当作当前性能验证。"
    },
    {
      "domain": "recommendation-system",
      "a": "recommendation-system:embedding-infrastructure-005",
      "b": "kvcache:lifecycle-001",
      "verdict": "历史分歧条件化，已关闭",
      "reason": "字典保持训练参数身份，缓存接受可从后备状态恢复的驻留淘汰；分阶段落位，非同一前提下相反结论。HKV源码快照已stale，不当作当前性能验证。"
    }
  ]
}