{
  "schema_version": 1,
  "reviewed_on": "2026-10-15",
  "note": "Independent priority/evidence checks; no new implementation or local cost measurement.",
  "reviews": [
    {
      "id": "2610.20293",
      "title": "CASI: Credit Assignment via Set Intervention for Agentic Retrieval",
      "published": "2026-10-14T14:47:46Z",
      "revision": 1,
      "priority": "P0",
      "decision": "reviewed_deferred",
      "resolution": "not_reproduced_high_cost",
      "reason": "CASI 在固定检索快照上做集合干预，经 Shapley 归因和有界零和投影构造动作信用；正文 §4.5 随机七天、10% 生产流量 A/B，Advv +3.51%，工业证据通过。生产回放/教师与完整训练规模未公开，不能廉价核定忠实复现成本；按预算门槛不复现，审查 DONE。",
      "locations": [
        "§3 snapshot intervention and shaping",
        "§4.4 sensitivity backend",
        "§4.5 production A/B"
      ],
      "scale": "1200 请求×31 coalition 的离线审计、5 次独立训练；未公开生产回放/完整模型配置",
      "cost_type": "unresolved_material_cost_and_private_replay_not_local_measurement",
      "industrial_gate_passed": true,
      "abstract_review_completed": true,
      "fulltext_review_completed": true,
      "full_article_read": false,
      "source_url": "https://arxiv.org/html/2610.20293v1",
      "source_sha256": "b9d5f5ff912eb7a126b05d29030d3f633777783317fb2bd942cb912ec0e94ee7",
      "metadata_sha256": "f8325f572d0c5d0561946243bf10a0620c04ad3609ed1f9265d9f200239f6315",
      "review_level": "independent_bounded_priority_method_evaluation_and_cost_review",
      "read_scope": "Only named original passages, not complete article/appendices.",
      "review_done": true,
      "automatic_retry": false,
      "automatic_implementation_queue": false,
      "implementation_status": "not_implemented_by_review",
      "review_artifact": "research-audits/oct15-review-closure/root-priority-reviews.json",
      "cost_assessment": {
        "source_url": "https://arxiv.org/html/2610.20293v1",
        "basis": "§3 snapshot intervention and shaping; §4.4 sensitivity backend; §4.5 production A/B",
        "scale": "1200 请求×31 coalition 的离线审计、5 次独立训练；未公开生产回放/完整模型配置",
        "available_resource": "单卡 A100 或 A30，或本地 Mac",
        "cost_type": "unresolved_material_cost_and_private_replay_not_local_measurement"
      }
    },
    {
      "id": "2610.15696",
      "title": "DRanker: A Transformer-Based Multi-Task Ranking Model for the Disney+ Homepage",
      "published": "2026-10-11T15:29:51Z",
      "revision": 1,
      "priority": "P0",
      "decision": "reviewed_deferred",
      "resolution": "not_reproduced_high_cost",
      "reason": "DRanker 的约 7M 参数 encoder-only 多任务排序器结合因果去偏，正文有线上 A/B 与全局上线结果；不是只有摘要宣称部署。训练依赖数月 Disney 私有交互与数亿训练样本，无法在公开数据和单卡预算下忠实核验生产设置；不复现、审查 DONE。首发 arXiv 2026-10-11，属旧文补发现，不是本窗口新首发。",
      "locations": [
        "§3 training data construction",
        "§4 model architecture",
        "§5.5–5.6 offline selection and online A/B",
        "§6 serving",
        "§8 global deployment"
      ],
      "scale": "约 7M 参数、数亿私有训练样本、频繁重训",
      "cost_type": "published_configuration_and_private_data_not_local_measurement",
      "industrial_gate_passed": true,
      "abstract_review_completed": true,
      "fulltext_review_completed": true,
      "full_article_read": false,
      "source_url": "https://arxiv.org/html/2610.15696v1",
      "source_sha256": "c62972072026bf97913645e33c61299afbc1ae89186bcdd499a3df4f6c509850",
      "metadata_sha256": "16f1c1336e5d072c2c3ef2352d69c9ceb1db01c24e2de2cf9a4d4c0f11f9a828",
      "review_level": "independent_bounded_priority_method_evaluation_and_cost_review",
      "read_scope": "Only named original passages, not complete article/appendices.",
      "review_done": true,
      "automatic_retry": false,
      "automatic_implementation_queue": false,
      "implementation_status": "not_implemented_by_review",
      "review_artifact": "research-audits/oct15-review-closure/root-priority-reviews.json",
      "cost_assessment": {
        "source_url": "https://arxiv.org/html/2610.15696v1",
        "basis": "§3 training data construction; §4 model architecture; §5.5–5.6 offline selection and online A/B; §6 serving; §8 global deployment",
        "scale": "约 7M 参数、数亿私有训练样本、频繁重训",
        "available_resource": "单卡 A100 或 A30，或本地 Mac",
        "cost_type": "published_configuration_and_private_data_not_local_measurement"
      }
    },
    {
      "id": "2610.14394",
      "title": "Understanding and Mitigating Training Instability in Large-Scale Recommender Systems: A Cascade Perspective",
      "published": "2026-10-09T23:37:02Z",
      "revision": 1,
      "priority": "P2",
      "decision": "rejected_or_out_of_scope",
      "resolution": "industrial_evidence_gate_not_met",
      "reason": "Google 的 G→W→L 发散级联与 SmeLU/RMSNorm/gradient clipping、activation clipping、Huber loss 已核正文 §3；比较的是训练稳定性、AUC/RMSE，没有该方法的量化业务线上 A/B 或经批准的全流量收益。工业门槛不通过；不因 Google 署名或真实日志直接接受。",
      "locations": [
        "§2 cascade/interventions",
        "§3 experiments/RQ1–RQ2",
        "§3.1 experimental setup",
        "§4 conclusion"
      ],
      "industrial_gate_passed": false,
      "abstract_review_completed": true,
      "fulltext_review_completed": true,
      "full_article_read": false,
      "source_url": "https://arxiv.org/html/2610.14394v1",
      "source_sha256": "e0b6611fcd896dd6e0048a0fd7fc75b0287a995e99ab3adc8a70b06c84a52e96",
      "metadata_sha256": "8e3c3d8ef8c1f94760309ac92e3de7ae3728b0510437844ef428c117059c0208",
      "review_level": "independent_bounded_priority_method_evaluation_and_cost_review",
      "read_scope": "Only named original passages, not complete article/appendices.",
      "review_done": true,
      "automatic_retry": false,
      "automatic_implementation_queue": false,
      "implementation_status": "not_implemented_by_review",
      "review_artifact": "research-audits/oct15-review-closure/root-priority-reviews.json"
    },
    {
      "id": "2610.13989",
      "title": "PsyBer-Agent: A Psychology-Driven User Agent for Recommender Systems",
      "published": "2026-09-26T14:42:02Z",
      "revision": 1,
      "priority": "P2",
      "decision": "reviewed_deferred",
      "resolution": "not_reproduced_high_cost",
      "reason": "PsyBer-Agent 的三状态 Psy-Engine、UGW 对齐与 WebSim 是用户模拟。§5.2 虽称 A/B，实际固定后端比较 500 人与 500 模拟用户的行为一致性，不是推荐器生产业务 A/B；不能通过工业门槛。原评测使用 Qwen3-VL-235B 等大量闭环模拟，成本过高不复现，审查 DONE。",
      "locations": [
        "§3 Psy-Engine and alignment",
        "§5.2 fixed-backend agreement",
        "Appendix D.6",
        "Appendix E.9"
      ],
      "scale": "235B VLM、长闭环和大规模模拟",
      "cost_type": "published_configuration_not_local_measurement",
      "industrial_gate_passed": false,
      "abstract_review_completed": true,
      "fulltext_review_completed": true,
      "full_article_read": false,
      "source_url": "https://arxiv.org/html/2610.13989v1",
      "source_sha256": "9b69089f30900262727eef42aa3c9bd39f0ea61d6663fc9ffafc4737c831a31c",
      "metadata_sha256": "1394ac4c2099275704912d0b3a5081cf5dde5dc434e099613625f29df094c81e",
      "review_level": "independent_bounded_priority_method_evaluation_and_cost_review",
      "read_scope": "Only named original passages, not complete article/appendices.",
      "review_done": true,
      "automatic_retry": false,
      "automatic_implementation_queue": false,
      "implementation_status": "not_implemented_by_review",
      "review_artifact": "research-audits/oct15-review-closure/root-priority-reviews.json",
      "cost_assessment": {
        "source_url": "https://arxiv.org/html/2610.13989v1",
        "basis": "§3 Psy-Engine and alignment; §5.2 fixed-backend agreement; Appendix D.6; Appendix E.9",
        "scale": "235B VLM、长闭环和大规模模拟",
        "available_resource": "单卡 A100 或 A30，或本地 Mac",
        "cost_type": "published_configuration_not_local_measurement"
      }
    },
    {
      "id": "2610.19454",
      "title": "AdCoEvo: A Co-Evolving Multi-Agent Advertising Delivery System",
      "published": "2026-10-14T02:40:56Z",
      "revision": 1,
      "priority": "P1",
      "decision": "reviewed_deferred",
      "resolution": "not_reproduced_high_cost",
      "reason": "AdCoEvo 是三角色广告策略共进化；§5.1 的 AuctionNet 是模拟器，14.02% ADV 不是生产线上 A/B。可归 Agent/RSI，不作工业 P0。8B 三角色、5 条×50-period 序列及交替更新的完整实验成本高，不复现、审查 DONE。",
      "locations": [
        "§3 co-evolution",
        "§5.1 AuctionNet setup",
        "§5.3"
      ],
      "scale": "8B 三角色、重复50-period序列和在线交替训练",
      "cost_type": "published_configuration_not_local_measurement",
      "industrial_gate_passed": false,
      "abstract_review_completed": true,
      "fulltext_review_completed": true,
      "full_article_read": false,
      "source_url": "https://arxiv.org/html/2610.19454v1",
      "source_sha256": "3b40fabf72dd5010b500a636067fb6231f3741e11cec1f141b0584d161ccc15e",
      "metadata_sha256": "5ce997cdca4af2c4a4e9c686941845a6b5434bb0047446e1a62e4a29f5688f34",
      "review_level": "independent_bounded_priority_method_evaluation_and_cost_review",
      "read_scope": "Only named original passages, not complete article/appendices.",
      "review_done": true,
      "automatic_retry": false,
      "automatic_implementation_queue": false,
      "implementation_status": "not_implemented_by_review",
      "review_artifact": "research-audits/oct15-review-closure/root-priority-reviews.json",
      "cost_assessment": {
        "source_url": "https://arxiv.org/html/2610.19454v1",
        "basis": "§3 co-evolution; §5.1 AuctionNet setup; §5.3",
        "scale": "8B 三角色、重复50-period序列和在线交替训练",
        "available_resource": "单卡 A100 或 A30，或本地 Mac",
        "cost_type": "published_configuration_not_local_measurement"
      }
    },
    {
      "id": "2610.14709",
      "title": "Vela: User Behavior as Supervision Sparks LLM Intelligence",
      "published": "2026-10-10T08:28:00Z",
      "revision": 1,
      "priority": "P1",
      "decision": "reviewed_deferred",
      "resolution": "not_reproduced_high_cost",
      "reason": "Vela 把长期用户行为变为任务与轨迹；属于 LLM 数据/训练方案而非已证实的工业业务线上 A/B。35B-A3B、131072 长度、75K 轨迹，两 epoch；附录 RL 使用 48 GPUs（32 训练/16 rollout）。成本过高，不复现、审查 DONE。",
      "locations": [
        "§3.1 setup",
        "Appendix E RL configuration"
      ],
      "scale": "35B-A3B、131072 context、75K tasks；RL 48 GPUs",
      "cost_type": "published_configuration_not_local_measurement",
      "abstract_review_completed": true,
      "fulltext_review_completed": true,
      "full_article_read": false,
      "source_url": "https://arxiv.org/html/2610.14709v1",
      "source_sha256": "dd866c580d5616aeecde0dc3208dd5179daf0d6922cadd81379ac6e20d2a0efc",
      "metadata_sha256": "72289f5f3e3ce866be2413053811efa7c60cbcf227a2d79e40764b08d0650d36",
      "review_level": "independent_bounded_priority_method_evaluation_and_cost_review",
      "read_scope": "Only named original passages, not complete article/appendices.",
      "review_done": true,
      "automatic_retry": false,
      "automatic_implementation_queue": false,
      "implementation_status": "not_implemented_by_review",
      "review_artifact": "research-audits/oct15-review-closure/root-priority-reviews.json",
      "cost_assessment": {
        "source_url": "https://arxiv.org/html/2610.14709v1",
        "basis": "§3.1 setup; Appendix E RL configuration",
        "scale": "35B-A3B、131072 context、75K tasks；RL 48 GPUs",
        "available_resource": "单卡 A100 或 A30，或本地 Mac",
        "cost_type": "published_configuration_not_local_measurement"
      }
    },
    {
      "id": "2610.19302",
      "title": "DTDD: Divergence-Triggered Dynamic Distillation for Reliable On-Policy Supervision",
      "published": "2026-10-13T23:26:26Z",
      "revision": 1,
      "priority": "P1",
      "decision": "reviewed_deferred",
      "resolution": "not_reproduced_high_cost",
      "reason": "DTDD 用片段分歧触发教师接管，学生前缀继续 OPD、教师后缀行为克隆；原研究23实验臂、约22 serial hours和43M教师生成 tokens，成本过高，不复现、审查 DONE。",
      "locations": [
        "§3 DTDD",
        "§4 cost comparison"
      ],
      "scale": "23 arms / 22 serial hours / 43M teacher-generated tokens",
      "cost_type": "published_configuration_not_local_measurement",
      "abstract_review_completed": true,
      "fulltext_review_completed": true,
      "full_article_read": false,
      "source_url": "https://arxiv.org/html/2610.19302v1",
      "source_sha256": "4e3ac1546d09b4a1cb29dd8d9a4d952d288b509409733f9c851446bfc3df26c8",
      "metadata_sha256": "5b963e2c68366941b38e00825efe9ac10414938f1a2e72a5b67da76e83d64c5a",
      "review_level": "independent_bounded_priority_method_evaluation_and_cost_review",
      "read_scope": "Only named original passages, not complete article/appendices.",
      "review_done": true,
      "automatic_retry": false,
      "automatic_implementation_queue": false,
      "implementation_status": "not_implemented_by_review",
      "review_artifact": "research-audits/oct15-review-closure/root-priority-reviews.json",
      "cost_assessment": {
        "source_url": "https://arxiv.org/html/2610.19302v1",
        "basis": "§3 DTDD; §4 cost comparison",
        "scale": "23 arms / 22 serial hours / 43M teacher-generated tokens",
        "available_resource": "单卡 A100 或 A30，或本地 Mac",
        "cost_type": "published_configuration_not_local_measurement"
      }
    },
    {
      "id": "2610.20174",
      "title": "RM-OPD: On-Policy Distillation for Generative Reward Models",
      "published": "2026-10-14T13:20:55Z",
      "revision": 1,
      "priority": "P1",
      "decision": "reviewed_deferred",
      "resolution": "not_reproduced_high_cost",
      "reason": "RM-OPD 先训练领域 GRPO 教师，再通过路由 OPD/领域权重合并生成奖励模型；74K 偏好、训练2节点×8 H800、评测8 H800，成本过高，不复现、审查 DONE。",
      "locations": [
        "§3 RM-OPD",
        "§6.1 experimental setup",
        "Appendix B training hardware"
      ],
      "scale": "7B 教师/学生、74K preferences、16 H800 training / 8 H800 evaluation",
      "cost_type": "published_configuration_not_local_measurement",
      "abstract_review_completed": true,
      "fulltext_review_completed": true,
      "full_article_read": false,
      "source_url": "https://arxiv.org/html/2610.20174v1",
      "source_sha256": "e3e2b330b7beb2940bfea379e1a736ad1fb312b932b073f4da45f04e346cd4e9",
      "metadata_sha256": "cc31cfc5b1fc144510d8e0373a9fcdc40f805b4635fcfcefc760df1ea3f09fe2",
      "review_level": "independent_bounded_priority_method_evaluation_and_cost_review",
      "read_scope": "Only named original passages, not complete article/appendices.",
      "review_done": true,
      "automatic_retry": false,
      "automatic_implementation_queue": false,
      "implementation_status": "not_implemented_by_review",
      "review_artifact": "research-audits/oct15-review-closure/root-priority-reviews.json",
      "cost_assessment": {
        "source_url": "https://arxiv.org/html/2610.20174v1",
        "basis": "§3 RM-OPD; §6.1 experimental setup; Appendix B training hardware",
        "scale": "7B 教师/学生、74K preferences、16 H800 training / 8 H800 evaluation",
        "available_resource": "单卡 A100 或 A30，或本地 Mac",
        "cost_type": "published_configuration_not_local_measurement"
      }
    },
    {
      "id": "2610.20572",
      "title": "Recovering Reliable Supervision via Critique-Guided On-Policy Distillation",
      "published": "2026-10-14T17:54:34Z",
      "revision": 1,
      "priority": "P1",
      "decision": "reviewed_deferred",
      "resolution": "not_reproduced_high_cost",
      "reason": "CriticOPD 让教师批评错误轨迹，学生修订，教师用特权 critique 打分；训练4张A100（学生2/教师2）、最长16384响应，评测8–32次采样。即使学生较小也不是低成本完整复现，不复现、审查 DONE。作者仓库已核为 kzhao5/CriticOPD。",
      "locations": [
        "§3 CriticOPD",
        "Appendix C.1–C.2"
      ],
      "scale": "0.6B/1.7B students、4B teacher；4 A100；长响应和重复采样",
      "cost_type": "published_configuration_not_local_measurement",
      "abstract_review_completed": true,
      "fulltext_review_completed": true,
      "full_article_read": false,
      "source_url": "https://arxiv.org/html/2610.20572v1",
      "source_sha256": "f14b87e88e0c21648936b5455c8a6bfb6f9e73b67feb5b754e9aa98aedcf29a9",
      "metadata_sha256": "6acad8623f1e79ee2a220f8126bbabd9b9b221fa7487106837932b0eeae52cd2",
      "review_level": "independent_bounded_priority_method_evaluation_and_cost_review",
      "read_scope": "Only named original passages, not complete article/appendices.",
      "review_done": true,
      "automatic_retry": false,
      "automatic_implementation_queue": false,
      "implementation_status": "not_implemented_by_review",
      "review_artifact": "research-audits/oct15-review-closure/root-priority-reviews.json",
      "cost_assessment": {
        "source_url": "https://arxiv.org/html/2610.20572v1",
        "basis": "§3 CriticOPD; Appendix C.1–C.2",
        "scale": "0.6B/1.7B students、4B teacher；4 A100；长响应和重复采样",
        "available_resource": "单卡 A100 或 A30，或本地 Mac",
        "cost_type": "published_configuration_not_local_measurement"
      }
    },
    {
      "id": "2610.20516",
      "title": "VASPO: Self-Distillation with Privileged Value Estimation",
      "published": "2026-10-14T17:17:18Z",
      "revision": 1,
      "priority": "P1",
      "decision": "reviewed_deferred",
      "resolution": "not_reproduced_high_cost",
      "reason": "VASPO 将同一特权上下文下 critic 的 PPO advantage 与自教师的有界概率比结合；原配置8 H200、full-parameter、200 rollout更新、每题8响应，成本过高，不复现、审查 DONE。",
      "locations": [
        "§3 VASPO",
        "Hyperparameter Record"
      ],
      "scale": "4B/7B actor、critic与self-teacher；8 H200、200 rollout updates",
      "cost_type": "published_configuration_not_local_measurement",
      "abstract_review_completed": true,
      "fulltext_review_completed": true,
      "full_article_read": false,
      "source_url": "https://arxiv.org/html/2610.20516v1",
      "source_sha256": "c096e8aef4eb876e1808c56b1a12292e2ba6f31d1b4c4d1b6f6d5c23617e27a3",
      "metadata_sha256": "6e5d4cba5fe13d9788809df22e6d401ebe30570fe3d222f8e4cdf714861a431c",
      "review_level": "independent_bounded_priority_method_evaluation_and_cost_review",
      "read_scope": "Only named original passages, not complete article/appendices.",
      "review_done": true,
      "automatic_retry": false,
      "automatic_implementation_queue": false,
      "implementation_status": "not_implemented_by_review",
      "review_artifact": "research-audits/oct15-review-closure/root-priority-reviews.json",
      "cost_assessment": {
        "source_url": "https://arxiv.org/html/2610.20516v1",
        "basis": "§3 VASPO; Hyperparameter Record",
        "scale": "4B/7B actor、critic与self-teacher；8 H200、200 rollout updates",
        "available_resource": "单卡 A100 或 A30，或本地 Mac",
        "cost_type": "published_configuration_not_local_measurement"
      }
    },
    {
      "id": "2610.20497",
      "title": "Hindsight-Privileged RLVR: Single-Rollout Token Credit Without Group Sampling",
      "published": "2026-10-14T17:06:59Z",
      "revision": 1,
      "priority": "P1",
      "decision": "reviewed_deferred",
      "resolution": "not_reproduced_high_cost",
      "reason": "HP-RLVR 用生成快照在因果/后见上下文中的 JS 差构造单轨迹 token credit，不读取参考答案。8B三任务100-step配对训练仍有大量绝对运行成本，不能因比GRPO相对更快当便宜；不复现、审查 DONE。",
      "locations": [
        "§3 token credit",
        "§4.2 fixed-budget timing",
        "Table 1 absolute timings"
      ],
      "scale": "Qwen3-4B/8B，数学/代码/ALFWorld，100-step matched runs",
      "cost_type": "published_configuration_not_local_measurement",
      "abstract_review_completed": true,
      "fulltext_review_completed": true,
      "full_article_read": false,
      "source_url": "https://arxiv.org/html/2610.20497v1",
      "source_sha256": "46b185fb3b5bee63e1206286936d843f13327d9b718f61486faf34984774674b",
      "metadata_sha256": "764ac63b4dc7369163c65f2ad022887b76579ea513fea636a940d1645259c93b",
      "review_level": "independent_bounded_priority_method_evaluation_and_cost_review",
      "read_scope": "Only named original passages, not complete article/appendices.",
      "review_done": true,
      "automatic_retry": false,
      "automatic_implementation_queue": false,
      "implementation_status": "not_implemented_by_review",
      "review_artifact": "research-audits/oct15-review-closure/root-priority-reviews.json",
      "cost_assessment": {
        "source_url": "https://arxiv.org/html/2610.20497v1",
        "basis": "§3 token credit; §4.2 fixed-budget timing; Table 1 absolute timings",
        "scale": "Qwen3-4B/8B，数学/代码/ALFWorld，100-step matched runs",
        "available_resource": "单卡 A100 或 A30，或本地 Mac",
        "cost_type": "published_configuration_not_local_measurement"
      }
    },
    {
      "id": "2610.19103",
      "title": "Contextual Credit Policy Optimization",
      "published": "2026-10-13T19:39:52Z",
      "revision": 1,
      "priority": "P1",
      "decision": "reviewed_deferred",
      "resolution": "not_reproduced_high_cost",
      "reason": "CCPO 使用上下文相似度加权 observation-matched peers，并结合历史与未来信用；附录每次8 H200、约19200轨迹，完整复现成本高，不复现、审查 DONE。",
      "locations": [
        "§3 contextual credit",
        "Appendix B training/ablation configurations"
      ],
      "scale": "1.5B/7B、ALFWorld/WebShop、8 H200 per run",
      "cost_type": "published_configuration_not_local_measurement",
      "abstract_review_completed": true,
      "fulltext_review_completed": true,
      "full_article_read": false,
      "source_url": "https://arxiv.org/html/2610.19103v1",
      "source_sha256": "bb452925a35f3fc74922d1aeef29b980f46cc8d04d4ae60c6962f92fcf74c806",
      "metadata_sha256": "04510cb06f95438e1d1281d013ebaf3a519f79a7a9b66ede7adf596fbdf57ff1",
      "review_level": "independent_bounded_priority_method_evaluation_and_cost_review",
      "read_scope": "Only named original passages, not complete article/appendices.",
      "review_done": true,
      "automatic_retry": false,
      "automatic_implementation_queue": false,
      "implementation_status": "not_implemented_by_review",
      "review_artifact": "research-audits/oct15-review-closure/root-priority-reviews.json",
      "cost_assessment": {
        "source_url": "https://arxiv.org/html/2610.19103v1",
        "basis": "§3 contextual credit; Appendix B training/ablation configurations",
        "scale": "1.5B/7B、ALFWorld/WebShop、8 H200 per run",
        "available_resource": "单卡 A100 或 A30，或本地 Mac",
        "cost_type": "published_configuration_not_local_measurement"
      }
    },
    {
      "id": "2610.19145",
      "title": "Demystifying On-Policy Distillation: Where Does the Teacher's Advantage Live?",
      "published": "2026-10-13T20:11:48Z",
      "revision": 1,
      "priority": "P1",
      "decision": "reviewed_deferred",
      "resolution": "not_reproduced_high_cost",
      "reason": "OPD 教师优势分析用受控 token flips 与分叉续写拆解教师纠正收益；每个分歧又生成多次长续写，整套多模型/多分叉配置无法廉价核定运行预算，按成本门槛不复现、审查 DONE，不自动重跑。",
      "locations": [
        "§3 controlled token flips",
        "Appendix A.3 Single Flip"
      ],
      "scale": "最多8192 tokens、分叉后4次最多3072 tokens续写，多模型选择策略",
      "cost_type": "unresolved_material_cost_after_primary_source_check",
      "abstract_review_completed": true,
      "fulltext_review_completed": true,
      "full_article_read": false,
      "source_url": "https://arxiv.org/html/2610.19145v1",
      "source_sha256": "0893daa56c1200cce16fde56b4bb12440b6dc61d0ba7d18196e86ff74db46d7d",
      "metadata_sha256": "cbd93aa9eecf4eebf80b7190d61671fdb0be5413c2c409cf0415d768df79118b",
      "review_level": "independent_bounded_priority_method_evaluation_and_cost_review",
      "read_scope": "Only named original passages, not complete article/appendices.",
      "review_done": true,
      "automatic_retry": false,
      "automatic_implementation_queue": false,
      "implementation_status": "not_implemented_by_review",
      "review_artifact": "research-audits/oct15-review-closure/root-priority-reviews.json",
      "cost_assessment": {
        "source_url": "https://arxiv.org/html/2610.19145v1",
        "basis": "§3 controlled token flips; Appendix A.3 Single Flip",
        "scale": "最多8192 tokens、分叉后4次最多3072 tokens续写，多模型选择策略",
        "available_resource": "单卡 A100 或 A30，或本地 Mac",
        "cost_type": "unresolved_material_cost_after_primary_source_check"
      }
    },
    {
      "id": "2610.19789",
      "title": "More Methods, Better Agents? A Reality Check on Self-Distillation in Agentic Reinforcement Learning",
      "published": "2026-10-14T08:36:31Z",
      "revision": 1,
      "priority": "P1",
      "decision": "reviewed_deferred",
      "resolution": "not_reproduced_high_cost",
      "reason": "Self-distillation reality check 比较十种方法与GRPO，132训练运行、12124.4 H800 GPU-hours；只保留公平比较结论，绝不整套重跑。成本过高、不复现，审查 DONE。",
      "locations": [
        "abstract study scale",
        "experimental setup and compute accounting"
      ],
      "scale": "132 runs / 12124.4 H800 GPU-hours / four environments / three model scales",
      "cost_type": "published_measurement_not_local_measurement",
      "abstract_review_completed": true,
      "fulltext_review_completed": true,
      "full_article_read": false,
      "source_url": "https://arxiv.org/pdf/2610.19789v1",
      "source_sha256": "2038e0601b8cab431720f40b3542991236e8a47f25d5a213d6c3716c1e445b80",
      "metadata_sha256": "522e17385bd2acdfa6654fc1995d8fd681b4dd9d946d643493b4179cab43a98a",
      "review_level": "independent_bounded_priority_method_evaluation_and_cost_review",
      "read_scope": "Only named original passages, not complete article/appendices.",
      "review_done": true,
      "automatic_retry": false,
      "automatic_implementation_queue": false,
      "implementation_status": "not_implemented_by_review",
      "review_artifact": "research-audits/oct15-review-closure/root-priority-reviews.json",
      "cost_assessment": {
        "source_url": "https://arxiv.org/pdf/2610.19789v1",
        "basis": "abstract study scale; experimental setup and compute accounting",
        "scale": "132 runs / 12124.4 H800 GPU-hours / four environments / three model scales",
        "available_resource": "单卡 A100 或 A30，或本地 Mac",
        "cost_type": "published_measurement_not_local_measurement"
      }
    },
    {
      "id": "2610.19951",
      "title": "AtomSearch: Evidence-Grounded Tree Search for Agent Skill Optimization",
      "published": "2026-10-14T10:33:30Z",
      "revision": 1,
      "priority": "P1",
      "decision": "reviewed_deferred",
      "resolution": "not_reproduced_high_cost",
      "reason": "AtomSearch 对证据支持的原子技能修改做 skill-state 树搜索与回归/修复记忆；单实验约923–17200 item evaluations且多个目标模型，完整搜索成本高，不复现、审查 DONE。",
      "locations": [
        "method atomic rules and search",
        "experiments evaluation budget"
      ],
      "scale": "923–17200 item evaluations per search; multiple frozen models",
      "cost_type": "published_configuration_not_local_measurement",
      "abstract_review_completed": true,
      "fulltext_review_completed": true,
      "full_article_read": false,
      "source_url": "https://arxiv.org/html/2610.19951v1",
      "source_sha256": "155ed8f9af500a4811c4930a2c849aa7a6be647e39c11188d83ac16c7d346d76",
      "metadata_sha256": "8940b6e97b8ca4ae13747eb2c3397adedea3a61594865665809ab1ebb502f671",
      "review_level": "independent_bounded_priority_method_evaluation_and_cost_review",
      "read_scope": "Only named original passages, not complete article/appendices.",
      "review_done": true,
      "automatic_retry": false,
      "automatic_implementation_queue": false,
      "implementation_status": "not_implemented_by_review",
      "review_artifact": "research-audits/oct15-review-closure/root-priority-reviews.json",
      "cost_assessment": {
        "source_url": "https://arxiv.org/html/2610.19951v1",
        "basis": "method atomic rules and search; experiments evaluation budget",
        "scale": "923–17200 item evaluations per search; multiple frozen models",
        "available_resource": "单卡 A100 或 A30，或本地 Mac",
        "cost_type": "published_configuration_not_local_measurement"
      }
    },
    {
      "id": "2610.18616",
      "title": "Knowledge Graphs as Files: An Agent Harness Where Tool Creation and Tool Use Co-Evolve",
      "published": "2026-10-13T16:00:50Z",
      "revision": 1,
      "priority": "P1",
      "decision": "reviewed_deferred",
      "resolution": "not_reproduced_high_cost",
      "reason": "DKGI 将完整 KG 作为文件，让工具创建和使用共进化；三图谱和frontier模型，每次最多300 turns/1800秒，多次评测成本高，不复现、审查 DONE。作者仓库 leopoldwhite/DKGI 已核。",
      "locations": [
        "agent harness",
        "evaluation turn/time limits"
      ],
      "scale": "STaRK three graphs, up to300 turns/1800 seconds per attempt",
      "cost_type": "published_configuration_not_local_measurement",
      "abstract_review_completed": true,
      "fulltext_review_completed": true,
      "full_article_read": false,
      "source_url": "https://arxiv.org/html/2610.18616v1",
      "source_sha256": "433d8655ee643291af99d9de728f96cf4a6865194856a1e1422d41fddbbe6587",
      "metadata_sha256": "98d4361d6f8d69e22879fd0aeedf450a61843f25b8ea51f366906b3fd22fa26d",
      "review_level": "independent_bounded_priority_method_evaluation_and_cost_review",
      "read_scope": "Only named original passages, not complete article/appendices.",
      "review_done": true,
      "automatic_retry": false,
      "automatic_implementation_queue": false,
      "implementation_status": "not_implemented_by_review",
      "review_artifact": "research-audits/oct15-review-closure/root-priority-reviews.json",
      "cost_assessment": {
        "source_url": "https://arxiv.org/html/2610.18616v1",
        "basis": "agent harness; evaluation turn/time limits",
        "scale": "STaRK three graphs, up to300 turns/1800 seconds per attempt",
        "available_resource": "单卡 A100 或 A30，或本地 Mac",
        "cost_type": "published_configuration_not_local_measurement"
      }
    },
    {
      "id": "2610.18654",
      "title": "Kwai Summary Attention V2 Technical Report",
      "published": "2026-10-13T16:18:44Z",
      "revision": 1,
      "priority": "P1",
      "decision": "reviewed_deferred",
      "resolution": "not_reproduced_high_cost",
      "reason": "KSA V2 结合 Summary→Meta KV 压缩与 summary-guided 原始块稀疏索引；1.9B/4B、128K上下文及数十亿tokens分阶段训练，成本过高，不复现、审查 DONE。",
      "locations": [
        "C-KSA",
        "S-KSA",
        "training stages"
      ],
      "scale": "1.9B/4B,128K,6.7B/3.4B/18.5B-token stages",
      "cost_type": "published_configuration_not_local_measurement",
      "abstract_review_completed": true,
      "fulltext_review_completed": true,
      "full_article_read": false,
      "source_url": "https://arxiv.org/html/2610.18654v1",
      "source_sha256": "263d27029ae4438e464d219cfdd8d89c46efcd7bc4e2d341f81c4e93df7cb079",
      "metadata_sha256": "5f754651bc4f7caf8afb383f89c08fdf52357333aa70ac1608b6a887fb91298b",
      "review_level": "independent_bounded_priority_method_evaluation_and_cost_review",
      "read_scope": "Only named original passages, not complete article/appendices.",
      "review_done": true,
      "automatic_retry": false,
      "automatic_implementation_queue": false,
      "implementation_status": "not_implemented_by_review",
      "review_artifact": "research-audits/oct15-review-closure/root-priority-reviews.json",
      "cost_assessment": {
        "source_url": "https://arxiv.org/html/2610.18654v1",
        "basis": "C-KSA; S-KSA; training stages",
        "scale": "1.9B/4B,128K,6.7B/3.4B/18.5B-token stages",
        "available_resource": "单卡 A100 或 A30，或本地 Mac",
        "cost_type": "published_configuration_not_local_measurement"
      }
    },
    {
      "id": "2610.20456",
      "title": "The COUNT: Counting N-grams Unlocks Scaling on Host Memory and Training-Free Adaptation",
      "published": "2026-10-14T16:38:19Z",
      "revision": 1,
      "priority": "P1",
      "decision": "reviewed_deferred",
      "resolution": "not_reproduced_high_cost",
      "reason": "COUNT 把 CPU n-gram 计数映射到 unembedding 输入特征；主表约797GB内存，小模型表仍约17.1GB，另有训练成本。成本过高，不下载计数表、不复现，审查 DONE。",
      "locations": [
        "method count features",
        "host memory and training setup"
      ],
      "scale": "797GB host table; smaller table17.1GB plus pretraining",
      "cost_type": "published_configuration_not_local_measurement",
      "abstract_review_completed": true,
      "fulltext_review_completed": true,
      "full_article_read": false,
      "source_url": "https://arxiv.org/html/2610.20456v1",
      "source_sha256": "11b3a80afb9109329b1c56aaca61887bdce1b1024df73edafdefdf6ce9dbfe59",
      "metadata_sha256": "4b7779dc164ae24a168e8bc3dce3ca33d5610fd7600e5f02703d0fc9512450a6",
      "review_level": "independent_bounded_priority_method_evaluation_and_cost_review",
      "read_scope": "Only named original passages, not complete article/appendices.",
      "review_done": true,
      "automatic_retry": false,
      "automatic_implementation_queue": false,
      "implementation_status": "not_implemented_by_review",
      "review_artifact": "research-audits/oct15-review-closure/root-priority-reviews.json",
      "cost_assessment": {
        "source_url": "https://arxiv.org/html/2610.20456v1",
        "basis": "method count features; host memory and training setup",
        "scale": "797GB host table; smaller table17.1GB plus pretraining",
        "available_resource": "单卡 A100 或 A30，或本地 Mac",
        "cost_type": "published_configuration_not_local_measurement"
      }
    },
    {
      "id": "2610.20399",
      "title": "Cortex: Consolidating Context into Parametric Memory",
      "published": "2026-10-14T16:03:01Z",
      "revision": 1,
      "priority": "P1",
      "decision": "reviewed_deferred",
      "resolution": "not_reproduced_high_cost",
      "reason": "Cortex 用情景查询与上下文条件蒸馏将外部上下文写入参数记忆；4B–120B、多批次/顺序编辑。完整训练/编辑成本高，不复现、审查 DONE；HTML不可用后已取得官方PDF并核指定方法/资源段落，不把有界核对冒称整篇通读。",
      "locations": [
        "abstract model/editing scale",
        "official PDF fallback resource check"
      ],
      "scale": "4B–120B, batch/sequential parametric-memory editing",
      "cost_type": "published_configuration_not_local_measurement",
      "abstract_review_completed": true,
      "fulltext_review_completed": true,
      "full_article_read": false,
      "source_url": "https://arxiv.org/pdf/2610.20399v1",
      "source_sha256": "6123bd40b6572a520050e8af36982e1d3471a5fb2f85e168aa313d549b93752c",
      "metadata_sha256": "abc57665d700f2a1aa6c421e892df31c8c3b47eca06148add8295ed8097c66a8",
      "review_level": "independent_bounded_priority_method_evaluation_and_cost_review",
      "read_scope": "Only named original passages, not complete article/appendices.",
      "review_done": true,
      "automatic_retry": false,
      "automatic_implementation_queue": false,
      "implementation_status": "not_implemented_by_review",
      "review_artifact": "research-audits/oct15-review-closure/root-priority-reviews.json",
      "cost_assessment": {
        "source_url": "https://arxiv.org/pdf/2610.20399v1",
        "basis": "abstract model/editing scale; official PDF fallback resource check",
        "scale": "4B–120B, batch/sequential parametric-memory editing",
        "available_resource": "单卡 A100 或 A30，或本地 Mac",
        "cost_type": "published_configuration_not_local_measurement"
      }
    },
    {
      "id": "2610.18883",
      "title": "Token Utility for On-Policy Distillation via Local Trust-Region Optimization",
      "published": "2026-10-01T20:01:20Z",
      "revision": 1,
      "priority": "P1",
      "decision": "reviewed_deferred",
      "resolution": "not_reproduced_high_cost",
      "reason": "FUW/Token Utility 用 Fisher 几何的 token utility 调整蒸馏权重；1.7B/4B学生、8B/14B/30B教师，279步×256序列×8192上下文，成本过高不复现、审查 DONE。arXiv首发2026-10-01，必须记旧文补漏而非10月14日新首发。",
      "locations": [
        "method Fisher utility weighting",
        "experimental setup"
      ],
      "scale": "1.7B/4B students,8B/14B/30B teachers,279×256×8192",
      "cost_type": "published_configuration_not_local_measurement",
      "abstract_review_completed": true,
      "fulltext_review_completed": true,
      "full_article_read": false,
      "source_url": "https://arxiv.org/html/2610.18883v1",
      "source_sha256": "e91a7e761f5d7f6f895865743358573bb38d658c886e562d92575ee572787845",
      "metadata_sha256": "b830af9574069d7fdb60ba06114f2d78d639a3a1c584905ef7d70ef676688959",
      "review_level": "independent_bounded_priority_method_evaluation_and_cost_review",
      "read_scope": "Only named original passages, not complete article/appendices.",
      "review_done": true,
      "automatic_retry": false,
      "automatic_implementation_queue": false,
      "implementation_status": "not_implemented_by_review",
      "review_artifact": "research-audits/oct15-review-closure/root-priority-reviews.json",
      "cost_assessment": {
        "source_url": "https://arxiv.org/html/2610.18883v1",
        "basis": "method Fisher utility weighting; experimental setup",
        "scale": "1.7B/4B students,8B/14B/30B teachers,279×256×8192",
        "available_resource": "单卡 A100 或 A30，或本地 Mac",
        "cost_type": "published_configuration_not_local_measurement"
      }
    },
    {
      "id": "2610.19844",
      "title": "Memory-Efficient Subspace Optimization with Residual Correction: From Momentum SGD to Muon",
      "published": "2026-10-14T09:12:57Z",
      "revision": 1,
      "priority": "P1",
      "decision": "reviewed_deferred",
      "resolution": "not_reproduced_high_cost",
      "reason": "SMRC/PMRC 保留子空间外残差并修正换基边界；原文60M/150M用2GPU、350M用4GPU、1.3B用8GPU预训练。完整多规模实证成本高，不复现、审查 DONE；不把单次数值测试冒称论文能力验证。",
      "locations": [
        "method residual/basis correction",
        "§6.1 LLM pretraining"
      ],
      "scale": "2–8 GPUs;1.3B/26.84B-token original configuration",
      "cost_type": "published_configuration_not_local_measurement",
      "abstract_review_completed": true,
      "fulltext_review_completed": true,
      "full_article_read": false,
      "source_url": "https://arxiv.org/html/2610.19844v1",
      "source_sha256": "9113da882de1534c246d302a71f6901e401baaa64e9384d36909e520b6ea498d",
      "metadata_sha256": "7a5ba60333760b25e4bd7e8f0f682869cccd87431ca9a8975d336de9eae21dbb",
      "review_level": "independent_bounded_priority_method_evaluation_and_cost_review",
      "read_scope": "Only named original passages, not complete article/appendices.",
      "review_done": true,
      "automatic_retry": false,
      "automatic_implementation_queue": false,
      "implementation_status": "not_implemented_by_review",
      "review_artifact": "research-audits/oct15-review-closure/root-priority-reviews.json",
      "cost_assessment": {
        "source_url": "https://arxiv.org/html/2610.19844v1",
        "basis": "method residual/basis correction; §6.1 LLM pretraining",
        "scale": "2–8 GPUs;1.3B/26.84B-token original configuration",
        "available_resource": "单卡 A100 或 A30，或本地 Mac",
        "cost_type": "published_configuration_not_local_measurement"
      }
    },
    {
      "id": "2610.19605",
      "title": "PN-Mix: Pair-Norm Guided Mixed-Precision Key Cache Quantization for Efficient Long-Output LLM Inference",
      "published": "2026-10-14T05:29:08Z",
      "revision": 1,
      "priority": "P1",
      "decision": "reviewed_deferred",
      "resolution": "not_reproduced_high_cost",
      "reason": "PN-Mix 用 prefill Q/K RoPE 通道对范数分配混合精度、配合1-bit残差与sink保护；原文多模型8B–32B、长输出多任务评测的整体成本未能廉价核定，按预算门槛不复现、审查 DONE。Kitty 链接是基线，不冒称作者代码。",
      "locations": [
        "method pair norm and residual correction",
        "hardware/software",
        "§6.4 inference efficiency"
      ],
      "scale": "8B–32B, long outputs up to32K, multi-model/multi-task sweep; singleA10040GB timings are only one slice",
      "cost_type": "unresolved_material_cost_after_primary_source_check",
      "abstract_review_completed": true,
      "fulltext_review_completed": true,
      "full_article_read": false,
      "source_url": "https://arxiv.org/html/2610.19605v1",
      "source_sha256": "592d70ee87547babba026b45ecef3dcd18af4397e784dc385bb02b09c1257f9f",
      "metadata_sha256": "282e663010c24c86dbb11a506cc3427f9dfcb5c9ced57dbb9f209c3bfef0299f",
      "review_level": "independent_bounded_priority_method_evaluation_and_cost_review",
      "read_scope": "Only named original passages, not complete article/appendices.",
      "review_done": true,
      "automatic_retry": false,
      "automatic_implementation_queue": false,
      "implementation_status": "not_implemented_by_review",
      "review_artifact": "research-audits/oct15-review-closure/root-priority-reviews.json",
      "cost_assessment": {
        "source_url": "https://arxiv.org/html/2610.19605v1",
        "basis": "method pair norm and residual correction; hardware/software; §6.4 inference efficiency",
        "scale": "8B–32B, long outputs up to32K, multi-model/multi-task sweep; singleA10040GB timings are only one slice",
        "available_resource": "单卡 A100 或 A30，或本地 Mac",
        "cost_type": "unresolved_material_cost_after_primary_source_check"
      }
    },
    {
      "id": "2610.19591",
      "title": "L3M: A Verifiable Harness and Runtime for Self-Extending and Self-Improving Agents",
      "published": "2026-10-14T05:03:06Z",
      "revision": 1,
      "priority": "P1",
      "decision": "reviewed_deferred",
      "resolution": "not_reproduced_high_cost",
      "reason": "L3M 包含 Git 状态轨迹、能力收窄授权和 Lean 内核检查的运行时；未核定可直接复用的完整上游实现，Terminal-Bench全套调用成本也未能廉价确定。按预算门槛不复现、审查 DONE；不能用Python状态机假装Lean证明。",
      "locations": [
        "harness/runtime design",
        "kernel-checked theorem boundary",
        "Terminal-Bench evaluation"
      ],
      "scale": "long-horizon terminal agent evaluation and Lean runtime integration",
      "cost_type": "unresolved_material_cost_and_source_access_after_primary_source_check",
      "abstract_review_completed": true,
      "fulltext_review_completed": true,
      "full_article_read": false,
      "source_url": "https://arxiv.org/html/2610.19591v1",
      "source_sha256": "315f66eef2c36deb0ea45924fe2b26857f7eaa8c2703dd68210c9ae4bffde6cf",
      "metadata_sha256": "6c4f6e6439ee7dca0b7c526975b54dc09c813983393c8f1f4649cb1c356ed675",
      "review_level": "independent_bounded_priority_method_evaluation_and_cost_review",
      "read_scope": "Only named original passages, not complete article/appendices.",
      "review_done": true,
      "automatic_retry": false,
      "automatic_implementation_queue": false,
      "implementation_status": "not_implemented_by_review",
      "review_artifact": "research-audits/oct15-review-closure/root-priority-reviews.json",
      "cost_assessment": {
        "source_url": "https://arxiv.org/html/2610.19591v1",
        "basis": "harness/runtime design; kernel-checked theorem boundary; Terminal-Bench evaluation",
        "scale": "long-horizon terminal agent evaluation and Lean runtime integration",
        "available_resource": "单卡 A100 或 A30，或本地 Mac",
        "cost_type": "unresolved_material_cost_and_source_access_after_primary_source_check"
      }
    },
    {
      "id": "2610.20586",
      "title": "Continuum Attention",
      "published": "2026-10-14T17:57:40Z",
      "revision": 1,
      "priority": "P1",
      "decision": "reviewed_deferred",
      "resolution": "not_reproduced_high_cost",
      "reason": "Continuum Attention 在连续分段线性K/V曲线上用闭式指数矩精确积分，含半token边界与解析梯度；研究共80训练运行、每模型3.84M tokens、5种子。完整多臂Metal复现总时长不能廉价核定，按保守成本门槛不复现、审查 DONE；不拿便宜算子单测顶替研究结论。",
      "locations": [
        "§2.1–2.7 exact integral and stable moments",
        "optimization protocol",
        "§3.5 performance"
      ],
      "scale": "80 completed runs,8 configurations,5 seeds,2 context lengths,3.84M tokens/model",
      "cost_type": "unresolved_total_campaign_cost_after_primary_source_check",
      "abstract_review_completed": true,
      "fulltext_review_completed": true,
      "full_article_read": false,
      "source_url": "https://arxiv.org/html/2610.20586v1",
      "source_sha256": "bf2c493e27d779bf8b7dbe44ae5d3a2faaa5bb0401f417584edf9ae9a8254ccf",
      "metadata_sha256": "72716c3fbfec3c4d1fb934979ff1e6c7d1459bd6042b04dcb6776ab3f5646fc6",
      "review_level": "independent_bounded_priority_method_evaluation_and_cost_review",
      "read_scope": "Only named original passages, not complete article/appendices.",
      "review_done": true,
      "automatic_retry": false,
      "automatic_implementation_queue": false,
      "implementation_status": "not_implemented_by_review",
      "review_artifact": "research-audits/oct15-review-closure/root-priority-reviews.json",
      "cost_assessment": {
        "source_url": "https://arxiv.org/html/2610.20586v1",
        "basis": "§2.1–2.7 exact integral and stable moments; optimization protocol; §3.5 performance",
        "scale": "80 completed runs,8 configurations,5 seeds,2 context lengths,3.84M tokens/model",
        "available_resource": "单卡 A100 或 A30，或本地 Mac",
        "cost_type": "unresolved_total_campaign_cost_after_primary_source_check"
      }
    },
    {
      "id": "2610.17727",
      "title": "SkillScreen: High-Recall Screening of Malicious Agent Skills with Decision Models",
      "published": "2026-10-13T05:03:43Z",
      "revision": 2,
      "priority": "P1",
      "decision": "reviewed_deferred",
      "resolution": "not_reproduced_high_cost",
      "reason": "SkillScreen 用2B决策模型的NLI分数筛查技能，以验证集冻结99%召回阈值；必须保留陌生良性来源FPR升高与仅扫描SKILL.md的限制。v2明确每次单epoch仍需3.4–3.9小时（10GB H100 MIG），还有多划分/多种子；成本过高不复现、审查 DONE，不能因显存小就当便宜。",
      "locations": [
        "§3 NLI scoring",
        "§4 split protocols",
        "Appendix optimization/hardware",
        "limitations"
      ],
      "scale": "2B,6.8K labelled skills,426/470 updates,3.4–3.9h per run; multiple splits/seeds",
      "cost_type": "published_measurement_not_local_measurement",
      "abstract_review_completed": true,
      "fulltext_review_completed": true,
      "full_article_read": false,
      "source_url": "https://arxiv.org/html/2610.17727v2",
      "source_sha256": "0c8e07447c3a70d3ed6a91387b5da8e26aca3ccd4441973d94662af630588640",
      "metadata_sha256": "6703f15479548dc2d5cc7897797b6b8301fde5d07fa41c4036c109aa22caa6b2",
      "review_level": "independent_bounded_priority_method_evaluation_and_cost_review",
      "read_scope": "Only named original passages, not complete article/appendices.",
      "review_done": true,
      "automatic_retry": false,
      "automatic_implementation_queue": false,
      "implementation_status": "not_implemented_by_review",
      "review_artifact": "research-audits/oct15-review-closure/root-priority-reviews.json",
      "cost_assessment": {
        "source_url": "https://arxiv.org/html/2610.17727v2",
        "basis": "§3 NLI scoring; §4 split protocols; Appendix optimization/hardware; limitations",
        "scale": "2B,6.8K labelled skills,426/470 updates,3.4–3.9h per run; multiple splits/seeds",
        "available_resource": "单卡 A100 或 A30，或本地 Mac",
        "cost_type": "published_measurement_not_local_measurement"
      }
    },
    {
      "id": "2610.18140",
      "title": "SparPR: Runtime KV-Cache Recovery for Resource-Constrained LLM Inference on Consumer Hardware",
      "published": "2026-10-13T10:47:05Z",
      "revision": 1,
      "priority": "P1",
      "decision": "reviewed_deferred",
      "resolution": "not_reproduced_high_cost",
      "reason": "SparPR 用 query-aware 块选择、动态预算与按需局部重算恢复KV；虽然仅8GB笔记本，原文8K负载单次仍需44分钟，完整多长度/比例/消融会显著增长。按完整实验成本门槛不复现、审查 DONE，不启动小探针证明它慢。",
      "locations": [
        "§3.2–3.5 recovery decisions",
        "§4.1 setup",
        "§4.2 system performance"
      ],
      "scale": "OPT-6.7B offloaded inference,batch4,8K workload44min plus sweep/ablations",
      "cost_type": "published_measurement_and_total_campaign_estimate_not_local_measurement",
      "abstract_review_completed": true,
      "fulltext_review_completed": true,
      "full_article_read": false,
      "source_url": "https://arxiv.org/html/2610.18140v1",
      "source_sha256": "236a73aab06f7af8fdd0338a187aeeebb109ed053a44e38b7af83552f07a652e",
      "metadata_sha256": "4e1b1527627c35850dc4339bc8727dfde8d5145f9f594aacc474c15cbec74541",
      "review_level": "independent_bounded_priority_method_evaluation_and_cost_review",
      "read_scope": "Only named original passages, not complete article/appendices.",
      "review_done": true,
      "automatic_retry": false,
      "automatic_implementation_queue": false,
      "implementation_status": "not_implemented_by_review",
      "review_artifact": "research-audits/oct15-review-closure/root-priority-reviews.json",
      "cost_assessment": {
        "source_url": "https://arxiv.org/html/2610.18140v1",
        "basis": "§3.2–3.5 recovery decisions; §4.1 setup; §4.2 system performance",
        "scale": "OPT-6.7B offloaded inference,batch4,8K workload44min plus sweep/ablations",
        "available_resource": "单卡 A100 或 A30，或本地 Mac",
        "cost_type": "published_measurement_and_total_campaign_estimate_not_local_measurement"
      }
    },
    {
      "id": "2610.19139",
      "title": "Online Tokenization with Optimal Delay in Linear Time",
      "published": "2026-10-13T20:08:40Z",
      "revision": 1,
      "priority": "P1",
      "decision": "accepted",
      "resolution": "affordable_candidate_not_implemented",
      "reason": "在线tokenization以稳定输出前缀与delay形式化为基础，组合special-token、pretokenizer和BPE，并给出更低delay的联合算法与不可解边界；公开Morgan Stanley上游实现可做CPU级精确ID/流式等价性验收。接受为低成本算法候选，本轮不实现、不跑完整吞吐研究。",
      "locations": [
        "§1.1–1.6",
        "§2–§5 streaming/stable-prefix/delay",
        "§6–§7 implementation and validation"
      ],
      "scale": "CPU streaming tokenizer, R50K/CL100K/O200K,public upstream implementation",
      "cost_type": "bounded_cpu_mechanism_review_not_runtime_measurement",
      "abstract_review_completed": true,
      "fulltext_review_completed": true,
      "full_article_read": false,
      "source_url": "https://arxiv.org/html/2610.19139v1",
      "source_sha256": "9c31536f09d15157267db79a34838a7e7b717acc5000609fcb986a0c17ddeef1",
      "metadata_sha256": "fd38b34f9b0e6d662a56efdfeb935bd0a366ffee906eb5ec3764c79b87f4cd50",
      "review_level": "independent_bounded_priority_method_evaluation_and_cost_review",
      "read_scope": "Only named original passages, not complete article/appendices.",
      "review_done": true,
      "automatic_retry": false,
      "automatic_implementation_queue": false,
      "implementation_status": "not_implemented_by_review",
      "review_artifact": "research-audits/oct15-review-closure/root-priority-reviews.json"
    },
    {
      "id": "2610.19375",
      "title": "One-Step Prediction Fidelity Is a Poor Proxy for Planning Performance in JEPA World Models",
      "published": "2026-10-14T01:19:50Z",
      "revision": 1,
      "priority": "P1",
      "decision": "reviewed_deferred",
      "resolution": "not_reproduced_high_cost",
      "reason": "JEPA prediction-vs-planning 研究采用冻结物理状态decoder、配对规划种子、分离编码/预测误差及匹配继续训练对照；可参考其评测协议，但三环境、模型梯度阶梯与重复训练的完整总成本未能廉价核定。不重训、不复现，审查 DONE；不将协议建议当已接入evolve。",
      "locations": [
        "§3 frozen-decoder protocol",
        "§4 experimental setup",
        "§6–§8 controls and confounds"
      ],
      "scale": "three environments,decoder training,checkpoint ladders,paired planning/training seeds",
      "cost_type": "unresolved_total_campaign_cost_after_primary_source_check",
      "abstract_review_completed": true,
      "fulltext_review_completed": true,
      "full_article_read": false,
      "source_url": "https://arxiv.org/html/2610.19375v1",
      "source_sha256": "630c89199ddc5b71e784021b310106e531f48c9eb9dc5783d5b2be8d074cfdf3",
      "metadata_sha256": "8b5dd7a3e51cd9c5a89f3aa66f6f83f074cb3859927de7bf2aa89866ec0b3a13",
      "review_level": "independent_bounded_priority_method_evaluation_and_cost_review",
      "read_scope": "Only named original passages, not complete article/appendices.",
      "review_done": true,
      "automatic_retry": false,
      "automatic_implementation_queue": false,
      "implementation_status": "not_implemented_by_review",
      "review_artifact": "research-audits/oct15-review-closure/root-priority-reviews.json",
      "cost_assessment": {
        "source_url": "https://arxiv.org/html/2610.19375v1",
        "basis": "§3 frozen-decoder protocol; §4 experimental setup; §6–§8 controls and confounds",
        "scale": "three environments,decoder training,checkpoint ladders,paired planning/training seeds",
        "available_resource": "单卡 A100 或 A30，或本地 Mac",
        "cost_type": "unresolved_total_campaign_cost_after_primary_source_check"
      }
    },
    {
      "id": "2610.20294",
      "title": "MMLA-RSI: Recursive Self-Improvement through RTT and RTMA",
      "published": "2026-10-14T14:47:50Z",
      "revision": 1,
      "priority": "P2",
      "decision": "reviewed_deferred",
      "resolution": "reviewed_boundary_not_capability",
      "reason": "MMLA-RSI 描述 RTT/RTMA 受限规则语言、trusted committer与独立接纳；原文明确有益递归链和发现效率尚待评估。保留为RSI构造/边界材料，不登记为已验证自进化收益，也不生成执行型operator。",
      "locations": [
        "RTT/RTMA construction",
        "evaluation and limitations"
      ],
      "abstract_review_completed": true,
      "fulltext_review_completed": true,
      "full_article_read": false,
      "source_url": "https://arxiv.org/html/2610.20294v1",
      "source_sha256": "80464e88a4d6ea184c095dc5aff0aeb15d2cd8d789586401b9eb94cbe753c0bd",
      "metadata_sha256": "ab1c2c90c66c6e535c17c0517e0c5e953df08910bf9e985caa44b5985e7c4dc8",
      "review_level": "independent_bounded_priority_method_evaluation_and_cost_review",
      "read_scope": "Only named original passages, not complete article/appendices.",
      "review_done": true,
      "automatic_retry": false,
      "automatic_implementation_queue": false,
      "implementation_status": "not_implemented_by_review",
      "review_artifact": "research-audits/oct15-review-closure/root-priority-reviews.json"
    }
  ]
}