{
  "schema_version": 1,
  "batch_id": "google39-primary-closeout-20261021",
  "snapshot_created_at_utc": "2026-10-20T18:57:03Z",
  "scope": {
    "source": "../historical-backlog-reconciliation-20261021/reconciliation.json#/current_2026_google_review_leads",
    "expected_unique_ids": 39,
    "reviewed_unique_ids": 39,
    "id_set_sha256": "8623baeb7c2e1132ce70ede629aef201b24ab6cbdfa5038e9848acb7beae6f3b",
    "id_set_hash_encoding": "UTF-8 IDs in frozen source order, one ID per line with final newline",
    "all_497_backlog_reviewed": false,
    "overall_review_watermark_updated": false,
    "implementation_scope": "No implementations in this 39-paper audit; only SpecGuard and Video Encoder Selection are authorized elsewhere in the parent batch."
  },
  "input_sha256": {
    "google_review": "9fe69fc47bf55fd1a270e4b0353133669435c10a2ce36fb42098db3385654697",
    "reconciliation": "e658dec8fbca8e2a98fca612bb00565dfbb6778f5bec3c9b79f79dfb2f3e6d9a"
  },
  "summary": {
    "accepted_core_not_implemented": 16,
    "deferred_full_reproduction_terminal": 14,
    "rejected_scope": 9,
    "review_done": 39,
    "implemented": 0,
    "automatic_implementation_queue": 0,
    "unread_fulltext_source_exceptions": 0
  },
  "semantics_zh": [
    "review DONE是本轮有界审查完成，不等于逐段通读、算法已实现或实验已验证。",
    "accepted只指有实质算法/评测核心且可负担的候选机会，不授权本轮实施，也不豁免工业论文真实A/B/明确全流量门槛。",
    "完整实验高成本/私有资源终态与低成本核心候选分轴；不构造廉价fixture代替论文定义算法。",
    "无作者代码链接时只记已读主源未找到，不声称全网不存在或作者确认未发布。",
    "source_sha256是公开URL/定位/UTC/中文转述摘要回执的哈希；原始浏览捕获仅保存在本轮私有临时目录，不提交，公开回执以original_capture_sha256绑定原捕获。两类哈希均不伪称完整HTTP原文哈希。"
  ],
  "retrieval_failures": [
    {
      "source_url": "https://arxiv.org/html/2602.07918v1",
      "method": "terminal curl -IL --max-time 20",
      "observed_result": "curl: (6) Could not resolve host: arxiv.org",
      "resolution": "official browser retrieval succeeded; not a paper-level source exception",
      "utc_window": "2026-10-20T18:48:00Z/2026-10-20T18:49:00Z"
    }
  ],
  "records": [
    {
      "id": "2601.00372",
      "title": "LLM-Powered Analysis of IoT User Reviews: Tracking and Ranking Security and Privacy Concerns",
      "track": "recommendation/llm-application",
      "status": "rejected_scope",
      "review_decision": "rejected_scope",
      "review_done": true,
      "resolution": "out_of_scope",
      "implementation_status": "not_implemented",
      "implementation_authorized_this_run": false,
      "automatic_implementation_queue": false,
      "automatic_retry": false,
      "evolve_integrated": false,
      "promotion_eligible": false,
      "review_decision_scope": "paper_method_scope",
      "full_experiment_resolution": "not_applicable_scope",
      "full_experiment_review_done": true,
      "arxiv_v1_date": "2026-01-01",
      "date_evidence": "official_arxiv_abs_or_prior_exact_identity_date",
      "earlier_public_release_exhaustively_checked": false,
      "authorship": {
        "first_author_institution_or_role": "North Carolina State University",
        "google_priority_basis": "current_arxiv_author_block_has_google_affiliation_or_collaboration",
        "first_author_google_required": false
      },
      "method_summary_zh": "GPT-3.5-Turbo两阶段SFT：CRC分类/解释/低层分类，TM主题映射，动态few-shot；属于IoT评论安全隐私主题实证，不提出通用推荐/Agent/基础模型机制。",
      "experiment_summary_zh": "91,149条无标注Amazon评论；2,454条人工转换标签；SFT 3 epochs，Linux16GB RAM与RTX3070M用于BERT embedding；正文结果是离线分类/主题统计。",
      "public_code": {
        "status": "project_or_data_link_only_no_algorithm_code_confirmed",
        "url": "https://anonymous.4open.science/r/IoT-Products-SecurityAndPrivacy-5F5D/README.md",
        "notes_zh": "正文给出匿名IoT-Products-SecurityAndPrivacy仓库，明确产品信息/数据；未核实算法训练代码已发布。",
        "search_scope": "reviewed arXiv v1 article text, code/repository mentions and linked primary repository when explicitly provided; no exhaustive web-wide absence claim"
      },
      "industrial_online_evidence": {
        "strict_online_ab_confirmed": false,
        "explicit_full_traffic_confirmed": false,
        "notes_zh": "未报告真实生产A/B或全流量；离线Amazon评论不是线上实验。"
      },
      "full_text_review": {
        "scope": "bounded_primary_fulltext_method_experiment_resource_and_release_review",
        "source_url": "https://arxiv.org/html/2601.00372v1",
        "read_sections": [
          "§3 Methodology",
          "§4 Results导航/摘要指标",
          "正文作者/日期"
        ],
        "source_locations": "L119-L145",
        "full_article_read": false,
        "proofs_read_in_full": false,
        "all_tables_read": false,
        "latest_revision_reviewed": false,
        "version": "v1",
        "limitations": "只审读列明关键段与检索返回内容，不声称逐段通读、穷尽全部附录或确认最新修订。"
      },
      "cost_assessment": {
        "available_resource": "one NVIDIA A100 or A30, or local Mac; no multi-GPU assumption",
        "paper_configuration_and_scale": "91,149条无标注Amazon评论；2,454条人工转换标签；SFT 3 epochs，Linux16GB RAM与RTX3070M用于BERT embedding；正文结果是离线分类/主题统计。",
        "decision_zh": "不启动91K评论批量API与SFT；本项scope终态，不以成本代替方法审查。",
        "basis": "reviewed primary-paper configuration, directly reported resource/runtime where available, otherwise qualitative bounded estimate",
        "local_cost_measurement_performed": false,
        "training_or_model_download_performed": false
      },
      "source_url": "https://arxiv.org/html/2601.00372v1",
      "source_sha256": "144665c0eee9b8eca2f7554bc52ef4d8a4de6c1a76d08feabc8e7549a8ce0f3e",
      "source_hash_kind": "sha256_of_public_metadata_locations_and_chinese_paraphrase_receipt_not_raw_capture_or_article",
      "source_retrieval_reference": "evidence/google39/2601.00372.json",
      "retrieval_utc": {
        "start": "2026-10-20T18:48:00Z",
        "end": "2026-10-20T18:56:04Z",
        "precision": "bounded_session_capture_window_not_per_request_timestamp"
      },
      "official_identity_sources": [
        "https://research.google/pubs/llm-powered-analysis-of-iot-user-reviews-tracking-and-ranking-security-and-privacy-concerns/"
      ],
      "original_capture_sha256": "c12fd7008447e0dff827264972eb33e7c73a7868d8a6c41e3574c10fa58a60cf"
    },
    {
      "id": "2601.19066",
      "title": "Dynamic Cogeneration of Bug Reproduction Test in Agentic Program Repair",
      "track": "agent",
      "status": "accepted",
      "review_decision": "accepted",
      "review_done": true,
      "resolution": "accepted_core_not_implemented_this_run",
      "implementation_status": "not_implemented",
      "implementation_authorized_this_run": false,
      "automatic_implementation_queue": false,
      "automatic_retry": false,
      "evolve_integrated": false,
      "promotion_eligible": false,
      "review_decision_scope": "bounded_core_or_evaluation_candidate_only_not_industrial_admission",
      "full_experiment_resolution": "not_reproduced_high_cost",
      "full_experiment_review_done": true,
      "arxiv_v1_date": "2026-01-27",
      "date_evidence": "official_arxiv_abs_or_prior_exact_identity_date",
      "earlier_public_release_exhaustively_checked": false,
      "authorship": {
        "first_author_institution_or_role": "Google",
        "google_priority_basis": "current_arxiv_author_block_has_google_affiliation_or_collaboration",
        "first_author_google_required": false
      },
      "method_summary_zh": "Passerine中TDD/TLD/Freeform联合生成fix+BRT；BRT在原bug失败、修后通过验证；test-aware patch selection结合reviewer过滤、归一化分组及多数选择。",
      "experiment_summary_zh": "120个Google私有bug、六语言、5配置，每bug20条trajectory、每条25步，Gemini2.5 Pro；内部部署与人工代码审查被描述，但对照是离线重放，不是量化生产A/B。",
      "public_code": {
        "status": "not_found_in_reviewed_primary_sources",
        "url": null,
        "notes_zh": "未在已读正文找到作者公开Passerine/私有bug代码。",
        "search_scope": "reviewed arXiv v1 article text, code/repository mentions and linked primary repository when explicitly provided; no exhaustive web-wide absence claim"
      },
      "industrial_online_evidence": {
        "strict_online_ab_confirmed": false,
        "explicit_full_traffic_confirmed": false,
        "notes_zh": "正文明确Passerine内部部署；未找到本方法的量化真实生产A/B或明确全流量采纳。"
      },
      "full_text_review": {
        "scope": "bounded_primary_fulltext_method_experiment_resource_and_release_review",
        "source_url": "https://arxiv.org/html/2601.19066v1",
        "read_sections": [
          "§2.1 Our Agentic APR System",
          "§3 Dynamic Cogeneration",
          "§4.2配置",
          "§5结果摘要"
        ],
        "source_locations": "L88-L139;L256;L329",
        "full_article_read": false,
        "proofs_read_in_full": false,
        "all_tables_read": false,
        "latest_revision_reviewed": false,
        "version": "v1",
        "limitations": "只审读列明关键段与检索返回内容，不声称逐段通读、穷尽全部附录或确认最新修订。"
      },
      "cost_assessment": {
        "available_resource": "one NVIDIA A100 or A30, or local Mac; no multi-GPU assumption",
        "paper_configuration_and_scale": "120个Google私有bug、六语言、5配置，每bug20条trajectory、每条25步，Gemini2.5 Pro；内部部署与人工代码审查被描述，但对照是离线重放，不是量化生产A/B。",
        "decision_zh": "完整5×120×20=12,000轨迹、最多300,000 agent步及私有monorepo测试，明确超出单卡/Mac短验预算；估算基于步数上界，不是实测。小规模真实patch/BRT选择核心可负担，未获本轮实施授权。",
        "basis": "reviewed primary-paper configuration, directly reported resource/runtime where available, otherwise qualitative bounded estimate",
        "local_cost_measurement_performed": false,
        "training_or_model_download_performed": false
      },
      "source_url": "https://arxiv.org/html/2601.19066v1",
      "source_sha256": "a3556e9374f3ded4cf6da3efccb95d34c58491866202094acc694d6d4df316d5",
      "source_hash_kind": "sha256_of_public_metadata_locations_and_chinese_paraphrase_receipt_not_raw_capture_or_article",
      "source_retrieval_reference": "evidence/google39/2601.19066.json",
      "retrieval_utc": {
        "start": "2026-10-20T18:48:00Z",
        "end": "2026-10-20T18:56:04Z",
        "precision": "bounded_session_capture_window_not_per_request_timestamp"
      },
      "official_identity_sources": [
        "https://research.google/pubs/dynamic-cogeneration-of-bug-reproduction-test-in-agentic-program-repair/"
      ],
      "original_capture_sha256": "8e9ae5f473a945c0780e42ebb52ab5514fea16aaaa6bf0bf61eee711fd2c2fb6"
    },
    {
      "id": "2601.21149",
      "title": "Mobility-Embedded POIs: Learning What a Place Is and How It’s Used from Human Movement",
      "track": "foundation-model",
      "status": "accepted",
      "review_decision": "accepted",
      "review_done": true,
      "resolution": "accepted_core_not_implemented_this_run",
      "implementation_status": "not_implemented",
      "implementation_authorized_this_run": false,
      "automatic_implementation_queue": false,
      "automatic_retry": false,
      "evolve_integrated": false,
      "promotion_eligible": false,
      "review_decision_scope": "bounded_core_or_evaluation_candidate_only_not_industrial_admission",
      "full_experiment_resolution": "not_reproduced_private_data",
      "full_experiment_review_done": true,
      "arxiv_v1_date": "2026-01-29",
      "date_evidence": "official_arxiv_abs_or_prior_exact_identity_date",
      "earlier_public_release_exhaustively_checked": false,
      "authorship": {
        "first_author_institution_or_role": "University of Southern California",
        "google_priority_basis": "current_arxiv_author_block_has_google_affiliation_or_collaboration",
        "first_author_google_required": false
      },
      "method_summary_zh": "visit Transformer与POI原型对比学习；多尺度Gaussian anchor分布迁移＋KL，anchor直接监督与text cosine alignment；冻结表示后轻量头。",
      "experiment_summary_zh": "Veraset LA一年/Houston20天私有移动数据；53.7M参数、4层8头d512；20 epoch预训练、100 epoch微调早停，V100 16GB；5项地图属性probe。",
      "public_code": {
        "status": "not_found_in_reviewed_primary_sources",
        "url": null,
        "notes_zh": "已读正文/附录未找到作者训练实现仓库。",
        "search_scope": "reviewed arXiv v1 article text, code/repository mentions and linked primary repository when explicitly provided; no exhaustive web-wide absence claim"
      },
      "industrial_online_evidence": {
        "strict_online_ab_confirmed": false,
        "explicit_full_traffic_confirmed": false,
        "notes_zh": "5个离线地图任务，不是线上A/B；未找到明确生产全流量。"
      },
      "full_text_review": {
        "scope": "bounded_primary_fulltext_method_experiment_resource_and_release_review",
        "source_url": "https://arxiv.org/html/2601.21149v1",
        "read_sections": [
          "§3.3-3.6",
          "§4.1",
          "Appendix A.1.4-A.3"
        ],
        "source_locations": "L125-L177;L406-L418",
        "full_article_read": false,
        "proofs_read_in_full": false,
        "all_tables_read": false,
        "latest_revision_reviewed": false,
        "version": "v1",
        "limitations": "只审读列明关键段与检索返回内容，不声称逐段通读、穷尽全部附录或确认最新修订。"
      },
      "cost_assessment": {
        "available_resource": "one NVIDIA A100 or A30, or local Mac; no multi-GPU assumption",
        "paper_configuration_and_scale": "Veraset LA一年/Houston20天私有移动数据；53.7M参数、4层8头d512；20 epoch预训练、100 epoch微调早停，V100 16GB；5项地图属性probe。",
        "decision_zh": "模型单卡可放下；原数据为付费/私有移动轨迹与Safegraph，原尺度训练与多个embedding API成本未量化，不能承诺完整复现。可负担有界机制候选，未实施；不把V100使用本身视为高成本证明。",
        "basis": "reviewed primary-paper configuration, directly reported resource/runtime where available, otherwise qualitative bounded estimate",
        "local_cost_measurement_performed": false,
        "training_or_model_download_performed": false
      },
      "source_url": "https://arxiv.org/html/2601.21149v1",
      "source_sha256": "bf0a285cb148b3b6ea24f6da6d5d7f25e595293085374a35faad4c1dea9ee2a4",
      "source_hash_kind": "sha256_of_public_metadata_locations_and_chinese_paraphrase_receipt_not_raw_capture_or_article",
      "source_retrieval_reference": "evidence/google39/2601.21149.json",
      "retrieval_utc": {
        "start": "2026-10-20T18:48:00Z",
        "end": "2026-10-20T18:56:04Z",
        "precision": "bounded_session_capture_window_not_per_request_timestamp"
      },
      "official_identity_sources": [
        "https://research.google/pubs/mobility-embedded-pois-learning-what-a-place-is-and-how-its-used-from-human-movement/"
      ],
      "original_capture_sha256": "a0d1323453538be224d3f721cac725f239d0df11332dffa7b0e1caad771b6c23"
    },
    {
      "id": "2602.02164",
      "title": "Co-RedTeam: Orchestrated Security Discovery and Exploitation with LLM Agents",
      "track": "agent",
      "status": "deferred",
      "review_decision": "deferred",
      "review_done": true,
      "resolution": "not_reproduced_high_cost",
      "implementation_status": "not_implemented",
      "implementation_authorized_this_run": false,
      "automatic_implementation_queue": false,
      "automatic_retry": false,
      "evolve_integrated": false,
      "promotion_eligible": false,
      "review_decision_scope": "full_paper_reproduction_closed_without_implementation",
      "full_experiment_resolution": "not_reproduced_high_cost",
      "full_experiment_review_done": true,
      "arxiv_v1_date": "2026-02-02",
      "date_evidence": "official_arxiv_abs_or_prior_exact_identity_date",
      "earlier_public_release_exhaustively_checked": false,
      "authorship": {
        "first_author_institution_or_role": "Michigan State University / Google collaboration",
        "google_priority_basis": "current_arxiv_author_block_has_google_affiliation_or_collaboration",
        "first_author_google_required": false
      },
      "method_summary_zh": "多Agent漏洞发现与隔离Docker执行：Planner/Validation/Execution/Evaluation循环，三层pattern/strategy/technical memory；代码上下文与真实执行反馈。",
      "experiment_summary_zh": "CyBench/BountyBench/CyberGym多模型、多对照及迭代消融；Gemini系列与GPT5-mini/o4-mini/Claude4.5/open20B/32B；正文报告安全benchmark而非线上用户A/B。",
      "public_code": {
        "status": "not_found_in_reviewed_primary_sources",
        "url": null,
        "notes_zh": "已读正文未找到作者Co-RedTeam实现仓库；OpenHands链接是baseline。",
        "search_scope": "reviewed arXiv v1 article text, code/repository mentions and linked primary repository when explicitly provided; no exhaustive web-wide absence claim"
      },
      "industrial_online_evidence": {
        "strict_online_ab_confirmed": false,
        "explicit_full_traffic_confirmed": false,
        "notes_zh": "仅隔离安全评测，无生产A/B或全流量证据。"
      },
      "full_text_review": {
        "scope": "bounded_primary_fulltext_method_experiment_resource_and_release_review",
        "source_url": "https://arxiv.org/html/2602.02164v1",
        "read_sections": [
          "§3.2-3.4",
          "§4 Experiments",
          "Appendix A.5/C"
        ],
        "source_locations": "L104-L131;L329-L370",
        "full_article_read": false,
        "proofs_read_in_full": false,
        "all_tables_read": false,
        "latest_revision_reviewed": false,
        "version": "v1",
        "limitations": "只审读列明关键段与检索返回内容，不声称逐段通读、穷尽全部附录或确认最新修订。"
      },
      "cost_assessment": {
        "available_resource": "one NVIDIA A100 or A30, or local Mac; no multi-GPU assumption",
        "paper_configuration_and_scale": "CyBench/BountyBench/CyberGym多模型、多对照及迭代消融；Gemini系列与GPT5-mini/o4-mini/Claude4.5/open20B/32B；正文报告安全benchmark而非线上用户A/B。",
        "decision_zh": "三benchmark×多backbone×多baseline长执行轨迹＋独立Docker目标环境，需付费API/20B-32B模型与多轮发现/利用；完整多臂成本明显超出短单卡预算，按配置推断未实测。",
        "basis": "reviewed primary-paper configuration, directly reported resource/runtime where available, otherwise qualitative bounded estimate",
        "local_cost_measurement_performed": false,
        "training_or_model_download_performed": false
      },
      "source_url": "https://arxiv.org/html/2602.02164v1",
      "source_sha256": "1fd8770e975ee9a8bcbdfacb2bae88cbd49113399f32f27ec29f091749ef735e",
      "source_hash_kind": "sha256_of_public_metadata_locations_and_chinese_paraphrase_receipt_not_raw_capture_or_article",
      "source_retrieval_reference": "evidence/google39/2602.02164.json",
      "retrieval_utc": {
        "start": "2026-10-20T18:48:00Z",
        "end": "2026-10-20T18:56:04Z",
        "precision": "bounded_session_capture_window_not_per_request_timestamp"
      },
      "official_identity_sources": [
        "https://research.google/pubs/co-redteam-orchestrated-security-discovery-and-exploitation-with-llm-agents/"
      ],
      "original_capture_sha256": "4ba5c48a3412393e7e69cbd272754f333bd1c955c8af905c846468ba8a465d5c"
    },
    {
      "id": "2602.02690",
      "title": "Outrunning LLM Cutoffs: A Live Kernel Crash Resolution Benchmark for All,",
      "track": "agent",
      "status": "deferred",
      "review_decision": "deferred",
      "review_done": true,
      "resolution": "not_reproduced_high_cost",
      "implementation_status": "not_implemented",
      "implementation_authorized_this_run": false,
      "automatic_implementation_queue": false,
      "automatic_retry": false,
      "evolve_integrated": false,
      "promotion_eligible": false,
      "review_decision_scope": "full_paper_reproduction_closed_without_implementation",
      "full_experiment_resolution": "not_reproduced_high_cost",
      "full_experiment_review_done": true,
      "arxiv_v1_date": "2026-02-02",
      "date_evidence": "current_official_arxiv_html_v1_header",
      "earlier_public_release_exhaustively_checked": false,
      "authorship": {
        "first_author_institution_or_role": "Columbia University",
        "google_priority_basis": "current_arxiv_author_block_has_google_affiliation_or_collaboration",
        "first_author_google_required": false
      },
      "method_summary_zh": "Live-kBench按时间持续采集内核bug；kEnv解耦agent与编译/执行环境，Crash Resolution Feedback与patch-equivalence judge。",
      "experiment_summary_zh": "534 Linux bugs；3 scaffolds、Gemini3Pro/Flash与ClaudeOpus4.5；judge每patch9票；单内核补丁编译测试至少30分钟。",
      "public_code": {
        "status": "not_found_in_reviewed_primary_sources",
        "url": null,
        "notes_zh": "正文未找到本框架的具名作者仓库；公开dashboard/benchmark声明不等于源码已验证。",
        "search_scope": "reviewed arXiv v1 article text, code/repository mentions and linked primary repository when explicitly provided; no exhaustive web-wide absence claim"
      },
      "industrial_online_evidence": {
        "strict_online_ab_confirmed": false,
        "explicit_full_traffic_confirmed": false,
        "notes_zh": "公开离线benchmark，不是生产A/B。"
      },
      "full_text_review": {
        "scope": "bounded_primary_fulltext_method_experiment_resource_and_release_review",
        "source_url": "https://arxiv.org/html/2602.02690v1",
        "read_sections": [
          "§3 kEnv导航",
          "§4 Live-kBench",
          "§5.1",
          "引言成本"
        ],
        "source_locations": "L45-L59;L140-L190",
        "full_article_read": false,
        "proofs_read_in_full": false,
        "all_tables_read": false,
        "latest_revision_reviewed": false,
        "version": "v1",
        "limitations": "只审读列明关键段与检索返回内容，不声称逐段通读、穷尽全部附录或确认最新修订。"
      },
      "cost_assessment": {
        "available_resource": "one NVIDIA A100 or A30, or local Mac; no multi-GPU assumption",
        "paper_configuration_and_scale": "534 Linux bugs；3 scaffolds、Gemini3Pro/Flash与ClaudeOpus4.5；judge每patch9票；单内核补丁编译测试至少30分钟。",
        "decision_zh": "仅534 bug一次30分钟即≥267串行小时，尚未含多agent/模型/重复验证；这是正文单例下界外推而非测量。内核VM与云执行也非Mac/单A100替代可忠实覆盖。",
        "basis": "reviewed primary-paper configuration, directly reported resource/runtime where available, otherwise qualitative bounded estimate",
        "local_cost_measurement_performed": false,
        "training_or_model_download_performed": false
      },
      "source_url": "https://arxiv.org/html/2602.02690v1",
      "source_sha256": "4e68b191ee36a17f566160667907aa2a2afc368109dd61a83100d4823440b1b5",
      "source_hash_kind": "sha256_of_public_metadata_locations_and_chinese_paraphrase_receipt_not_raw_capture_or_article",
      "source_retrieval_reference": "evidence/google39/2602.02690.json",
      "retrieval_utc": {
        "start": "2026-10-20T18:48:00Z",
        "end": "2026-10-20T18:56:04Z",
        "precision": "bounded_session_capture_window_not_per_request_timestamp"
      },
      "official_identity_sources": [
        "https://research.google/pubs/outrunning-llm-cutoffs-a-live-kernel-crash-resolution-benchmark-for-all/"
      ],
      "original_capture_sha256": "39443f9dc580df3b2d8760f7bb342a69129f7cc52c3588ceea351471ced669a6"
    },
    {
      "id": "2602.06276",
      "title": "Statistical Learning from Attribution Sets",
      "track": "recommendation",
      "status": "accepted",
      "review_decision": "accepted",
      "review_done": true,
      "resolution": "accepted_core_not_implemented_this_run",
      "implementation_status": "not_implemented",
      "implementation_authorized_this_run": false,
      "automatic_implementation_queue": false,
      "automatic_retry": false,
      "evolve_integrated": false,
      "promotion_eligible": false,
      "review_decision_scope": "bounded_core_or_evaluation_candidate_only_not_industrial_admission",
      "full_experiment_resolution": "not_reproduced_high_cost",
      "full_experiment_review_done": true,
      "arxiv_v1_date": "2026-02-06",
      "date_evidence": "current_official_arxiv_html_v1_header",
      "earlier_public_release_exhaustively_checked": false,
      "authorship": {
        "first_author_institution_or_role": "Google Research",
        "google_priority_basis": "current_arxiv_author_block_has_google_affiliation_or_collaboration",
        "first_author_google_required": false
      },
      "method_summary_zh": "由匿名attribution sets和先验构造无偏population-loss estimator，再ERM；与随机分配/最大先验伪标签baseline比较。",
      "experiment_summary_zh": "MNIST、CIFAR10、Higgs；MLP/CNN；k=2^0…2^8；10学习率×10重复，MNIST200epoch、其余100epoch。正文是初步离线模拟归因集合实验。",
      "public_code": {
        "status": "not_found_in_reviewed_primary_sources",
        "url": null,
        "notes_zh": "已读正文未发现作者实现仓库。",
        "search_scope": "reviewed arXiv v1 article text, code/repository mentions and linked primary repository when explicitly provided; no exhaustive web-wide absence claim"
      },
      "industrial_online_evidence": {
        "strict_online_ab_confirmed": false,
        "explicit_full_traffic_confirmed": false,
        "notes_zh": "无真实广告生产A/B或全流量；不能把行业启发式baseline当线上对照。"
      },
      "full_text_review": {
        "scope": "bounded_primary_fulltext_method_experiment_resource_and_release_review",
        "source_url": "https://arxiv.org/html/2602.06276v1",
        "read_sections": [
          "§3-4章节导航",
          "§5 Experiments"
        ],
        "source_locations": "L235-L253",
        "full_article_read": false,
        "proofs_read_in_full": false,
        "all_tables_read": false,
        "latest_revision_reviewed": false,
        "version": "v1",
        "limitations": "只审读列明关键段与检索返回内容，不声称逐段通读、穷尽全部附录或确认最新修订。"
      },
      "cost_assessment": {
        "available_resource": "one NVIDIA A100 or A30, or local Mac; no multi-GPU assumption",
        "paper_configuration_and_scale": "MNIST、CIFAR10、Higgs；MLP/CNN；k=2^0…2^8；10学习率×10重复，MNIST200epoch、其余100epoch。正文是初步离线模拟归因集合实验。",
        "decision_zh": "完整网格约(9+9+8)×3×10×10=7,800次训练，100–200epoch，远超单卡短运行；核心loss小数据CPU机制可负担，未授权本轮实现。",
        "basis": "reviewed primary-paper configuration, directly reported resource/runtime where available, otherwise qualitative bounded estimate",
        "local_cost_measurement_performed": false,
        "training_or_model_download_performed": false
      },
      "source_url": "https://arxiv.org/html/2602.06276v1",
      "source_sha256": "dde635f3f538fead75f4540ef5ac3c6605c8ce64640f5bfeb333609bec728354",
      "source_hash_kind": "sha256_of_public_metadata_locations_and_chinese_paraphrase_receipt_not_raw_capture_or_article",
      "source_retrieval_reference": "evidence/google39/2602.06276.json",
      "retrieval_utc": {
        "start": "2026-10-20T18:48:00Z",
        "end": "2026-10-20T18:56:04Z",
        "precision": "bounded_session_capture_window_not_per_request_timestamp"
      },
      "official_identity_sources": [
        "https://research.google/pubs/learning-from-attribution-sets/"
      ],
      "original_capture_sha256": "8b3820deb18a2e82858c17586190a97a53a6c709149f643579e70d0775eb076e",
      "google_listing_title": "Learning from Attribution Sets",
      "title_source": "https://arxiv.org/html/2602.06276v1"
    },
    {
      "id": "2602.07266",
      "title": "ADCanvas: Accessible and Conversational Audio Description Authoring for Blind and Low Vision Creators",
      "track": "agent",
      "status": "rejected_scope",
      "review_decision": "rejected_scope",
      "review_done": true,
      "resolution": "out_of_scope",
      "implementation_status": "not_implemented",
      "implementation_authorized_this_run": false,
      "automatic_implementation_queue": false,
      "automatic_retry": false,
      "evolve_integrated": false,
      "promotion_eligible": false,
      "review_decision_scope": "paper_method_scope",
      "full_experiment_resolution": "not_applicable_scope",
      "full_experiment_review_done": true,
      "arxiv_v1_date": "2026-02-06",
      "date_evidence": "current_official_arxiv_html_v1_header",
      "earlier_public_release_exhaustively_checked": false,
      "authorship": {
        "first_author_institution_or_role": "Carnegie Mellon University / Google Research",
        "google_priority_basis": "current_arxiv_author_block_has_google_affiliation_or_collaboration",
        "first_author_google_required": false
      },
      "method_summary_zh": "ADCanvas屏幕阅读器/键盘WebVTT编辑器，视频问答与音频描述脚本协作界面；贡献是无障碍交互设计与用户研究，不是独立可比较的Agent算法。",
      "experiment_summary_zh": "12名盲人/低视力创作者的形成性authoring study；三个视频任务，质性观察与可用性反馈。",
      "public_code": {
        "status": "not_found_in_reviewed_primary_sources",
        "url": null,
        "notes_zh": "已读方法/用户研究未核到作者公开系统代码。",
        "search_scope": "reviewed arXiv v1 article text, code/repository mentions and linked primary repository when explicitly provided; no exhaustive web-wide absence claim"
      },
      "industrial_online_evidence": {
        "strict_online_ab_confirmed": false,
        "explicit_full_traffic_confirmed": false,
        "notes_zh": "小样本用户研究不是生产A/B或全流量。"
      },
      "full_text_review": {
        "scope": "bounded_primary_fulltext_method_experiment_resource_and_release_review",
        "source_url": "https://arxiv.org/html/2602.07266v1",
        "read_sections": [
          "§3.2 Design Rationale",
          "§3.3技术实现定位",
          "§4用户研究任务/流程"
        ],
        "source_locations": "L227-L247;L273;L316-L328",
        "full_article_read": false,
        "proofs_read_in_full": false,
        "all_tables_read": false,
        "latest_revision_reviewed": false,
        "version": "v1",
        "limitations": "只审读列明关键段与检索返回内容，不声称逐段通读、穷尽全部附录或确认最新修订。"
      },
      "cost_assessment": {
        "available_resource": "one NVIDIA A100 or A30, or local Mac; no multi-GPU assumption",
        "paper_configuration_and_scale": "12名盲人/低视力创作者的形成性authoring study；三个视频任务，质性观察与可用性反馈。",
        "decision_zh": "scope终态；API原型可编写不等于本论文用户研究能在单卡自动复现；不安排受试者招募。",
        "basis": "reviewed primary-paper configuration, directly reported resource/runtime where available, otherwise qualitative bounded estimate",
        "local_cost_measurement_performed": false,
        "training_or_model_download_performed": false
      },
      "source_url": "https://arxiv.org/html/2602.07266v1",
      "source_sha256": "48461145d250e77c956113131a1275558986abd7a9585b4d06ca2b053fd7f108",
      "source_hash_kind": "sha256_of_public_metadata_locations_and_chinese_paraphrase_receipt_not_raw_capture_or_article",
      "source_retrieval_reference": "evidence/google39/2602.07266.json",
      "retrieval_utc": {
        "start": "2026-10-20T18:48:00Z",
        "end": "2026-10-20T18:56:04Z",
        "precision": "bounded_session_capture_window_not_per_request_timestamp"
      },
      "official_identity_sources": [
        "https://research.google/pubs/adcanvas-accessible-and-conversational-audio-description-authoring-for-blind-and-low-vision-creators/"
      ],
      "original_capture_sha256": "2573707e534f48efa609b55e6e257249461e67f50699f63ad9682412afc72dd7"
    },
    {
      "id": "2602.07918",
      "title": "Causal Armor: Efficient Indirect Prompt Injection Guardrails via Causal Attribution",
      "track": "agent",
      "status": "accepted",
      "review_decision": "accepted",
      "review_done": true,
      "resolution": "accepted_core_not_implemented_this_run",
      "implementation_status": "not_implemented",
      "implementation_authorized_this_run": false,
      "automatic_implementation_queue": false,
      "automatic_retry": false,
      "evolve_integrated": false,
      "promotion_eligible": false,
      "review_decision_scope": "bounded_core_or_evaluation_candidate_only_not_industrial_admission",
      "full_experiment_resolution": "not_reproduced_high_cost",
      "full_experiment_review_done": true,
      "arxiv_v1_date": "2026-02-08",
      "date_evidence": "current_official_arxiv_html_v1_header",
      "earlier_public_release_exhaustively_checked": false,
      "authorship": {
        "first_author_institution_or_role": "Seoul National University / Google collaboration",
        "google_priority_basis": "current_arxiv_author_block_has_google_affiliation_or_collaboration",
        "first_author_google_required": false
      },
      "method_summary_zh": "privileged action处LOO因果消融，比较用户与不可信段落影响，触发选择性sanitization；清除注入后assistant CoT并重新生成action。",
      "experiment_summary_zh": "AgentDojo1.2.2四环境/三攻击template与DoomArena TauBench115场景×5seed；VertexAI多backbone/sanitizer，原环境8×A100。",
      "public_code": {
        "status": "not_found_in_reviewed_primary_sources",
        "url": null,
        "notes_zh": "已读正文未核到作者CausalArmor仓库；AgentDojo链接为benchmark而非作者实现。",
        "search_scope": "reviewed arXiv v1 article text, code/repository mentions and linked primary repository when explicitly provided; no exhaustive web-wide absence claim"
      },
      "industrial_online_evidence": {
        "strict_online_ab_confirmed": false,
        "explicit_full_traffic_confirmed": false,
        "notes_zh": "隔离Agent安全评测，未报告生产A/B或全流量。"
      },
      "full_text_review": {
        "scope": "bounded_primary_fulltext_method_experiment_resource_and_release_review",
        "source_url": "https://arxiv.org/html/2602.07918v1",
        "read_sections": [
          "§3-4机制导航",
          "Algorithm 1",
          "Appendix C/D.1"
        ],
        "source_locations": "L545-L608;L632-L634",
        "full_article_read": false,
        "proofs_read_in_full": false,
        "all_tables_read": false,
        "latest_revision_reviewed": false,
        "version": "v1",
        "limitations": "只审读列明关键段与检索返回内容，不声称逐段通读、穷尽全部附录或确认最新修订。"
      },
      "cost_assessment": {
        "available_resource": "one NVIDIA A100 or A30, or local Mac; no multi-GPU assumption",
        "paper_configuration_and_scale": "AgentDojo1.2.2四环境/三攻击template与DoomArena TauBench115场景×5seed；VertexAI多backbone/sanitizer，原环境8×A100。",
        "decision_zh": "原文8×A100与多模型/攻击/seed完整矩阵成本过高；不尝试下载或启动。LOO/masking有界组件可单卡/CPU模型接入，未在本轮批准实施。",
        "basis": "reviewed primary-paper configuration, directly reported resource/runtime where available, otherwise qualitative bounded estimate",
        "local_cost_measurement_performed": false,
        "training_or_model_download_performed": false
      },
      "source_url": "https://arxiv.org/html/2602.07918v1",
      "source_sha256": "5eca3db81e91f363b5eb4da5a10247f56796693890f579788561287c00bb17f6",
      "source_hash_kind": "sha256_of_public_metadata_locations_and_chinese_paraphrase_receipt_not_raw_capture_or_article",
      "source_retrieval_reference": "evidence/google39/2602.07918.json",
      "retrieval_utc": {
        "start": "2026-10-20T18:48:00Z",
        "end": "2026-10-20T18:56:04Z",
        "precision": "bounded_session_capture_window_not_per_request_timestamp"
      },
      "official_identity_sources": [
        "https://research.google/pubs/causal-armor-efficient-indirect-prompt-injection-guardrails-via-causal-attribution/"
      ],
      "original_capture_sha256": "427c8b5f3d0f4811c46c4a32edc3963f066610ce62ed7ec9805b0c1f9614d1f2"
    },
    {
      "id": "2602.12481",
      "title": "Online Advertising with Spatial Interactions",
      "track": "industrial",
      "status": "accepted",
      "review_decision": "accepted",
      "review_done": true,
      "resolution": "accepted_core_not_implemented_this_run",
      "implementation_status": "not_implemented",
      "implementation_authorized_this_run": false,
      "automatic_implementation_queue": false,
      "automatic_retry": false,
      "evolve_integrated": false,
      "promotion_eligible": false,
      "review_decision_scope": "bounded_core_or_evaluation_candidate_only_not_industrial_admission",
      "full_experiment_resolution": "not_reproduced_high_cost",
      "full_experiment_review_done": true,
      "arxiv_v1_date": "2026-02-12",
      "date_evidence": "current_official_arxiv_html_v1_header",
      "earlier_public_release_exhaustively_checked": false,
      "authorship": {
        "first_author_institution_or_role": "Google Research",
        "google_priority_basis": "current_arxiv_author_block_has_google_affiliation_or_collaboration",
        "first_author_google_required": false
      },
      "method_summary_zh": "广告slot建模为metric points；nearest-neighbor外部性转grouped pseudo-disk、LP rounding得到18近似并论证truthful；另有简单log近似与2D PTAS。",
      "experiment_summary_zh": "理论算法与证明；正文目录至hardness/附录，无实验章，全文检索experiment无结果；不是生产广告实验。",
      "public_code": {
        "status": "not_found_in_reviewed_primary_sources",
        "url": null,
        "notes_zh": "已读原文未发现作者代码链接。",
        "search_scope": "reviewed arXiv v1 article text, code/repository mentions and linked primary repository when explicitly provided; no exhaustive web-wide absence claim"
      },
      "industrial_online_evidence": {
        "strict_online_ab_confirmed": false,
        "explicit_full_traffic_confirmed": false,
        "notes_zh": "未报告生产A/B或全流量；理论广告机制不满足工业上线证据门槛。"
      },
      "full_text_review": {
        "scope": "bounded_primary_fulltext_method_experiment_resource_and_release_review",
        "source_url": "https://arxiv.org/html/2602.12481v1",
        "read_sections": [
          "§2.1-2.3",
          "Algorithm简单slot选择",
          "全文章节/experiment检索"
        ],
        "source_locations": "L178-L246;L308-L354",
        "full_article_read": false,
        "proofs_read_in_full": false,
        "all_tables_read": false,
        "latest_revision_reviewed": false,
        "version": "v1",
        "limitations": "只审读列明关键段与检索返回内容，不声称逐段通读、穷尽全部附录或确认最新修订。"
      },
      "cost_assessment": {
        "available_resource": "one NVIDIA A100 or A30, or local Mac; no multi-GPU assumption",
        "paper_configuration_and_scale": "理论算法与证明；正文目录至hardness/附录，无实验章，全文检索experiment无结果；不是生产广告实验。",
        "decision_zh": "小规模LP/几何实例CPU可负担；无必须大模型/数据；未在本轮批准实现，不用缺A/B否认理论算法。",
        "basis": "reviewed primary-paper configuration, directly reported resource/runtime where available, otherwise qualitative bounded estimate",
        "local_cost_measurement_performed": false,
        "training_or_model_download_performed": false
      },
      "source_url": "https://arxiv.org/html/2602.12481v1",
      "source_sha256": "205f87ef58138591d91b3919a3d20e4d9c6d8478ae81ff64fa24efd18b51315c",
      "source_hash_kind": "sha256_of_public_metadata_locations_and_chinese_paraphrase_receipt_not_raw_capture_or_article",
      "source_retrieval_reference": "evidence/google39/2602.12481.json",
      "retrieval_utc": {
        "start": "2026-10-20T18:48:00Z",
        "end": "2026-10-20T18:56:04Z",
        "precision": "bounded_session_capture_window_not_per_request_timestamp"
      },
      "official_identity_sources": [
        "https://research.google/pubs/online-advertising-with-spatial-interactions/"
      ],
      "original_capture_sha256": "6f97ad40b46cdcf416c5b38b96cad0de95888194519e52ace1a94434a0ff922a"
    },
    {
      "id": "2602.16938",
      "title": "ConvApparel: A Benchmark Dataset and Validation Framework for User Simulators in Conversational Recommenders",
      "track": "agent",
      "status": "accepted",
      "review_decision": "accepted",
      "review_done": true,
      "resolution": "accepted_core_not_implemented_this_run",
      "implementation_status": "not_implemented",
      "implementation_authorized_this_run": false,
      "automatic_implementation_queue": false,
      "automatic_retry": false,
      "evolve_integrated": false,
      "promotion_eligible": false,
      "review_decision_scope": "bounded_core_or_evaluation_candidate_only_not_industrial_admission",
      "full_experiment_resolution": "not_reproduced_high_cost",
      "full_experiment_review_done": true,
      "arxiv_v1_date": "2026-02-18",
      "date_evidence": "current_official_arxiv_html_v1_header",
      "earlier_public_release_exhaustively_checked": false,
      "authorship": {
        "first_author_institution_or_role": "Google",
        "google_priority_basis": "current_arxiv_author_block_has_google_affiliation_or_collaboration",
        "first_author_google_required": false
      },
      "method_summary_zh": "ConvApparel好/坏recommender采集数据，PLSA行为统计、训练discriminator的HLS与跨策略counterfactual validation；Prompt/ICL(k3)/SFT模拟器。",
      "experiment_summary_zh": "4,146真人对话＋3,549模拟对话训练Gemini2.5FlashLite discriminator；各simulator生成300好＋300坏对话；90年代Turing类比不是生产证据；KS/MWU检验及HLS显示realism gap。",
      "public_code": {
        "status": "not_found_in_reviewed_primary_sources",
        "url": null,
        "notes_zh": "已读正文未找到作者公开代码仓库；数据发布声明不自动等于代码。",
        "search_scope": "reviewed arXiv v1 article text, code/repository mentions and linked primary repository when explicitly provided; no exhaustive web-wide absence claim"
      },
      "industrial_online_evidence": {
        "strict_online_ab_confirmed": false,
        "explicit_full_traffic_confirmed": false,
        "notes_zh": "受试者交互采集好/坏系统并非生产真实流量A/B或明确全流量。"
      },
      "full_text_review": {
        "scope": "bounded_primary_fulltext_method_experiment_resource_and_release_review",
        "source_url": "https://arxiv.org/html/2602.16938v1",
        "read_sections": [
          "§4-6",
          "§6.1 discriminator配置",
          "§6.2结果"
        ],
        "source_locations": "L120-L153;L169-L202",
        "full_article_read": false,
        "proofs_read_in_full": false,
        "all_tables_read": false,
        "latest_revision_reviewed": false,
        "version": "v1",
        "limitations": "只审读列明关键段与检索返回内容，不声称逐段通读、穷尽全部附录或确认最新修订。"
      },
      "cost_assessment": {
        "available_resource": "one NVIDIA A100 or A30, or local Mac; no multi-GPU assumption",
        "paper_configuration_and_scale": "4,146真人对话＋3,549模拟对话训练Gemini2.5FlashLite discriminator；各simulator生成300好＋300坏对话；90年代Turing类比不是生产证据；KS/MWU检验及HLS显示realism gap。",
        "decision_zh": "有界统计/validation框架可负担；完整真人数据采集、多个GeminiSFT与数千多轮rollout不在单卡短预算，原完整实验不复现。",
        "basis": "reviewed primary-paper configuration, directly reported resource/runtime where available, otherwise qualitative bounded estimate",
        "local_cost_measurement_performed": false,
        "training_or_model_download_performed": false
      },
      "source_url": "https://arxiv.org/html/2602.16938v1",
      "source_sha256": "9c5a79260f66684ce72cb66467b78de5b39b3b1416de333862dbc41d3853b9c7",
      "source_hash_kind": "sha256_of_public_metadata_locations_and_chinese_paraphrase_receipt_not_raw_capture_or_article",
      "source_retrieval_reference": "evidence/google39/2602.16938.json",
      "retrieval_utc": {
        "start": "2026-10-20T18:48:00Z",
        "end": "2026-10-20T18:56:04Z",
        "precision": "bounded_session_capture_window_not_per_request_timestamp"
      },
      "official_identity_sources": [
        "https://research.google/pubs/convapparel-a-benchmark-dataset-and-validation-framework-for-user-simulators-in-conversational-recommenders-2/"
      ],
      "original_capture_sha256": "4d9ca84530b634b1899f927637b95cf90b1acc44ee0b677aecd669d99214cb0f"
    },
    {
      "id": "2602.17554",
      "title": "A Theoretical Framework for Modular Learning of Robust Generative Models",
      "track": "foundation",
      "status": "accepted",
      "review_decision": "accepted",
      "review_done": true,
      "resolution": "accepted_core_not_implemented_this_run",
      "implementation_status": "not_implemented",
      "implementation_authorized_this_run": false,
      "automatic_implementation_queue": false,
      "automatic_retry": false,
      "evolve_integrated": false,
      "promotion_eligible": false,
      "review_decision_scope": "bounded_core_or_evaluation_candidate_only_not_industrial_admission",
      "full_experiment_resolution": "not_implemented",
      "full_experiment_review_done": true,
      "arxiv_v1_date": "2026-02-19",
      "date_evidence": "current_official_arxiv_html_v1_header",
      "earlier_public_release_exhaustively_checked": false,
      "authorship": {
        "first_author_institution_or_role": "Google Research",
        "google_priority_basis": "current_arxiv_author_block_has_google_affiliation_or_collaboration",
        "first_author_google_required": false
      },
      "method_summary_zh": "normalized robust gate最小化最坏mixture KL；stochastic primal-dual训练，SIR/rejection sampling与cached-logit结构蒸馏causal router；不是普通expert均值。",
      "experiment_summary_zh": "synthetic vocab100/seq10、1层2头ff32、embedding6/8/16/20；expert/baseline/gate800步，5重复；结构蒸馏5K序列。已读部分可明确CPU小机制可行。",
      "public_code": {
        "status": "not_found_in_reviewed_primary_sources",
        "url": null,
        "notes_zh": "已读正文未发现作者代码仓库。",
        "search_scope": "reviewed arXiv v1 article text, code/repository mentions and linked primary repository when explicitly provided; no exhaustive web-wide absence claim"
      },
      "industrial_online_evidence": {
        "strict_online_ab_confirmed": false,
        "explicit_full_traffic_confirmed": false,
        "notes_zh": "已读为synthetic/离线建模，无生产A/B或全流量证据。"
      },
      "full_text_review": {
        "scope": "bounded_primary_fulltext_method_experiment_resource_and_release_review",
        "source_url": "https://arxiv.org/html/2602.17554v1",
        "read_sections": [
          "§7.3 Algorithm5",
          "§8.1-8.3"
        ],
        "source_locations": "L827-L915",
        "full_article_read": false,
        "proofs_read_in_full": false,
        "all_tables_read": false,
        "latest_revision_reviewed": false,
        "version": "v1",
        "limitations": "只审读列明关键段与检索返回内容，不声称逐段通读、穷尽全部附录或确认最新修订。"
      },
      "cost_assessment": {
        "available_resource": "one NVIDIA A100 or A30, or local Mac; no multi-GPU assumption",
        "paper_configuration_and_scale": "synthetic vocab100/seq10、1层2头ff32、embedding6/8/16/20；expert/baseline/gate800步，5重复；结构蒸馏5K序列。已读部分可明确CPU小机制可行。",
        "decision_zh": "正文小型Transformer实验可在Mac/单卡负担；不因为标题涉及LLM而判高成本；本轮未批准实施。真实数据§8.4-8.5尚未逐段审读，不据此声称完整实验复现可负担。",
        "basis": "reviewed primary-paper configuration, directly reported resource/runtime where available, otherwise qualitative bounded estimate",
        "local_cost_measurement_performed": false,
        "training_or_model_download_performed": false
      },
      "source_url": "https://arxiv.org/html/2602.17554v1",
      "source_sha256": "eef22aa59cc164eccd4f7baf425753daa5c87bafa48c74103d4a94d33b269d03",
      "source_hash_kind": "sha256_of_public_metadata_locations_and_chinese_paraphrase_receipt_not_raw_capture_or_article",
      "source_retrieval_reference": "evidence/google39/2602.17554.json",
      "retrieval_utc": {
        "start": "2026-10-20T18:48:00Z",
        "end": "2026-10-20T18:56:04Z",
        "precision": "bounded_session_capture_window_not_per_request_timestamp"
      },
      "official_identity_sources": [
        "https://research.google/pubs/a-theoretical-framework-for-modular-learning-of-robust-generative-models/"
      ],
      "original_capture_sha256": "3b9e529a76e8b8d3efb1fa54297b4784ae7d635b56e4a74b20b4172cc9563615"
    },
    {
      "id": "2603.01293",
      "title": "Theoretical Perspectives on Data Quality and Synergistic Effects in Pre- and Post-Training Reasoning Models",
      "track": "foundation/post-training",
      "status": "accepted",
      "review_decision": "accepted",
      "review_done": true,
      "resolution": "accepted_core_not_implemented_this_run",
      "implementation_status": "not_implemented",
      "implementation_authorized_this_run": false,
      "automatic_implementation_queue": false,
      "automatic_retry": false,
      "evolve_integrated": false,
      "promotion_eligible": false,
      "review_decision_scope": "bounded_core_or_evaluation_candidate_only_not_industrial_admission",
      "full_experiment_resolution": "not_reproduced_high_cost",
      "full_experiment_review_done": true,
      "arxiv_v1_date": "2026-03-01",
      "date_evidence": "current_official_arxiv_html_v1_header",
      "earlier_public_release_exhaustively_checked": false,
      "authorship": {
        "first_author_institution_or_role": "USC / Google Research",
        "google_priority_basis": "current_arxiv_author_block_has_google_affiliation_or_collaboration",
        "first_author_google_required": false
      },
      "method_summary_zh": "线性回归in-context weight预测理论：pretrain covariance与SFT/OS数据分配；OS是outcome-supervised regression，明确不是完整policy-gradient RL。",
      "experiment_summary_zh": "LSA d100/130epochs；GPT2架构9.5M、12层8头d256、batch64，SFT100步/OS12K步，10重复；synthetic covariance shift。",
      "public_code": {
        "status": "not_found_in_reviewed_primary_sources",
        "url": null,
        "notes_zh": "已读正文未发现作者代码。",
        "search_scope": "reviewed arXiv v1 article text, code/repository mentions and linked primary repository when explicitly provided; no exhaustive web-wide absence claim"
      },
      "industrial_online_evidence": {
        "strict_online_ab_confirmed": false,
        "explicit_full_traffic_confirmed": false,
        "notes_zh": "synthetic理论验证，无生产A/B或全流量。"
      },
      "full_text_review": {
        "scope": "bounded_primary_fulltext_method_experiment_resource_and_release_review",
        "source_url": "https://arxiv.org/html/2603.01293v1",
        "read_sections": [
          "§1任务/OS边界",
          "§8 Experiments"
        ],
        "source_locations": "L59-L61;L275-L317",
        "full_article_read": false,
        "proofs_read_in_full": false,
        "all_tables_read": false,
        "latest_revision_reviewed": false,
        "version": "v1",
        "limitations": "只审读列明关键段与检索返回内容，不声称逐段通读、穷尽全部附录或确认最新修订。"
      },
      "cost_assessment": {
        "available_resource": "one NVIDIA A100 or A30, or local Mac; no multi-GPU assumption",
        "paper_configuration_and_scale": "LSA d100/130epochs；GPT2架构9.5M、12层8头d256、batch64，SFT100步/OS12K步，10重复；synthetic covariance shift。",
        "decision_zh": "LSA核心CPU可负担；完整GPT2×多B/n/k×10run×12K OS步预计长训练，按网格规模推断超短预算，不启动；未在本轮批准LSA组件。",
        "basis": "reviewed primary-paper configuration, directly reported resource/runtime where available, otherwise qualitative bounded estimate",
        "local_cost_measurement_performed": false,
        "training_or_model_download_performed": false
      },
      "source_url": "https://arxiv.org/html/2603.01293v1",
      "source_sha256": "b8065a6fd16e805d23e3de4352dc184ad3b6ba1b031fc892f4494946a33bc89c",
      "source_hash_kind": "sha256_of_public_metadata_locations_and_chinese_paraphrase_receipt_not_raw_capture_or_article",
      "source_retrieval_reference": "evidence/google39/2603.01293.json",
      "retrieval_utc": {
        "start": "2026-10-20T18:48:00Z",
        "end": "2026-10-20T18:56:04Z",
        "precision": "bounded_session_capture_window_not_per_request_timestamp"
      },
      "official_identity_sources": [
        "https://research.google/pubs/theoretical-perspectives-on-data-quality-and-synergistic-effects-in-pre-and-post-training-reasoning-models/"
      ],
      "original_capture_sha256": "8d010186478e59b750728171a506b48afb98c1bb6a0990bef5b90dcd02834f6c"
    },
    {
      "id": "2603.06981",
      "title": "Diffusion Controller: Framework, Algorithms and Parameterization",
      "track": "multimodal/post-training",
      "status": "deferred",
      "review_decision": "deferred",
      "review_done": true,
      "resolution": "not_reproduced_high_cost",
      "implementation_status": "not_implemented",
      "implementation_authorized_this_run": false,
      "automatic_implementation_queue": false,
      "automatic_retry": false,
      "evolve_integrated": false,
      "promotion_eligible": false,
      "review_decision_scope": "full_paper_reproduction_closed_without_implementation",
      "full_experiment_resolution": "not_reproduced_high_cost",
      "full_experiment_review_done": true,
      "arxiv_v1_date": "2026-03-07",
      "date_evidence": "current_official_arxiv_html_v1_header",
      "earlier_public_release_exhaustively_checked": false,
      "authorship": {
        "first_author_institution_or_role": "CMU / Google Research",
        "google_priority_basis": "current_arxiv_author_block_has_google_affiliation_or_collaboration",
        "first_author_google_required": false
      },
      "method_summary_zh": "DiffCon将reverse diffusion看作LS-MDP控制：f-divergence regularized PG/PPO与reward-weighted regression；冻结backbone+暴露中间输出的轻量side control score修正。",
      "experiment_summary_zh": "SD1.4，5种参数化×SFT/RWL/PPO；side12M，LoRA17M；SFT1000/RWL2000/PPO2400步；HPS/CLIP/PickScore及人工评测。",
      "public_code": {
        "status": "not_found_in_reviewed_primary_sources",
        "url": null,
        "notes_zh": "已读正文未核到作者代码仓库。",
        "search_scope": "reviewed arXiv v1 article text, code/repository mentions and linked primary repository when explicitly provided; no exhaustive web-wide absence claim"
      },
      "industrial_online_evidence": {
        "strict_online_ab_confirmed": false,
        "explicit_full_traffic_confirmed": false,
        "notes_zh": "离线图像偏好比较，不是生产A/B。"
      },
      "full_text_review": {
        "scope": "bounded_primary_fulltext_method_experiment_resource_and_release_review",
        "source_url": "https://arxiv.org/html/2603.06981v1",
        "read_sections": [
          "§1模型定义",
          "§5.1-5.2配置及Table1"
        ],
        "source_locations": "L77-L81;L319-L354",
        "full_article_read": false,
        "proofs_read_in_full": false,
        "all_tables_read": false,
        "latest_revision_reviewed": false,
        "version": "v1",
        "limitations": "只审读列明关键段与检索返回内容，不声称逐段通读、穷尽全部附录或确认最新修订。"
      },
      "cost_assessment": {
        "available_resource": "one NVIDIA A100 or A30, or local Mac; no multi-GPU assumption",
        "paper_configuration_and_scale": "SD1.4，5种参数化×SFT/RWL/PPO；side12M，LoRA17M；SFT1000/RWL2000/PPO2400步；HPS/CLIP/PickScore及人工评测。",
        "decision_zh": "完整15训练臂、SD1.4与多个reward模型/图像生成评测及数千步PPO，按配置预计多小时以上，不在单A100/A30短预算；非本地测量。",
        "basis": "reviewed primary-paper configuration, directly reported resource/runtime where available, otherwise qualitative bounded estimate",
        "local_cost_measurement_performed": false,
        "training_or_model_download_performed": false
      },
      "source_url": "https://arxiv.org/html/2603.06981v1",
      "source_sha256": "225d3c64ba529e58ca3ebbd26192ef238252266dcfd8ad17198f92ceb70268e9",
      "source_hash_kind": "sha256_of_public_metadata_locations_and_chinese_paraphrase_receipt_not_raw_capture_or_article",
      "source_retrieval_reference": "evidence/google39/2603.06981.json",
      "retrieval_utc": {
        "start": "2026-10-20T18:48:00Z",
        "end": "2026-10-20T18:56:04Z",
        "precision": "bounded_session_capture_window_not_per_request_timestamp"
      },
      "official_identity_sources": [
        "https://research.google/pubs/diffusion-controller-framework-algorithms-and-parameterization/"
      ],
      "original_capture_sha256": "09305364d1b95f1f79fc04c4e3429f03ea1bdadb4ee2de8901a6199f944ef2d9"
    },
    {
      "id": "2603.10465",
      "title": "MoXaRt: Audio-Visual Object-Guided Sound Interaction for XR",
      "track": "cross-domain",
      "status": "rejected_scope",
      "review_decision": "rejected_scope",
      "review_done": true,
      "resolution": "out_of_scope",
      "implementation_status": "not_implemented",
      "implementation_authorized_this_run": false,
      "automatic_implementation_queue": false,
      "automatic_retry": false,
      "evolve_integrated": false,
      "promotion_eligible": false,
      "review_decision_scope": "paper_method_scope",
      "full_experiment_resolution": "not_applicable_scope",
      "full_experiment_review_done": true,
      "arxiv_v1_date": "2026-03-11",
      "date_evidence": "current_official_arxiv_html_v1_header",
      "earlier_public_release_exhaustively_checked": false,
      "authorship": {
        "first_author_institution_or_role": "Google",
        "google_priority_basis": "current_arxiv_author_block_has_google_affiliation_or_collaboration",
        "first_author_google_required": false
      },
      "method_summary_zh": "MoXaRt XR音频：AudioScopeV2粗分离、视觉人脸/乐器anchors、分支refinement与teacher-student distillation；贡献偏XR声学交互系统。",
      "experiment_summary_zh": "30条一分钟混音录制与22人user study；视觉3.2M/33M、粗音6M、speech9M/music7M，约2秒交互时延。",
      "public_code": {
        "status": "not_found_in_reviewed_primary_sources",
        "url": null,
        "notes_zh": "已读正文未核到作者完整系统/蒸馏代码。",
        "search_scope": "reviewed arXiv v1 article text, code/repository mentions and linked primary repository when explicitly provided; no exhaustive web-wide absence claim"
      },
      "industrial_online_evidence": {
        "strict_online_ab_confirmed": false,
        "explicit_full_traffic_confirmed": false,
        "notes_zh": "技术评测/受试者study，不是生产A/B。"
      },
      "full_text_review": {
        "scope": "bounded_primary_fulltext_method_experiment_resource_and_release_review",
        "source_url": "https://arxiv.org/html/2603.10465v1",
        "read_sections": [
          "§3.2.1-3.2.4",
          "§5/6技术与用户研究摘要"
        ],
        "source_locations": "L187-L228;L93-L95",
        "full_article_read": false,
        "proofs_read_in_full": false,
        "all_tables_read": false,
        "latest_revision_reviewed": false,
        "version": "v1",
        "limitations": "只审读列明关键段与检索返回内容，不声称逐段通读、穷尽全部附录或确认最新修订。"
      },
      "cost_assessment": {
        "available_resource": "one NVIDIA A100 or A30, or local Mac; no multi-GPU assumption",
        "paper_configuration_and_scale": "30条一分钟混音录制与22人user study；视觉3.2M/33M、粗音6M、speech9M/music7M，约2秒交互时延。",
        "decision_zh": "scope终态，不启动YFCC100M预训练、多专模teacher distillation或受试者研究。",
        "basis": "reviewed primary-paper configuration, directly reported resource/runtime where available, otherwise qualitative bounded estimate",
        "local_cost_measurement_performed": false,
        "training_or_model_download_performed": false
      },
      "source_url": "https://arxiv.org/html/2603.10465v1",
      "source_sha256": "8101409b163ea7e86c321094f2ab2d0606f66f42ebfb15ba6652bed06f1b8772",
      "source_hash_kind": "sha256_of_public_metadata_locations_and_chinese_paraphrase_receipt_not_raw_capture_or_article",
      "source_retrieval_reference": "evidence/google39/2603.10465.json",
      "retrieval_utc": {
        "start": "2026-10-20T18:48:00Z",
        "end": "2026-10-20T18:56:04Z",
        "precision": "bounded_session_capture_window_not_per_request_timestamp"
      },
      "official_identity_sources": [
        "https://research.google/pubs/moxart-audio-visual-object-guided-sound-interaction-for-xr-2/"
      ],
      "original_capture_sha256": "c3ee1863e3e158a2eb138f56722a5ea7083524a6b9e2944ac42819fbd49a090a"
    },
    {
      "id": "2603.14609",
      "title": "GroundSet: A Cadastral-Grounded Dataset for Spatial Understanding with Vector Data",
      "track": "multimodal",
      "status": "deferred",
      "review_decision": "deferred",
      "review_done": true,
      "resolution": "not_reproduced_high_cost",
      "implementation_status": "not_implemented",
      "implementation_authorized_this_run": false,
      "automatic_implementation_queue": false,
      "automatic_retry": false,
      "evolve_integrated": false,
      "promotion_eligible": false,
      "review_decision_scope": "full_paper_reproduction_closed_without_implementation",
      "full_experiment_resolution": "not_reproduced_high_cost",
      "full_experiment_review_done": true,
      "arxiv_v1_date": "2026-03-15",
      "date_evidence": "current_official_arxiv_html_v1_header",
      "earlier_public_release_exhaustively_checked": false,
      "authorship": {
        "first_author_institution_or_role": "官方Google来源，原文作者机构需以源页面为准",
        "google_priority_basis": "official_google_or_deepmind_listing_inherited_not_current_author_block_confirmation",
        "first_author_google_required": false
      },
      "method_summary_zh": "GroundSet将地籍vector对象与航空影像对齐，自动生成spatial理解任务；vanilla VLM LoRA作为数据贡献baseline，无额外空间网络。",
      "experiment_summary_zh": "85,864km²/20行政区、20cm影像；多VLM在caption/VQA/classification/detection/REC/segmentation上比较；1 epoch LoRA用8×A100约72h。",
      "public_code": {
        "status": "not_found_in_reviewed_primary_sources",
        "url": null,
        "notes_zh": "已读原文未找到作者实现仓库链接。",
        "search_scope": "reviewed arXiv v1 article text, code/repository mentions and linked primary repository when explicitly provided; no exhaustive web-wide absence claim"
      },
      "industrial_online_evidence": {
        "strict_online_ab_confirmed": false,
        "explicit_full_traffic_confirmed": false,
        "notes_zh": "离线遥感benchmark，无生产A/B。"
      },
      "full_text_review": {
        "scope": "bounded_primary_fulltext_method_experiment_resource_and_release_review",
        "source_url": "https://arxiv.org/html/2603.14609v1",
        "read_sections": [
          "§3数据构建",
          "§4.2-4.3"
        ],
        "source_locations": "L93;L221-L238",
        "full_article_read": false,
        "proofs_read_in_full": false,
        "all_tables_read": false,
        "latest_revision_reviewed": false,
        "version": "v1",
        "limitations": "只审读列明关键段与检索返回内容，不声称逐段通读、穷尽全部附录或确认最新修订。"
      },
      "cost_assessment": {
        "available_resource": "one NVIDIA A100 or A30, or local Mac; no multi-GPU assumption",
        "paper_configuration_and_scale": "85,864km²/20行政区、20cm影像；多VLM在caption/VQA/classification/detection/REC/segmentation上比较；1 epoch LoRA用8×A100约72h。",
        "decision_zh": "论文已测8×A100×72h=576 GPUh，仅主训练就远超1×A100/A30预算；不下载遥感语料或模型。",
        "basis": "reviewed primary-paper configuration, directly reported resource/runtime where available, otherwise qualitative bounded estimate",
        "local_cost_measurement_performed": false,
        "training_or_model_download_performed": false
      },
      "source_url": "https://arxiv.org/html/2603.14609v1",
      "source_sha256": "0028b9f91234a8ecab28b888a0e14cd2afcd0793e3d6b2057f168a6e182022b7",
      "source_hash_kind": "sha256_of_public_metadata_locations_and_chinese_paraphrase_receipt_not_raw_capture_or_article",
      "source_retrieval_reference": "evidence/google39/2603.14609.json",
      "retrieval_utc": {
        "start": "2026-10-20T18:48:00Z",
        "end": "2026-10-20T18:56:04Z",
        "precision": "bounded_session_capture_window_not_per_request_timestamp"
      },
      "official_identity_sources": [
        "https://research.google/pubs/groundset-a-cadastral-grounded-dataset-for-spatial-understanding-with-vector-data/"
      ],
      "original_capture_sha256": "615263cb567a68ee2fdd256d05e8e729fb926e0c70f12832580962069929d790"
    },
    {
      "id": "2603.14769",
      "title": "POLCA: Stochastic Generative Optimization with LLM",
      "track": "agent",
      "status": "accepted",
      "review_decision": "accepted",
      "review_done": true,
      "resolution": "accepted_core_not_implemented_this_run",
      "implementation_status": "not_implemented",
      "implementation_authorized_this_run": false,
      "automatic_implementation_queue": false,
      "automatic_retry": false,
      "evolve_integrated": false,
      "promotion_eligible": false,
      "review_decision_scope": "bounded_core_or_evaluation_candidate_only_not_industrial_admission",
      "full_experiment_resolution": "not_reproduced_high_cost",
      "full_experiment_review_done": true,
      "arxiv_v1_date": "2026-03-16",
      "date_evidence": "current_official_arxiv_html_v1_header",
      "earlier_public_release_exhaustively_checked": false,
      "authorship": {
        "first_author_institution_or_role": "University of Wisconsin-Madison",
        "google_priority_basis": "current_arxiv_author_block_has_google_affiliation_or_collaboration",
        "first_author_google_required": false
      },
      "method_summary_zh": "POLCA优先队列追踪候选评价历史、epsilon-Net embedding保持多样性、LLM Summarizer聚合局部历史；按噪声分配探索/利用，非简单evolution标签。",
      "experiment_summary_zh": "tau-bench/HotpotQA/VeriBench/KernelBench；Veri41题每题最多50 compiler calls×3run×多baseline；Kernel16题、Claude3.7与Gemini embedding。",
      "public_code": {
        "status": "released_primary_repository_reachable",
        "url": "https://github.com/rlx-lab/POLCA",
        "notes_zh": "作者论文明确开源，已核论文链接到rlx-lab/POLCA；仓库运行要求尚未下载/执行。",
        "search_scope": "reviewed arXiv v1 article text, code/repository mentions and linked primary repository when explicitly provided; no exhaustive web-wide absence claim",
        "primary_repository_receipt": "evidence/google39-code-primary-receipts.json"
      },
      "industrial_online_evidence": {
        "strict_online_ab_confirmed": false,
        "explicit_full_traffic_confirmed": false,
        "notes_zh": "离线优化任务，无生产A/B或全流量。"
      },
      "full_text_review": {
        "scope": "bounded_primary_fulltext_method_experiment_resource_and_release_review",
        "source_url": "https://arxiv.org/html/2603.14769v1",
        "read_sections": [
          "§1机制",
          "Appendix D VeriBench/KernelBench配置与预算"
        ],
        "source_locations": "L99-L106;L725-L745;L769-L775",
        "full_article_read": false,
        "proofs_read_in_full": false,
        "all_tables_read": false,
        "latest_revision_reviewed": false,
        "version": "v1",
        "limitations": "只审读列明关键段与检索返回内容，不声称逐段通读、穷尽全部附录或确认最新修订。"
      },
      "cost_assessment": {
        "available_resource": "one NVIDIA A100 or A30, or local Mac; no multi-GPU assumption",
        "paper_configuration_and_scale": "tau-bench/HotpotQA/VeriBench/KernelBench；Veri41题每题最多50 compiler calls×3run×多baseline；Kernel16题、Claude3.7与Gemini embedding。",
        "decision_zh": "完整多benchmark多baseline×3/6seed、每题50迭代与付费LLM/API的搜索矩阵成本过高；epsilon-Net/队列有界core可负担，未批准本轮实现。",
        "basis": "reviewed primary-paper configuration, directly reported resource/runtime where available, otherwise qualitative bounded estimate",
        "local_cost_measurement_performed": false,
        "training_or_model_download_performed": false
      },
      "source_url": "https://arxiv.org/html/2603.14769v1",
      "source_sha256": "936c7c6cc3242b8bf6b04e8aaf2402390e960601bed9d535049e424f3017e4d9",
      "source_hash_kind": "sha256_of_public_metadata_locations_and_chinese_paraphrase_receipt_not_raw_capture_or_article",
      "source_retrieval_reference": "evidence/google39/2603.14769.json",
      "retrieval_utc": {
        "start": "2026-10-20T18:48:00Z",
        "end": "2026-10-20T18:56:04Z",
        "precision": "bounded_session_capture_window_not_per_request_timestamp"
      },
      "official_identity_sources": [
        "https://research.google/pubs/polca-stochastic-generative-optimization-with-llm-2/"
      ],
      "original_capture_sha256": "f8fed134d833d46d20cf67e47ad218dbd5babf5210f56120de8d8e418d543112"
    },
    {
      "id": "2603.24591",
      "title": "Vibe Coding XR: Accelerating AI + XR Prototyping with XR Blocks and Gemini",
      "track": "agent",
      "status": "rejected_scope",
      "review_decision": "rejected_scope",
      "review_done": true,
      "resolution": "out_of_scope",
      "implementation_status": "not_implemented",
      "implementation_authorized_this_run": false,
      "automatic_implementation_queue": false,
      "automatic_retry": false,
      "evolve_integrated": false,
      "promotion_eligible": false,
      "review_decision_scope": "paper_method_scope",
      "full_experiment_resolution": "not_applicable_scope",
      "full_experiment_review_done": true,
      "arxiv_v1_date": "2026-03-25",
      "date_evidence": "current_official_arxiv_html_v1_header",
      "earlier_public_release_exhaustively_checked": false,
      "authorship": {
        "first_author_institution_or_role": "Google XR Labs",
        "google_priority_basis": "current_arxiv_author_block_has_google_affiliation_or_collaboration",
        "first_author_google_required": false
      },
      "method_summary_zh": "XR Blocks将WebXR perception/interaction封装semantic primitives，Gemini系统prompt＋templates生成可运行XR脚本；重点是XR开发框架/交互工作流。",
      "experiment_summary_zh": "VCXR60 pilot技术评测，桌面模拟和Android XR headset场景；非通用agent学习/优化算法的受控能力比较。",
      "public_code": {
        "status": "released_primary_repository_reachable",
        "url": "https://github.com/google/xrblocks",
        "notes_zh": "作者官方google/xrblocks链接明确公开。",
        "search_scope": "reviewed arXiv v1 article text, code/repository mentions and linked primary repository when explicitly provided; no exhaustive web-wide absence claim",
        "primary_repository_receipt": "evidence/google39-code-primary-receipts.json"
      },
      "industrial_online_evidence": {
        "strict_online_ab_confirmed": false,
        "explicit_full_traffic_confirmed": false,
        "notes_zh": "pilot技术评测不是生产A/B或全流量。"
      },
      "full_text_review": {
        "scope": "bounded_primary_fulltext_method_experiment_resource_and_release_review",
        "source_url": "https://arxiv.org/html/2603.24591v1",
        "read_sections": [
          "§1/§3框架与workflow",
          "§4 Preliminary Technical Evaluation定位"
        ],
        "source_locations": "L55-L85",
        "full_article_read": false,
        "proofs_read_in_full": false,
        "all_tables_read": false,
        "latest_revision_reviewed": false,
        "version": "v1",
        "limitations": "只审读列明关键段与检索返回内容，不声称逐段通读、穷尽全部附录或确认最新修订。"
      },
      "cost_assessment": {
        "available_resource": "one NVIDIA A100 or A30, or local Mac; no multi-GPU assumption",
        "paper_configuration_and_scale": "VCXR60 pilot技术评测，桌面模拟和Android XR headset场景；非通用agent学习/优化算法的受控能力比较。",
        "decision_zh": "scope终态；不因API演示可用就另开XR实现，也不需大模型/训练来关闭审查。",
        "basis": "reviewed primary-paper configuration, directly reported resource/runtime where available, otherwise qualitative bounded estimate",
        "local_cost_measurement_performed": false,
        "training_or_model_download_performed": false
      },
      "source_url": "https://arxiv.org/html/2603.24591v1",
      "source_sha256": "b8e9b1c9aea42b8ee8efd861663ba85b2e1c96caa851943087fbc31d4f73fc40",
      "source_hash_kind": "sha256_of_public_metadata_locations_and_chinese_paraphrase_receipt_not_raw_capture_or_article",
      "source_retrieval_reference": "evidence/google39/2603.24591.json",
      "retrieval_utc": {
        "start": "2026-10-20T18:48:00Z",
        "end": "2026-10-20T18:56:04Z",
        "precision": "bounded_session_capture_window_not_per_request_timestamp"
      },
      "official_identity_sources": [
        "https://research.google/pubs/vibe-coding-xr-accelerating-ai-xr-prototyping-with-xr-blocks-and-gemini/"
      ],
      "original_capture_sha256": "8cf577ed584880a9fecb1ecb58fa641aba61e929d80a6d4d58d265394717a081"
    },
    {
      "id": "2603.28896",
      "title": "Fisheye3R: Adapting Unified 3D Feed-Forward Foundation Models to Fisheye Lenses",
      "track": "multimodal",
      "status": "deferred",
      "review_decision": "deferred",
      "review_done": true,
      "resolution": "not_reproduced_high_cost",
      "implementation_status": "not_implemented",
      "implementation_authorized_this_run": false,
      "automatic_implementation_queue": false,
      "automatic_retry": false,
      "evolve_integrated": false,
      "promotion_eligible": false,
      "review_decision_scope": "full_paper_reproduction_closed_without_implementation",
      "full_experiment_resolution": "not_reproduced_high_cost",
      "full_experiment_review_done": true,
      "arxiv_v1_date": "2026-03-30",
      "date_evidence": "current_official_arxiv_html_v1_header",
      "earlier_public_release_exhaustively_checked": false,
      "authorship": {
        "first_author_institution_or_role": "Yale University / Google XR",
        "google_priority_basis": "current_arxiv_author_block_has_google_affiliation_or_collaboration",
        "first_author_google_required": false
      },
      "method_summary_zh": "冻结VGGT/pi3/MapAnything，用calibration tokens适配fisheye；camera-type linear classifier控制attention mask保持perspective兼容，支持SL/SSL。",
      "experiment_summary_zh": "ScanNet++/ADT/KITTI360多几何任务；每层K8 tokens、前12层除外；40K步AdamW，gradient checkpointing；4×RTX A6000 48GiB。",
      "public_code": {
        "status": "not_found_in_reviewed_primary_sources",
        "url": null,
        "notes_zh": "已读正文未找到作者Fisheye3R代码仓库。",
        "search_scope": "reviewed arXiv v1 article text, code/repository mentions and linked primary repository when explicitly provided; no exhaustive web-wide absence claim"
      },
      "industrial_online_evidence": {
        "strict_online_ab_confirmed": false,
        "explicit_full_traffic_confirmed": false,
        "notes_zh": "离线3D评测，无生产A/B。"
      },
      "full_text_review": {
        "scope": "bounded_primary_fulltext_method_experiment_resource_and_release_review",
        "source_url": "https://arxiv.org/html/2603.28896v1",
        "read_sections": [
          "§3 masked calibration",
          "训练/评估/硬件配置"
        ],
        "source_locations": "L142-L149;L332-L347",
        "full_article_read": false,
        "proofs_read_in_full": false,
        "all_tables_read": false,
        "latest_revision_reviewed": false,
        "version": "v1",
        "limitations": "只审读列明关键段与检索返回内容，不声称逐段通读、穷尽全部附录或确认最新修订。"
      },
      "cost_assessment": {
        "available_resource": "one NVIDIA A100 or A30, or local Mac; no multi-GPU assumption",
        "paper_configuration_and_scale": "ScanNet++/ADT/KITTI360多几何任务；每层K8 tokens、前12层除外；40K步AdamW，gradient checkpointing；4×RTX A6000 48GiB。",
        "decision_zh": "三个大3Dbackbone×多适配配置×40K步、原文4×48GiB GPU与多数据集图像，完整实验成本过高；不以1分钟CPU相机分类器冒充整个适配算法。",
        "basis": "reviewed primary-paper configuration, directly reported resource/runtime where available, otherwise qualitative bounded estimate",
        "local_cost_measurement_performed": false,
        "training_or_model_download_performed": false
      },
      "source_url": "https://arxiv.org/html/2603.28896v1",
      "source_sha256": "c8ba18b1d695e0193276c207d941d884308b370b35bd65f2cae18d53038daabe",
      "source_hash_kind": "sha256_of_public_metadata_locations_and_chinese_paraphrase_receipt_not_raw_capture_or_article",
      "source_retrieval_reference": "evidence/google39/2603.28896.json",
      "retrieval_utc": {
        "start": "2026-10-20T18:48:00Z",
        "end": "2026-10-20T18:56:04Z",
        "precision": "bounded_session_capture_window_not_per_request_timestamp"
      },
      "official_identity_sources": [
        "https://research.google/pubs/fisheye3r-adapting-unified-3d-feed-forward-foundation-models-to-fisheye-lenses/"
      ],
      "original_capture_sha256": "78773a73fe6ebbd248640a88dba77d20d85b3f0d78e37c20f4debfdd371adfb7"
    },
    {
      "id": "2603.29470",
      "title": "AI-Simulated Expert Panels for Socio-Technical Scenarios and Decision Guidance",
      "track": "agent",
      "status": "rejected_scope",
      "review_decision": "rejected_scope",
      "review_done": true,
      "resolution": "out_of_scope",
      "implementation_status": "not_implemented",
      "implementation_authorized_this_run": false,
      "automatic_implementation_queue": false,
      "automatic_retry": false,
      "evolve_integrated": false,
      "promotion_eligible": false,
      "review_decision_scope": "paper_method_scope",
      "full_experiment_resolution": "not_applicable_scope",
      "full_experiment_review_done": true,
      "arxiv_v1_date": "2026-03-31",
      "date_evidence": "current_official_arxiv_html_v1_header",
      "earlier_public_release_exhaustively_checked": false,
      "authorship": {
        "first_author_institution_or_role": "HTML作者机构不可见，未独立确认",
        "google_priority_basis": "official_google_or_deepmind_listing_inherited_not_current_author_block_confirmation",
        "first_author_google_required": false
      },
      "method_summary_zh": "单Gemini模拟能源专家，CIB cross-impact stochastic pathways＋MCDA stakeholder选择＋情景量化；德国能源转型proof-of-concept，不是agent能力benchmark。",
      "experiment_summary_zh": "5模拟专家、15×3 descriptor states；方法承认CIB一致性不能保证叙事合理；保留超过250K词讨论但缺真实专家对照能力实验。",
      "public_code": {
        "status": "not_found_in_reviewed_primary_sources",
        "url": null,
        "notes_zh": "已读正文仅accompanying resources/transcripts声明，未核到作者算法代码；Google机构身份在HTML未显示，保留原官方源归属，不声称第一作者Google。",
        "search_scope": "reviewed arXiv v1 article text, code/repository mentions and linked primary repository when explicitly provided; no exhaustive web-wide absence claim"
      },
      "industrial_online_evidence": {
        "strict_online_ab_confirmed": false,
        "explicit_full_traffic_confirmed": false,
        "notes_zh": "政策scenario模拟，非生产A/B。"
      },
      "full_text_review": {
        "scope": "bounded_primary_fulltext_method_experiment_resource_and_release_review",
        "source_url": "https://arxiv.org/html/2603.29470v1",
        "read_sections": [
          "§2 Methods",
          "§3.1-3.3 Results"
        ],
        "source_locations": "L69-L77;L138-L170",
        "full_article_read": false,
        "proofs_read_in_full": false,
        "all_tables_read": false,
        "latest_revision_reviewed": false,
        "version": "v1",
        "limitations": "只审读列明关键段与检索返回内容，不声称逐段通读、穷尽全部附录或确认最新修订。"
      },
      "cost_assessment": {
        "available_resource": "one NVIDIA A100 or A30, or local Mac; no multi-GPU assumption",
        "paper_configuration_and_scale": "5模拟专家、15×3 descriptor states；方法承认CIB一致性不能保证叙事合理；保留超过250K词讨论但缺真实专家对照能力实验。",
        "decision_zh": "scope终态；无需把250K词模拟panel重复为能力复现；不以高成本掩盖来源身份未确认。",
        "basis": "reviewed primary-paper configuration, directly reported resource/runtime where available, otherwise qualitative bounded estimate",
        "local_cost_measurement_performed": false,
        "training_or_model_download_performed": false
      },
      "source_url": "https://arxiv.org/html/2603.29470v1",
      "source_sha256": "a0791bb14f867ae57b07c93c835bf67dd3cc3720924dd8c9f9dbbad749fc9e4f",
      "source_hash_kind": "sha256_of_public_metadata_locations_and_chinese_paraphrase_receipt_not_raw_capture_or_article",
      "source_retrieval_reference": "evidence/google39/2603.29470.json",
      "retrieval_utc": {
        "start": "2026-10-20T18:48:00Z",
        "end": "2026-10-20T18:56:04Z",
        "precision": "bounded_session_capture_window_not_per_request_timestamp"
      },
      "official_identity_sources": [
        "https://research.google/pubs/ai-simulated-expert-panels-for-socio-technical-scenarios-and-decision-guidance/"
      ],
      "original_capture_sha256": "c675db6bfb5c45eb3623b55c1ebd898b407bf2acf15622f134a594e891d687ff"
    },
    {
      "id": "2603.29791",
      "title": "Reasoning-Driven Synthetic Data Generation and Evaluation",
      "track": "foundation/agent",
      "status": "accepted",
      "review_decision": "accepted",
      "review_done": true,
      "resolution": "accepted_core_not_implemented_this_run",
      "implementation_status": "not_implemented",
      "implementation_authorized_this_run": false,
      "automatic_implementation_queue": false,
      "automatic_retry": false,
      "evolve_integrated": false,
      "promotion_eligible": false,
      "review_decision_scope": "bounded_core_or_evaluation_candidate_only_not_industrial_admission",
      "full_experiment_resolution": "not_reproduced_high_cost",
      "full_experiment_review_done": true,
      "arxiv_v1_date": "2026-03-31",
      "date_evidence": "current_official_arxiv_html_v1_header",
      "earlier_public_release_exhaustively_checked": false,
      "authorship": {
        "first_author_institution_or_role": "EPFL / Google internship",
        "google_priority_basis": "current_arxiv_author_block_has_google_affiliation_or_collaboration",
        "first_author_google_required": false
      },
      "method_summary_zh": "Simula seedless taxonomy构造、局部meta-prompt/complexification、全局层次节点采样、double critic rejection与calibrated complexity/Elo；需真实M3生成/批判。",
      "experiment_summary_zh": "Gemini2.5Flash；CTI/法律/GSM8K/GlobalMMLU多语言，每数据集512K唯一点、5系统版本；13gram/Jaccard decontamination；BS=N5评分。",
      "public_code": {
        "status": "not_found_in_reviewed_primary_sources",
        "url": null,
        "notes_zh": "已读正文未发现作者实现仓库。",
        "search_scope": "reviewed arXiv v1 article text, code/repository mentions and linked primary repository when explicitly provided; no exhaustive web-wide absence claim"
      },
      "industrial_online_evidence": {
        "strict_online_ab_confirmed": false,
        "explicit_full_traffic_confirmed": false,
        "notes_zh": "离线合成数据/下游评测，未报告生产A/B。"
      },
      "full_text_review": {
        "scope": "bounded_primary_fulltext_method_experiment_resource_and_release_review",
        "source_url": "https://arxiv.org/html/2603.29791v1",
        "read_sections": [
          "§3 Experimental Setup",
          "§3.2数据规模",
          "Appendix E成本示例"
        ],
        "source_locations": "L133-L160;L674-L685",
        "full_article_read": false,
        "proofs_read_in_full": false,
        "all_tables_read": false,
        "latest_revision_reviewed": false,
        "version": "v1",
        "limitations": "只审读列明关键段与检索返回内容，不声称逐段通读、穷尽全部附录或确认最新修订。"
      },
      "cost_assessment": {
        "available_resource": "one NVIDIA A100 or A30, or local Mac; no multi-GPU assumption",
        "paper_configuration_and_scale": "Gemini2.5Flash；CTI/法律/GSM8K/GlobalMMLU多语言，每数据集512K唯一点、5系统版本；13gram/Jaccard decontamination；BS=N5评分。",
        "decision_zh": "每任务512K样本×多版本＋多次critic/复杂度与下游训练，付费多模态API数百万调用量级风险；完整实验成本过高。小有界taxonomy/critic核心可负担，未在本轮批准实现。",
        "basis": "reviewed primary-paper configuration, directly reported resource/runtime where available, otherwise qualitative bounded estimate",
        "local_cost_measurement_performed": false,
        "training_or_model_download_performed": false
      },
      "source_url": "https://arxiv.org/html/2603.29791v1",
      "source_sha256": "31797910eae4658e36df7db22237fecb2dc4b8ee39d5f0bb22fb719906ef0f8c",
      "source_hash_kind": "sha256_of_public_metadata_locations_and_chinese_paraphrase_receipt_not_raw_capture_or_article",
      "source_retrieval_reference": "evidence/google39/2603.29791.json",
      "retrieval_utc": {
        "start": "2026-10-20T18:48:00Z",
        "end": "2026-10-20T18:56:04Z",
        "precision": "bounded_session_capture_window_not_per_request_timestamp"
      },
      "official_identity_sources": [
        "https://research.google/pubs/reasoning-driven-synthetic-data-generation-and-evaluation/"
      ],
      "original_capture_sha256": "d375df94344a0ff1333c2c547bef2fadb196d8736b1146b1f5abc22bbb3fc25e"
    },
    {
      "id": "2604.05364",
      "title": "TFRBench: A Reasoning Benchmark for Evaluating Forecasting Systems",
      "track": "multimodal",
      "status": "accepted",
      "review_decision": "accepted",
      "review_done": true,
      "resolution": "accepted_core_not_implemented_this_run",
      "implementation_status": "not_implemented",
      "implementation_authorized_this_run": false,
      "automatic_implementation_queue": false,
      "automatic_retry": false,
      "evolve_integrated": false,
      "promotion_eligible": false,
      "review_decision_scope": "bounded_core_or_evaluation_candidate_only_not_industrial_admission",
      "full_experiment_resolution": "not_reproduced_high_cost",
      "full_experiment_review_done": true,
      "arxiv_v1_date": "2026-04-07",
      "date_evidence": "current_official_arxiv_html_v1_header",
      "earlier_public_release_exhaustively_checked": false,
      "authorship": {
        "first_author_institution_or_role": "University of Kentucky / Google collaboration",
        "google_priority_basis": "current_arxiv_author_block_has_google_affiliation_or_collaboration",
        "first_author_google_required": false
      },
      "method_summary_zh": "TFRBench reason/search/forecast/verify/summary多agent生成grounded traces，最多K3循环，按MASE与一致性选trace；这是有标签benchmark构建，不可把gold反馈循环当无泄露预测Agent。",
      "experiment_summary_zh": "10数据集/5领域，多Gemini/Claude预测和LLMjudge对比；候选生成验证访问未来ground truth，forecast本身不读；领域外部事件检索包含未来horizon。",
      "public_code": {
        "status": "project_or_data_link_only_no_algorithm_code_confirmed",
        "url": "https://tfrbench.github.io/",
        "notes_zh": "作者benchmark主页明确给出；未在已读正文核到独立算法源码发布。",
        "search_scope": "reviewed arXiv v1 article text, code/repository mentions and linked primary repository when explicitly provided; no exhaustive web-wide absence claim"
      },
      "industrial_online_evidence": {
        "strict_online_ab_confirmed": false,
        "explicit_full_traffic_confirmed": false,
        "notes_zh": "离线forecasting benchmark，无生产A/B。"
      },
      "full_text_review": {
        "scope": "bounded_primary_fulltext_method_experiment_resource_and_release_review",
        "source_url": "https://arxiv.org/html/2604.05364v1",
        "read_sections": [
          "§3 Algorithm1/6steps",
          "§4结果表片段"
        ],
        "source_locations": "L165-L224;L616-L672",
        "full_article_read": false,
        "proofs_read_in_full": false,
        "all_tables_read": false,
        "latest_revision_reviewed": false,
        "version": "v1",
        "limitations": "只审读列明关键段与检索返回内容，不声称逐段通读、穷尽全部附录或确认最新修订。"
      },
      "cost_assessment": {
        "available_resource": "one NVIDIA A100 or A30, or local Mac; no multi-GPU assumption",
        "paper_configuration_and_scale": "10数据集/5领域，多Gemini/Claude预测和LLMjudge对比；候选生成验证访问未来ground truth，forecast本身不读；领域外部事件检索包含未来horizon。",
        "decision_zh": "完整10数据集×多backbone×多候选/3loop＋检索/judge的API矩阵昂贵；小规模benchmark协议与评分接口可负担，未实施。标签隔离必须明确。",
        "basis": "reviewed primary-paper configuration, directly reported resource/runtime where available, otherwise qualitative bounded estimate",
        "local_cost_measurement_performed": false,
        "training_or_model_download_performed": false
      },
      "source_url": "https://arxiv.org/html/2604.05364v1",
      "source_sha256": "573ddf1fa3c2c4aa17ea4776f708c999a3ac96578e5c952f4b709caa48db9e57",
      "source_hash_kind": "sha256_of_public_metadata_locations_and_chinese_paraphrase_receipt_not_raw_capture_or_article",
      "source_retrieval_reference": "evidence/google39/2604.05364.json",
      "retrieval_utc": {
        "start": "2026-10-20T18:48:00Z",
        "end": "2026-10-20T18:56:04Z",
        "precision": "bounded_session_capture_window_not_per_request_timestamp"
      },
      "official_identity_sources": [
        "https://research.google/pubs/tfrbench-a-reasoning-benchmark-for-evaluating-forecasting-systems/"
      ],
      "original_capture_sha256": "9c6c944daeed049319dd0a9e0731743e92e2b8fe1b58b22779510cf928efca71"
    },
    {
      "id": "2604.09803",
      "title": "MAGE: Modality-Agnostic Music Generation and Editing",
      "track": "multimodal",
      "status": "deferred",
      "review_decision": "deferred",
      "review_done": true,
      "resolution": "not_reproduced_high_cost",
      "implementation_status": "not_implemented",
      "implementation_authorized_this_run": false,
      "automatic_implementation_queue": false,
      "automatic_retry": false,
      "evolve_integrated": false,
      "promotion_eligible": false,
      "review_decision_scope": "full_paper_reproduction_closed_without_implementation",
      "full_experiment_resolution": "not_reproduced_high_cost",
      "full_experiment_review_done": true,
      "arxiv_v1_date": "2026-04-10",
      "date_evidence": "current_official_arxiv_html_v1_header",
      "earlier_public_release_exhaustively_checked": false,
      "authorship": {
        "first_author_institution_or_role": "Google (internship)",
        "google_priority_basis": "current_arxiv_author_block_has_google_affiliation_or_collaboration",
        "first_author_google_required": false
      },
      "method_summary_zh": "MAGE Controlled Multimodal FluxFormer latent flow matching、audio-visual nexus时序对齐、乘性cross-gated modulation与dynamic modality masking；统一生成/mixture编辑。",
      "experiment_summary_zh": "MUSIC benchmark；AdamW1e-4，1K epochs且每epoch数据重复5次；70%mixture、20%zero-mixture、10%CFG fallback；Euler2eval与MixWavCodec。",
      "public_code": {
        "status": "not_found_in_reviewed_primary_sources",
        "url": null,
        "notes_zh": "已读正文未发现作者实现仓库。",
        "search_scope": "reviewed arXiv v1 article text, code/repository mentions and linked primary repository when explicitly provided; no exhaustive web-wide absence claim"
      },
      "industrial_online_evidence": {
        "strict_online_ab_confirmed": false,
        "explicit_full_traffic_confirmed": false,
        "notes_zh": "离线音乐评测，无生产A/B。"
      },
      "full_text_review": {
        "scope": "bounded_primary_fulltext_method_experiment_resource_and_release_review",
        "source_url": "https://arxiv.org/html/2604.09803v1",
        "read_sections": [
          "Appendix A模型/训练/推理",
          "评测段定位"
        ],
        "source_locations": "L315-L343",
        "full_article_read": false,
        "proofs_read_in_full": false,
        "all_tables_read": false,
        "latest_revision_reviewed": false,
        "version": "v1",
        "limitations": "只审读列明关键段与检索返回内容，不声称逐段通读、穷尽全部附录或确认最新修订。"
      },
      "cost_assessment": {
        "available_resource": "one NVIDIA A100 or A30, or local Mac; no multi-GPU assumption",
        "paper_configuration_and_scale": "MUSIC benchmark；AdamW1e-4，1K epochs且每epoch数据重复5次；70%mixture、20%zero-mixture、10%CFG fallback；Euler2eval与MixWavCodec。",
        "decision_zh": "1000epoch×5数据重复的音视频flow训练及多评测臂预计多小时/天，原模型/codec多资源，完整实验成本过高；未测量、不下载。",
        "basis": "reviewed primary-paper configuration, directly reported resource/runtime where available, otherwise qualitative bounded estimate",
        "local_cost_measurement_performed": false,
        "training_or_model_download_performed": false
      },
      "source_url": "https://arxiv.org/html/2604.09803v1",
      "source_sha256": "dbf1e594ba8530befc001b170937bb1667df54a0c1e1ddda516dfbcaeccebfd7",
      "source_hash_kind": "sha256_of_public_metadata_locations_and_chinese_paraphrase_receipt_not_raw_capture_or_article",
      "source_retrieval_reference": "evidence/google39/2604.09803.json",
      "retrieval_utc": {
        "start": "2026-10-20T18:48:00Z",
        "end": "2026-10-20T18:56:04Z",
        "precision": "bounded_session_capture_window_not_per_request_timestamp"
      },
      "official_identity_sources": [
        "https://research.google/pubs/mage-modality-agnostic-music-generation-and-editing/"
      ],
      "original_capture_sha256": "0afeb4ee32f6f29bd43f5881b34bc2a643b91fdb29bcaa6bc5b491c9939f8d7d"
    },
    {
      "id": "2604.20329",
      "title": "Image Generators are Generalist Vision Learners",
      "track": "cross-domain",
      "status": "deferred",
      "review_decision": "deferred",
      "review_done": true,
      "resolution": "not_reproduced_high_cost",
      "implementation_status": "not_implemented",
      "implementation_authorized_this_run": false,
      "automatic_implementation_queue": false,
      "automatic_retry": false,
      "evolve_integrated": false,
      "promotion_eligible": false,
      "review_decision_scope": "full_paper_reproduction_closed_without_implementation",
      "full_experiment_resolution": "not_reproduced_high_cost",
      "full_experiment_review_done": true,
      "arxiv_v1_date": "2026-04-22",
      "date_evidence": "current_official_arxiv_html_v1_header",
      "earlier_public_release_exhaustively_checked": false,
      "authorship": {
        "first_author_institution_or_role": "Google / Google DeepMind",
        "google_priority_basis": "current_arxiv_author_block_has_google_affiliation_or_collaboration",
        "first_author_google_required": false
      },
      "method_summary_zh": "Vision Banana对NanoBananaPro instruction-tuning，将segmentation/depth/normals编码为可逆/可解析RGB图像输出；保留原generation数据混合。",
      "experiment_summary_zh": "多2D/3D公开benchmark与生成editing benchmark；depth仅synthetic训练；基础NanoBananaPro原训练数据/可训练checkpoint非公开。",
      "public_code": {
        "status": "project_or_data_link_only_no_algorithm_code_confirmed",
        "url": "https://vision-banana.github.io/",
        "notes_zh": "项目主页公开；未在已读正文核到作者可训练代码、NBP权重或原始training mix。",
        "search_scope": "reviewed arXiv v1 article text, code/repository mentions and linked primary repository when explicitly provided; no exhaustive web-wide absence claim"
      },
      "industrial_online_evidence": {
        "strict_online_ab_confirmed": false,
        "explicit_full_traffic_confirmed": false,
        "notes_zh": "离线zero-shot视觉/生成评测，无生产A/B。"
      },
      "full_text_review": {
        "scope": "bounded_primary_fulltext_method_experiment_resource_and_release_review",
        "source_url": "https://arxiv.org/html/2604.20329v1",
        "read_sections": [
          "§2 Method",
          "Table1",
          "§3.2 metric depth编码/训练协议"
        ],
        "source_locations": "L61-L85;L219-L244",
        "full_article_read": false,
        "proofs_read_in_full": false,
        "all_tables_read": false,
        "latest_revision_reviewed": false,
        "version": "v1",
        "limitations": "只审读列明关键段与检索返回内容，不声称逐段通读、穷尽全部附录或确认最新修订。"
      },
      "cost_assessment": {
        "available_resource": "one NVIDIA A100 or A30, or local Mac; no multi-GPU assumption",
        "paper_configuration_and_scale": "多2D/3D公开benchmark与生成editing benchmark；depth仅synthetic训练；基础NanoBananaPro原训练数据/可训练checkpoint非公开。",
        "decision_zh": "核心要求闭源NBP训练接口与原始generation corpus，单卡无法取得相同训练链路；成本/资源不可获得，不用RGB编码小函数冒充generalist训练。",
        "basis": "reviewed primary-paper configuration, directly reported resource/runtime where available, otherwise qualitative bounded estimate",
        "local_cost_measurement_performed": false,
        "training_or_model_download_performed": false
      },
      "source_url": "https://arxiv.org/html/2604.20329v1",
      "source_sha256": "832a2b2e678fbed6c8a46ed3980d80d94c0d1976f3b8ffb46f152b84c7f0084d",
      "source_hash_kind": "sha256_of_public_metadata_locations_and_chinese_paraphrase_receipt_not_raw_capture_or_article",
      "source_retrieval_reference": "evidence/google39/2604.20329.json",
      "retrieval_utc": {
        "start": "2026-10-20T18:48:00Z",
        "end": "2026-10-20T18:56:04Z",
        "precision": "bounded_session_capture_window_not_per_request_timestamp"
      },
      "official_identity_sources": [
        "https://deepmind.google/research/publications/240658/"
      ],
      "original_capture_sha256": "355340a947a223c202774f95e969adda4b873d7487d64ae1c4e1fdd97eb80d79"
    },
    {
      "id": "2604.23099",
      "title": "ProEval: Proactive Failure Discovery and Efficient Performance Estimation for Generative AI Evaluation",
      "track": "cross-domain",
      "status": "accepted",
      "review_decision": "accepted",
      "review_done": true,
      "resolution": "accepted_core_not_implemented_this_run",
      "implementation_status": "not_implemented",
      "implementation_authorized_this_run": false,
      "automatic_implementation_queue": false,
      "automatic_retry": false,
      "evolve_integrated": false,
      "promotion_eligible": false,
      "review_decision_scope": "bounded_core_or_evaluation_candidate_only_not_industrial_admission",
      "full_experiment_resolution": "not_reproduced_high_cost",
      "full_experiment_review_done": true,
      "arxiv_v1_date": "2026-04-25",
      "date_evidence": "current_official_arxiv_html_v1_header",
      "earlier_public_release_exhaustively_checked": false,
      "authorship": {
        "first_author_institution_or_role": "Google DeepMind",
        "google_priority_basis": "current_arxiv_author_block_has_google_affiliation_or_collaboration",
        "first_author_google_required": false
      },
      "method_summary_zh": "ProEval预训练GP surrogate，Bayesian quadrature估计总体性能与superlevel set failure采样；TPF/RPF/SF迁移；topic UCB1＋LLM failure pattern转置。",
      "experiment_summary_zh": "16LLM/VLM×reasoning/knowledge/safety/text/image多benchmark；failure discovery100迭代；GP使用历史外部模型score，目标test gold不能给evaluated agent。",
      "public_code": {
        "status": "released_primary_repository_reachable",
        "url": "https://github.com/google-deepmind/proeval",
        "notes_zh": "作者明确开源google-deepmind/proeval。",
        "search_scope": "reviewed arXiv v1 article text, code/repository mentions and linked primary repository when explicitly provided; no exhaustive web-wide absence claim",
        "primary_repository_receipt": "evidence/google39-code-primary-receipts.json"
      },
      "industrial_online_evidence": {
        "strict_online_ab_confirmed": false,
        "explicit_full_traffic_confirmed": false,
        "notes_zh": "离线评价方法，无生产A/B或全流量。"
      },
      "full_text_review": {
        "scope": "bounded_primary_fulltext_method_experiment_resource_and_release_review",
        "source_url": "https://arxiv.org/html/2604.23099v1",
        "read_sections": [
          "§2 Algorithm2",
          "§3.1数据/模型/feature",
          "Table5"
        ],
        "source_locations": "L274-L336;L452-L462",
        "full_article_read": false,
        "proofs_read_in_full": false,
        "all_tables_read": false,
        "latest_revision_reviewed": false,
        "version": "v1",
        "limitations": "只审读列明关键段与检索返回内容，不声称逐段通读、穷尽全部附录或确认最新修订。"
      },
      "cost_assessment": {
        "available_resource": "one NVIDIA A100 or A30, or local Mac; no multi-GPU assumption",
        "paper_configuration_and_scale": "16LLM/VLM×reasoning/knowledge/safety/text/image多benchmark；failure discovery100迭代；GP使用历史外部模型score，目标test gold不能给evaluated agent。",
        "decision_zh": "现有历史score上的GP/BQ有界core可CPU负担；完整16模型多benchmark与LLM query synthesis昂贵，不本轮重跑；未批准实现。",
        "basis": "reviewed primary-paper configuration, directly reported resource/runtime where available, otherwise qualitative bounded estimate",
        "local_cost_measurement_performed": false,
        "training_or_model_download_performed": false
      },
      "source_url": "https://arxiv.org/html/2604.23099v1",
      "source_sha256": "5704d72b8d22dedb0b0411bc1ed476c103825a93ca54939f2ccec8e8e3907611",
      "source_hash_kind": "sha256_of_public_metadata_locations_and_chinese_paraphrase_receipt_not_raw_capture_or_article",
      "source_retrieval_reference": "evidence/google39/2604.23099.json",
      "retrieval_utc": {
        "start": "2026-10-20T18:48:00Z",
        "end": "2026-10-20T18:56:04Z",
        "precision": "bounded_session_capture_window_not_per_request_timestamp"
      },
      "official_identity_sources": [
        "https://deepmind.google/research/publications/238239/"
      ],
      "original_capture_sha256": "395e92d90fcc6b0f0fed11f404f21ac7a76330544b5a6ff09485900a560b94c7"
    },
    {
      "id": "2604.27723",
      "title": "Optimized Deferral for Imbalanced Settings",
      "track": "foundation",
      "status": "accepted",
      "review_decision": "accepted",
      "review_done": true,
      "resolution": "accepted_core_not_implemented_this_run",
      "implementation_status": "not_implemented",
      "implementation_authorized_this_run": false,
      "automatic_implementation_queue": false,
      "automatic_retry": false,
      "evolve_integrated": false,
      "promotion_eligible": false,
      "review_decision_scope": "bounded_core_or_evaluation_candidate_only_not_industrial_admission",
      "full_experiment_resolution": "not_implemented",
      "full_experiment_review_done": true,
      "arxiv_v1_date": "2026-04-30",
      "date_evidence": "current_official_arxiv_html_v1_header",
      "earlier_public_release_exhaustively_checked": false,
      "authorship": {
        "first_author_institution_or_role": "Google Research",
        "google_priority_basis": "current_arxiv_author_block_has_google_affiliation_or_collaboration",
        "first_author_google_required": false
      },
      "method_summary_zh": "mild把imbalanced expert deferral转为input-expert cost-sensitive分类，margin surrogate与一致性界；利用真实expert输出训练router，不能让inference读取标签。",
      "experiment_summary_zh": "CIFAR人为synthetic expert覆盖比例7/2/1等；真实LLM routing有Qwen2.5 0.5B/1.5B等；实验区别synthetic oracle专家与真实模型。",
      "public_code": {
        "status": "not_found_in_reviewed_primary_sources",
        "url": null,
        "notes_zh": "已读原文未发现作者代码仓库。",
        "search_scope": "reviewed arXiv v1 article text, code/repository mentions and linked primary repository when explicitly provided; no exhaustive web-wide absence claim"
      },
      "industrial_online_evidence": {
        "strict_online_ab_confirmed": false,
        "explicit_full_traffic_confirmed": false,
        "notes_zh": "离线classification/routing，无生产A/B或全流量。"
      },
      "full_text_review": {
        "scope": "bounded_primary_fulltext_method_experiment_resource_and_release_review",
        "source_url": "https://arxiv.org/html/2604.27723v1",
        "read_sections": [
          "§5.3 Deferral loss",
          "Appendix C.1/C.2"
        ],
        "source_locations": "L306-L326;L536-L572",
        "full_article_read": false,
        "proofs_read_in_full": false,
        "all_tables_read": false,
        "latest_revision_reviewed": false,
        "version": "v1",
        "limitations": "只审读列明关键段与检索返回内容，不声称逐段通读、穷尽全部附录或确认最新修订。"
      },
      "cost_assessment": {
        "available_resource": "one NVIDIA A100 or A30, or local Mac; no multi-GPU assumption",
        "paper_configuration_and_scale": "CIFAR人为synthetic expert覆盖比例7/2/1等；真实LLM routing有Qwen2.5 0.5B/1.5B等；实验区别synthetic oracle专家与真实模型。",
        "decision_zh": "loss/router有界CPU机制可负担；不必下载多个LLM就可审查，但不得把synthetic gold expert诊断当真实routing能力；本轮未批准实施。",
        "basis": "reviewed primary-paper configuration, directly reported resource/runtime where available, otherwise qualitative bounded estimate",
        "local_cost_measurement_performed": false,
        "training_or_model_download_performed": false
      },
      "source_url": "https://arxiv.org/html/2604.27723v1",
      "source_sha256": "69f6576ef83d97e8e782bee058a70029f9baa72f9c93b9d18da64a7b1f8c7a0c",
      "source_hash_kind": "sha256_of_public_metadata_locations_and_chinese_paraphrase_receipt_not_raw_capture_or_article",
      "source_retrieval_reference": "evidence/google39/2604.27723.json",
      "retrieval_utc": {
        "start": "2026-10-20T18:48:00Z",
        "end": "2026-10-20T18:56:04Z",
        "precision": "bounded_session_capture_window_not_per_request_timestamp"
      },
      "official_identity_sources": [
        "https://research.google/pubs/optimized-deferral-for-imbalanced-settings/"
      ],
      "original_capture_sha256": "5f64d901e1a59384b596792ee22bdbda7710258eeeb30dcf17d89bd318c8e7cb"
    },
    {
      "id": "2605.26340",
      "title": "Scientist-One: Verifiable Autonomous Research via Chain-of-Evidence",
      "track": "agent",
      "status": "accepted",
      "review_decision": "accepted",
      "review_done": true,
      "resolution": "accepted_core_not_implemented_this_run",
      "implementation_status": "not_implemented",
      "implementation_authorized_this_run": false,
      "automatic_implementation_queue": false,
      "automatic_retry": false,
      "evolve_integrated": false,
      "promotion_eligible": false,
      "review_decision_scope": "bounded_core_or_evaluation_candidate_only_not_industrial_admission",
      "full_experiment_resolution": "not_reproduced_high_cost",
      "full_experiment_review_done": true,
      "arxiv_v1_date": "2026-05-25",
      "date_evidence": "current_official_arxiv_html_v1_header",
      "earlier_public_release_exhaustively_checked": false,
      "authorship": {
        "first_author_institution_or_role": "Google Cloud AI Research",
        "google_priority_basis": "current_arxiv_author_block_has_google_affiliation_or_collaboration",
        "first_author_google_required": false
      },
      "method_summary_zh": "ScientistOne CoE将claim连接code/evaluator/bibliography，I1 score复算、I2 spec违规、I3引文查询、I4 method-code alignment与CPR；全自动research pipeline维护provenance。",
      "experiment_summary_zh": "ADRS五系统研究任务×五research系统×3run，Gemini3.1Pro，最多20 solver迭代与2小时code窗口；evaluator5次复算、judges多数投票。",
      "public_code": {
        "status": "not_found_in_reviewed_primary_sources",
        "url": null,
        "notes_zh": "已读正文未核到ScientistOne作者独立发布仓库；baseline开源不等于其开源。",
        "search_scope": "reviewed arXiv v1 article text, code/repository mentions and linked primary repository when explicitly provided; no exhaustive web-wide absence claim"
      },
      "industrial_online_evidence": {
        "strict_online_ab_confirmed": false,
        "explicit_full_traffic_confirmed": false,
        "notes_zh": "离线ADRS，不是生产A/B。"
      },
      "full_text_review": {
        "scope": "bounded_primary_fulltext_method_experiment_resource_and_release_review",
        "source_url": "https://arxiv.org/html/2605.26340v1",
        "read_sections": [
          "§5 Integrity Audit",
          "§6 Experiments",
          "Appendix paper investigation"
        ],
        "source_locations": "L200-L246;L463-L471",
        "full_article_read": false,
        "proofs_read_in_full": false,
        "all_tables_read": false,
        "latest_revision_reviewed": false,
        "version": "v1",
        "limitations": "只审读列明关键段与检索返回内容，不声称逐段通读、穷尽全部附录或确认最新修订。"
      },
      "cost_assessment": {
        "available_resource": "one NVIDIA A100 or A30, or local Mac; no multi-GPU assumption",
        "paper_configuration_and_scale": "ADRS五系统研究任务×五research系统×3run，Gemini3.1Pro，最多20 solver迭代与2小时code窗口；evaluator5次复算、judges多数投票。",
        "decision_zh": "5×5×3=75完整research runs，单run上限2小时且可能重试3次；多轮500篇pool/100notes的外部研究；完整实验明显高成本。有界provenance/审计接口可负担但本轮未授权实现。",
        "basis": "reviewed primary-paper configuration, directly reported resource/runtime where available, otherwise qualitative bounded estimate",
        "local_cost_measurement_performed": false,
        "training_or_model_download_performed": false
      },
      "source_url": "https://arxiv.org/html/2605.26340v1",
      "source_sha256": "5d965c86619f27647d82ee82b0170bb417e9b1443a13fe40eba517987abb9415",
      "source_hash_kind": "sha256_of_public_metadata_locations_and_chinese_paraphrase_receipt_not_raw_capture_or_article",
      "source_retrieval_reference": "evidence/google39/2605.26340.json",
      "retrieval_utc": {
        "start": "2026-10-20T18:48:00Z",
        "end": "2026-10-20T18:56:04Z",
        "precision": "bounded_session_capture_window_not_per_request_timestamp"
      },
      "official_identity_sources": [
        "https://research.google/pubs/scientist-one-verifiable-autonomous-research-via-chain-of-evidence/"
      ],
      "original_capture_sha256": "84d107ced83fa6e1c6c0458c7582833da044ab7ee7f933e435a70bd067d2f567"
    },
    {
      "id": "2605.29729",
      "title": "Realistic honeypot evaluations for scheming propensity",
      "track": "cross-domain",
      "status": "deferred",
      "review_decision": "deferred",
      "review_done": true,
      "resolution": "not_reproduced_resource_unavailable",
      "implementation_status": "not_implemented",
      "implementation_authorized_this_run": false,
      "automatic_implementation_queue": false,
      "automatic_retry": false,
      "evolve_integrated": false,
      "promotion_eligible": false,
      "review_decision_scope": "full_paper_reproduction_closed_without_implementation",
      "full_experiment_resolution": "not_reproduced_private_environment",
      "full_experiment_review_done": true,
      "arxiv_v1_date": "2026-05-28",
      "date_evidence": "current_official_arxiv_html_v1_header",
      "earlier_public_release_exhaustively_checked": false,
      "authorship": {
        "first_author_institution_or_role": "Google DeepMind",
        "google_priority_basis": "current_arxiv_author_block_has_google_affiliation_or_collaboration",
        "first_author_google_required": false
      },
      "method_summary_zh": "真实内部alignment代码honeypots，code review与agentic coding分别设sabotage机会，agency/hidden goal提示调节，检测思考/行为trigger；是评价框架非训练改进。",
      "experiment_summary_zh": "code review12honeypot+1baseline/10commit chain；coding11honeypot+2baseline、Antigravity Google code副本；多Gemini模型与提示变体。",
      "public_code": {
        "status": "not_found_in_reviewed_primary_sources",
        "url": null,
        "notes_zh": "Google内部alignment研究代码不公开；已读正文未找到对应可运行公开环境/实现。",
        "search_scope": "reviewed arXiv v1 article text, code/repository mentions and linked primary repository when explicitly provided; no exhaustive web-wide absence claim"
      },
      "industrial_online_evidence": {
        "strict_online_ab_confirmed": false,
        "explicit_full_traffic_confirmed": false,
        "notes_zh": "作者明确deployment-like，不是in-deployment生产A/B；未触发也不构成安全证明。"
      },
      "full_text_review": {
        "scope": "bounded_primary_fulltext_method_experiment_resource_and_release_review",
        "source_url": "https://arxiv.org/html/2605.29729v1",
        "read_sections": [
          "§3 Methods",
          "§3.1 Model prompting",
          "引言部署边界"
        ],
        "source_locations": "L99-L140;L48-L58",
        "full_article_read": false,
        "proofs_read_in_full": false,
        "all_tables_read": false,
        "latest_revision_reviewed": false,
        "version": "v1",
        "limitations": "只审读列明关键段与检索返回内容，不声称逐段通读、穷尽全部附录或确认最新修订。"
      },
      "cost_assessment": {
        "available_resource": "one NVIDIA A100 or A30, or local Mac; no multi-GPU assumption",
        "paper_configuration_and_scale": "code review12honeypot+1baseline/10commit chain；coding11honeypot+2baseline、Antigravity Google code副本；多Gemini模型与提示变体。",
        "decision_zh": "原内部代码环境不可获得，26场景×多model/prompt长coding轨迹与人工/CoT审查成本不明且显著；按成本gate不尝试外部昂贵近似，保留资源不可用原因。",
        "basis": "reviewed primary-paper configuration, directly reported resource/runtime where available, otherwise qualitative bounded estimate",
        "local_cost_measurement_performed": false,
        "training_or_model_download_performed": false
      },
      "source_url": "https://arxiv.org/html/2605.29729v1",
      "source_sha256": "845d4f3e64430e54cbfda1c8adde5b03e2a179a013af3497530fcba52aad7505",
      "source_hash_kind": "sha256_of_public_metadata_locations_and_chinese_paraphrase_receipt_not_raw_capture_or_article",
      "source_retrieval_reference": "evidence/google39/2605.29729.json",
      "retrieval_utc": {
        "start": "2026-10-20T18:48:00Z",
        "end": "2026-10-20T18:56:04Z",
        "precision": "bounded_session_capture_window_not_per_request_timestamp"
      },
      "official_identity_sources": [
        "https://deepmind.google/research/publications/253391/"
      ],
      "original_capture_sha256": "cfc102a6e4bc1bfe01b0cc8211593d92b24a8b5e9dd095fc7791da064bd0ddbe"
    },
    {
      "id": "2605.30174",
      "title": "LiveSVG: Zero-Shot SVG Animation via Video Generation",
      "track": "multimodal",
      "status": "deferred",
      "review_decision": "deferred",
      "review_done": true,
      "resolution": "not_reproduced_high_cost",
      "implementation_status": "not_implemented",
      "implementation_authorized_this_run": false,
      "automatic_implementation_queue": false,
      "automatic_retry": false,
      "evolve_integrated": false,
      "promotion_eligible": false,
      "review_decision_scope": "full_paper_reproduction_closed_without_implementation",
      "full_experiment_resolution": "not_reproduced_high_cost",
      "full_experiment_review_done": true,
      "arxiv_v1_date": "2026-05-28",
      "date_evidence": "current_official_arxiv_html_v1_header",
      "earlier_public_release_exhaustively_checked": false,
      "authorship": {
        "first_author_institution_or_role": "Google / Hebrew University",
        "google_priority_basis": "current_arxiv_author_block_has_google_affiliation_or_collaboration",
        "first_author_google_required": false
      },
      "method_summary_zh": "LiveSVG先视频生成，再用differentiable rasterizer拟合原SVG：group homography＋Bezier offsets，sphere-packing recoloring消歧；无骨骼先验。",
      "experiment_summary_zh": "AniClipart/ChallengeSVG，Veo3.1/Wan/LTX三video variants；每SVG10–20候选视频，Gemini3.1Pro两阶段rerank；GPU优化每成功样本5.2分钟，含AMT偏好study。",
      "public_code": {
        "status": "project_or_data_link_only_no_algorithm_code_confirmed",
        "url": "https://levymsn.github.io/LiveSVG",
        "notes_zh": "项目主页公开；已读正文未核到作者公开实现仓库。",
        "search_scope": "reviewed arXiv v1 article text, code/repository mentions and linked primary repository when explicitly provided; no exhaustive web-wide absence claim"
      },
      "industrial_online_evidence": {
        "strict_online_ab_confirmed": false,
        "explicit_full_traffic_confirmed": false,
        "notes_zh": "AMT pairwise A/B是离线人工偏好，非真实生产流量A/B。"
      },
      "full_text_review": {
        "scope": "bounded_primary_fulltext_method_experiment_resource_and_release_review",
        "source_url": "https://arxiv.org/html/2605.30174v1",
        "read_sections": [
          "§3机制",
          "Appendix F implementation",
          "运行效率与human study"
        ],
        "source_locations": "L87-L92;L404;L440-L475",
        "full_article_read": false,
        "proofs_read_in_full": false,
        "all_tables_read": false,
        "latest_revision_reviewed": false,
        "version": "v1",
        "limitations": "只审读列明关键段与检索返回内容，不声称逐段通读、穷尽全部附录或确认最新修订。"
      },
      "cost_assessment": {
        "available_resource": "one NVIDIA A100 or A30, or local Mac; no multi-GPU assumption",
        "paper_configuration_and_scale": "AniClipart/ChallengeSVG，Veo3.1/Wan/LTX三video variants；每SVG10–20候选视频，Gemini3.1Pro两阶段rerank；GPU优化每成功样本5.2分钟，含AMT偏好study。",
        "decision_zh": "原完整benchmark×10–20生成视频/每样本＋Gemini rerank＋多baseline拟合/人工study高成本；5.2分钟仅拟合成功样本不含视频生成，不能据此声称完整路径便宜。",
        "basis": "reviewed primary-paper configuration, directly reported resource/runtime where available, otherwise qualitative bounded estimate",
        "local_cost_measurement_performed": false,
        "training_or_model_download_performed": false
      },
      "source_url": "https://arxiv.org/html/2605.30174v1",
      "source_sha256": "f819c31e580dbb974239bb2e82076d89ba07c4829462226656c91b573821c351",
      "source_hash_kind": "sha256_of_public_metadata_locations_and_chinese_paraphrase_receipt_not_raw_capture_or_article",
      "source_retrieval_reference": "evidence/google39/2605.30174.json",
      "retrieval_utc": {
        "start": "2026-10-20T18:48:00Z",
        "end": "2026-10-20T18:56:04Z",
        "precision": "bounded_session_capture_window_not_per_request_timestamp"
      },
      "official_identity_sources": [
        "https://research.google/pubs/livesvg-zero-shot-svg-animation-via-video-generation/"
      ],
      "original_capture_sha256": "e77f6a0a179e72df6bacab490fd7c79c96818db35b3148a87d077ca740a067c6"
    },
    {
      "id": "2605.30322",
      "title": "Gram: Assessing sabotage propensities via automated alignment auditing",
      "track": "cross-domain",
      "status": "deferred",
      "review_decision": "deferred",
      "review_done": true,
      "resolution": "not_reproduced_resource_unavailable",
      "implementation_status": "not_implemented",
      "implementation_authorized_this_run": false,
      "automatic_implementation_queue": false,
      "automatic_retry": false,
      "evolve_integrated": false,
      "promotion_eligible": false,
      "review_decision_scope": "full_paper_reproduction_closed_without_implementation",
      "full_experiment_resolution": "not_reproduced_private_evaluation",
      "full_experiment_review_done": true,
      "arxiv_v1_date": "2026-05-28",
      "date_evidence": "current_official_arxiv_html_v1_header",
      "earlier_public_release_exhaustively_checked": false,
      "authorship": {
        "first_author_institution_or_role": "Google DeepMind",
        "google_priority_basis": "current_arxiv_author_block_has_google_affiliation_or_collaboration",
        "first_author_google_required": false
      },
      "method_summary_zh": "Gram对Petri auditor约束减少leading，17 sabotage seed scenarios；investigator将动态trajectory转静态环境、做causal interventions，区分overeagerness/roleplay。",
      "experiment_summary_zh": "Gemini2.5Pro/3.0Pro/3.1Pro；17scenarios仅4公开、13private heldout；LLMjudge与人工复核，原始2–3%sabotage不能泛化真实部署率。",
      "public_code": {
        "status": "not_found_in_reviewed_primary_sources",
        "url": null,
        "notes_zh": "已读正文未找到Gram完整公开实现；13/17 seed明确私有，不应补造。",
        "search_scope": "reviewed arXiv v1 article text, code/repository mentions and linked primary repository when explicitly provided; no exhaustive web-wide absence claim"
      },
      "industrial_online_evidence": {
        "strict_online_ab_confirmed": false,
        "explicit_full_traffic_confirmed": false,
        "notes_zh": "模拟部署前审计，不是生产真实流量A/B。"
      },
      "full_text_review": {
        "scope": "bounded_primary_fulltext_method_experiment_resource_and_release_review",
        "source_url": "https://arxiv.org/html/2605.30322v1",
        "read_sections": [
          "§2.1-2.2",
          "§1 evaluation边界"
        ],
        "source_locations": "L74-L139",
        "full_article_read": false,
        "proofs_read_in_full": false,
        "all_tables_read": false,
        "latest_revision_reviewed": false,
        "version": "v1",
        "limitations": "只审读列明关键段与检索返回内容，不声称逐段通读、穷尽全部附录或确认最新修订。"
      },
      "cost_assessment": {
        "available_resource": "one NVIDIA A100 or A30, or local Mac; no multi-GPU assumption",
        "paper_configuration_and_scale": "Gemini2.5Pro/3.0Pro/3.1Pro；17scenarios仅4公开、13private heldout；LLMjudge与人工复核，原始2–3%sabotage不能泛化真实部署率。",
        "decision_zh": "完整动态auditor×target多模型/多提示/多trajectory＋人工审查，13私有场景不可获得；成本与资源不确定，按gate不启动昂贵probe。",
        "basis": "reviewed primary-paper configuration, directly reported resource/runtime where available, otherwise qualitative bounded estimate",
        "local_cost_measurement_performed": false,
        "training_or_model_download_performed": false
      },
      "source_url": "https://arxiv.org/html/2605.30322v1",
      "source_sha256": "c3cd1ca60b5cd69dd32699f0e65337bc17df1b3c8a37d6c558e57914378156f0",
      "source_hash_kind": "sha256_of_public_metadata_locations_and_chinese_paraphrase_receipt_not_raw_capture_or_article",
      "source_retrieval_reference": "evidence/google39/2605.30322.json",
      "retrieval_utc": {
        "start": "2026-10-20T18:48:00Z",
        "end": "2026-10-20T18:56:04Z",
        "precision": "bounded_session_capture_window_not_per_request_timestamp"
      },
      "official_identity_sources": [
        "https://deepmind.google/research/publications/252981/"
      ],
      "original_capture_sha256": "0ac83fcba9679517b679bc9e535420e36121f9ce3cf92642b590ff352d1bb01e"
    },
    {
      "id": "2606.00131",
      "title": "AI-PROPELLER: Warehouse-Scale Inter-Procedural Code Layout Optimization with AlphaEvolve",
      "track": "agent",
      "status": "deferred",
      "review_decision": "deferred",
      "review_done": true,
      "resolution": "not_reproduced_high_cost",
      "implementation_status": "not_implemented",
      "implementation_authorized_this_run": false,
      "automatic_implementation_queue": false,
      "automatic_retry": false,
      "evolve_integrated": false,
      "promotion_eligible": false,
      "review_decision_scope": "full_paper_reproduction_closed_without_implementation",
      "full_experiment_resolution": "not_reproduced_high_cost",
      "full_experiment_review_done": true,
      "arxiv_v1_date": "2026-05-28",
      "date_evidence": "current_official_arxiv_html_v1_header",
      "earlier_public_release_exhaustively_checked": false,
      "authorship": {
        "first_author_institution_or_role": "Google",
        "google_priority_basis": "current_arxiv_author_block_has_google_affiliation_or_collaboration",
        "first_author_google_required": false
      },
      "method_summary_zh": "AI-Propeller Magellan让AlphaEvolve修改Propeller interprocedural code-layout heuristic，再Vizier调compiler flags；每candidate实际编译/执行硬件counter而非静态代理。",
      "experiment_summary_zh": "Clang100 modules训练12AlphaEvolve迭代/1200Vizier trials/2.7天，每binary10次；48cores96GB受控Xeon；Clang/LevelDB/Redis/Search；Search单独重训。",
      "public_code": {
        "status": "not_found_in_reviewed_primary_sources",
        "url": null,
        "notes_zh": "已读正文未核到AI-Propeller/Magellan完整公开代码；Propeller/LLVM具名上游不等于作者新policy全套已发布。",
        "search_scope": "reviewed arXiv v1 article text, code/repository mentions and linked primary repository when explicitly provided; no exhaustive web-wide absence claim"
      },
      "industrial_online_evidence": {
        "strict_online_ab_confirmed": false,
        "explicit_full_traffic_confirmed": false,
        "notes_zh": "§5.4明确相对deployed Ext-TSP生产Search提升0.23%；但未披露流量分组/规模或明确全流量，不标严格A/B/全流量已满足。",
        "reported": true,
        "metric": "execution performance improvement",
        "delta_percent": 0.23,
        "baseline": "deployed Ext-TSP"
      },
      "full_text_review": {
        "scope": "bounded_primary_fulltext_method_experiment_resource_and_release_review",
        "source_url": "https://arxiv.org/html/2606.00131v1",
        "read_sections": [
          "§4方法",
          "§5训练/评测",
          "§5.4 Warehouse Scale"
        ],
        "source_locations": "L88-L115;L124-L156",
        "full_article_read": false,
        "proofs_read_in_full": false,
        "all_tables_read": false,
        "latest_revision_reviewed": false,
        "version": "v1",
        "limitations": "只审读列明关键段与检索返回内容，不声称逐段通读、穷尽全部附录或确认最新修订。"
      },
      "cost_assessment": {
        "available_resource": "one NVIDIA A100 or A30, or local Mac; no multi-GPU assumption",
        "paper_configuration_and_scale": "Clang100 modules训练12AlphaEvolve迭代/1200Vizier trials/2.7天，每binary10次；48cores96GB受控Xeon；Clang/LevelDB/Redis/Search；Search单独重训。",
        "decision_zh": "原训练2.7天，1200trials×每binary10次、重编译Clang及私有Search大二进制；单A100/A30不替代所需受控CPU与内部workload，完整实验高成本。",
        "basis": "reviewed primary-paper configuration, directly reported resource/runtime where available, otherwise qualitative bounded estimate",
        "local_cost_measurement_performed": false,
        "training_or_model_download_performed": false
      },
      "source_url": "https://arxiv.org/html/2606.00131v1",
      "source_sha256": "08a3e83c916e8107808c35f0df816ffa145cf299a7b50003e30c40a6529499f6",
      "source_hash_kind": "sha256_of_public_metadata_locations_and_chinese_paraphrase_receipt_not_raw_capture_or_article",
      "source_retrieval_reference": "evidence/google39/2606.00131.json",
      "retrieval_utc": {
        "start": "2026-10-20T18:48:00Z",
        "end": "2026-10-20T18:56:04Z",
        "precision": "bounded_session_capture_window_not_per_request_timestamp"
      },
      "official_identity_sources": [
        "https://research.google/pubs/ai-propeller-warehouse-scale-inter-procedural-code-layout-optimization-with-alphaevolve/"
      ],
      "original_capture_sha256": "249d56934ed61afe04227be083adcf36adb42fad6edd7923a2186039f8f870bf"
    },
    {
      "id": "2606.00369",
      "title": "Quantifying the Salience of Geo-Cultural Values for Pluralistic Safety Alignment",
      "track": "cross-domain",
      "status": "rejected_scope",
      "review_decision": "rejected_scope",
      "review_done": true,
      "resolution": "out_of_scope",
      "implementation_status": "not_implemented",
      "implementation_authorized_this_run": false,
      "automatic_implementation_queue": false,
      "automatic_retry": false,
      "evolve_integrated": false,
      "promotion_eligible": false,
      "review_decision_scope": "paper_method_scope",
      "full_experiment_resolution": "not_applicable_scope",
      "full_experiment_review_done": true,
      "arxiv_v1_date": "2026-05-29",
      "date_evidence": "current_official_arxiv_html_v1_header",
      "earlier_public_release_exhaustively_checked": false,
      "date_correction": {
        "prior_audit_date": "2026-07-10",
        "current_v1_date": "2026-05-29",
        "basis": "official arXiv HTML v1 header; prior date was not first publication"
      },
      "authorship": {
        "first_author_institution_or_role": "Columbia University / Google DeepMind internship",
        "google_priority_basis": "current_arxiv_author_block_has_google_affiliation_or_collaboration",
        "first_author_google_required": false
      },
      "method_summary_zh": "对安全标注数据做meta-analysis，IW文化zone与demographics的multilevel logistic/linear mixed models；评价LLM rater surrogate/triage，非新的alignment训练或agent算法。",
      "experiment_summary_zh": "8安全数据集概览/6数据集显著性；item/rater随机效应与文化/人口固定效应；发现LLM不能可靠代替文化多样标注。",
      "public_code": {
        "status": "not_found_in_reviewed_primary_sources",
        "url": null,
        "notes_zh": "已读原文方法与结果未核到作者公开代码仓库。",
        "search_scope": "reviewed arXiv v1 article text, code/repository mentions and linked primary repository when explicitly provided; no exhaustive web-wide absence claim"
      },
      "industrial_online_evidence": {
        "strict_online_ab_confirmed": false,
        "explicit_full_traffic_confirmed": false,
        "notes_zh": "标注/安全评价meta-analysis，无生产A/B。"
      },
      "full_text_review": {
        "scope": "bounded_primary_fulltext_method_experiment_resource_and_release_review",
        "source_url": "https://arxiv.org/html/2606.00369v1",
        "read_sections": [
          "§4.1文化操作化",
          "§4.2 mixed models",
          "摘要实证范围"
        ],
        "source_locations": "L156-L181;L73-L88",
        "full_article_read": false,
        "proofs_read_in_full": false,
        "all_tables_read": false,
        "latest_revision_reviewed": false,
        "version": "v1",
        "limitations": "只审读列明关键段与检索返回内容，不声称逐段通读、穷尽全部附录或确认最新修订。"
      },
      "cost_assessment": {
        "available_resource": "one NVIDIA A100 or A30, or local Mac; no multi-GPU assumption",
        "paper_configuration_and_scale": "8安全数据集概览/6数据集显著性；item/rater随机效应与文化/人口固定效应；发现LLM不能可靠代替文化多样标注。",
        "decision_zh": "scope终态；统计分析本身可负担，不用高成本拒绝；不另开受试者/标注收集。",
        "basis": "reviewed primary-paper configuration, directly reported resource/runtime where available, otherwise qualitative bounded estimate",
        "local_cost_measurement_performed": false,
        "training_or_model_download_performed": false
      },
      "source_url": "https://arxiv.org/html/2606.00369v1",
      "source_sha256": "e4c94a79c808f4f0d536874bb36eac7ce830ebe9ba5d68508b1625a89ff4010d",
      "source_hash_kind": "sha256_of_public_metadata_locations_and_chinese_paraphrase_receipt_not_raw_capture_or_article",
      "source_retrieval_reference": "evidence/google39/2606.00369.json",
      "retrieval_utc": {
        "start": "2026-10-20T18:48:00Z",
        "end": "2026-10-20T18:56:04Z",
        "precision": "bounded_session_capture_window_not_per_request_timestamp"
      },
      "official_identity_sources": [
        "https://deepmind.google/research/publications/225819/"
      ],
      "original_capture_sha256": "d1af8745400e8ffb4c6fd51c07ffedc09969c244bbb733f024dd0f0d0d314d6e"
    },
    {
      "id": "2607.02515",
      "title": "PointDiT: Pixel-Space Diffusion for Monocular Geometry Estimation",
      "track": "multimodal",
      "status": "deferred",
      "review_decision": "deferred",
      "review_done": true,
      "resolution": "not_reproduced_high_cost",
      "implementation_status": "not_implemented",
      "implementation_authorized_this_run": false,
      "automatic_implementation_queue": false,
      "automatic_retry": false,
      "evolve_integrated": false,
      "promotion_eligible": false,
      "review_decision_scope": "full_paper_reproduction_closed_without_implementation",
      "full_experiment_resolution": "not_reproduced_high_cost",
      "full_experiment_review_done": true,
      "arxiv_v1_date": "2026-07-02",
      "date_evidence": "current_official_arxiv_html_v1_header",
      "earlier_public_release_exhaustively_checked": false,
      "authorship": {
        "first_author_institution_or_role": "ETH Zurich / Google",
        "google_priority_basis": "current_arxiv_author_block_has_google_affiliation_or_collaboration",
        "first_author_google_required": false
      },
      "method_summary_zh": "PointDiT从raw3D pointmap patch预测clean point map，frozenDINOv3四层features拼接条件；plainViT、velocity loss、sky球映射与rectified t0 sampling。",
      "experiment_summary_zh": "B223M/L771M/H1.807B；pretrain30epochs，16/16/64H100×12/21/22h；finetune64/64/128H100×2.5/7/5.5h。",
      "public_code": {
        "status": "not_found_in_reviewed_primary_sources",
        "url": null,
        "notes_zh": "已读正文未核到作者公开代码仓库；不据baseline链接断言开源。",
        "search_scope": "reviewed arXiv v1 article text, code/repository mentions and linked primary repository when explicitly provided; no exhaustive web-wide absence claim"
      },
      "industrial_online_evidence": {
        "strict_online_ab_confirmed": false,
        "explicit_full_traffic_confirmed": false,
        "notes_zh": "离线几何评测，无生产A/B。"
      },
      "full_text_review": {
        "scope": "bounded_primary_fulltext_method_experiment_resource_and_release_review",
        "source_url": "https://arxiv.org/html/2607.02515v1",
        "read_sections": [
          "§3.2-3.3",
          "Appendix训练成本Table6"
        ],
        "source_locations": "L137-L163;L455-L460",
        "full_article_read": false,
        "proofs_read_in_full": false,
        "all_tables_read": false,
        "latest_revision_reviewed": false,
        "version": "v1",
        "limitations": "只审读列明关键段与检索返回内容，不声称逐段通读、穷尽全部附录或确认最新修订。"
      },
      "cost_assessment": {
        "available_resource": "one NVIDIA A100 or A30, or local Mac; no multi-GPU assumption",
        "paper_configuration_and_scale": "B223M/L771M/H1.807B；pretrain30epochs，16/16/64H100×12/21/22h；finetune64/64/128H100×2.5/7/5.5h。",
        "decision_zh": "最小B主训练192H100h＋finetune160H100h=352H100h；L784/H2112H100h；均远超单A100/A30预算，数字按论文GPUs×walltime相乘，非本地测量。",
        "basis": "reviewed primary-paper configuration, directly reported resource/runtime where available, otherwise qualitative bounded estimate",
        "local_cost_measurement_performed": false,
        "training_or_model_download_performed": false
      },
      "source_url": "https://arxiv.org/html/2607.02515v1",
      "source_sha256": "c9fd2f768b7e39d7ddef8e1bf57016975cb23e06b042a3c7ca28baa9922efe0f",
      "source_hash_kind": "sha256_of_public_metadata_locations_and_chinese_paraphrase_receipt_not_raw_capture_or_article",
      "source_retrieval_reference": "evidence/google39/2607.02515.json",
      "retrieval_utc": {
        "start": "2026-10-20T18:48:00Z",
        "end": "2026-10-20T18:56:04Z",
        "precision": "bounded_session_capture_window_not_per_request_timestamp"
      },
      "official_identity_sources": [
        "https://research.google/pubs/pointdit-pixel-space-diffusion-for-monocular-geometry-estimation/"
      ],
      "original_capture_sha256": "6e100b01b5a80160cdbbf0a6379a6acd306a92c289877f309ea92cd8463cb1cb"
    },
    {
      "id": "2607.13614",
      "title": "VIP-MINGLE: A Corpus for Videoconference and In-Person Multimodal Interaction in Group Language Engagement",
      "track": "multimodal",
      "status": "rejected_scope",
      "review_decision": "rejected_scope",
      "review_done": true,
      "resolution": "out_of_scope",
      "implementation_status": "not_implemented",
      "implementation_authorized_this_run": false,
      "automatic_implementation_queue": false,
      "automatic_retry": false,
      "evolve_integrated": false,
      "promotion_eligible": false,
      "review_decision_scope": "paper_method_scope",
      "full_experiment_resolution": "not_applicable_scope",
      "full_experiment_review_done": true,
      "arxiv_v1_date": "2026-07-15",
      "date_evidence": "current_official_arxiv_html_v1_header",
      "earlier_public_release_exhaustively_checked": false,
      "authorship": {
        "first_author_institution_or_role": "New York University / Google collaboration",
        "google_priority_basis": "current_arxiv_author_block_has_google_affiliation_or_collaboration",
        "first_author_google_required": false
      },
      "method_summary_zh": "VIP-MINGLE是跨inperson/video conference受试者配对多模态语料，speech timing/language complexity/facial features统计与interaction annotation；非新的通用模型或Agent算法。",
      "experiment_summary_zh": "59h/32groups/105participants；7077个10秒片段、192annotators；mixed-effects/Wilcoxon比较环境domain shift。",
      "public_code": {
        "status": "not_found_in_reviewed_primary_sources",
        "url": null,
        "notes_zh": "已读原文未核到作者完整处理pipeline公开代码。",
        "search_scope": "reviewed arXiv v1 article text, code/repository mentions and linked primary repository when explicitly provided; no exhaustive web-wide absence claim"
      },
      "industrial_online_evidence": {
        "strict_online_ab_confirmed": false,
        "explicit_full_traffic_confirmed": false,
        "notes_zh": "受试者配对研究，非生产A/B。"
      },
      "full_text_review": {
        "scope": "bounded_primary_fulltext_method_experiment_resource_and_release_review",
        "source_url": "https://arxiv.org/html/2607.13614v1",
        "read_sections": [
          "§2数据设计摘要",
          "§3.1-3.3",
          "§4 annotations"
        ],
        "source_locations": "L44-L62;L94-L113",
        "full_article_read": false,
        "proofs_read_in_full": false,
        "all_tables_read": false,
        "latest_revision_reviewed": false,
        "version": "v1",
        "limitations": "只审读列明关键段与检索返回内容，不声称逐段通读、穷尽全部附录或确认最新修订。"
      },
      "cost_assessment": {
        "available_resource": "one NVIDIA A100 or A30, or local Mac; no multi-GPU assumption",
        "paper_configuration_and_scale": "59h/32groups/105participants；7077个10秒片段、192annotators；mixed-effects/Wilcoxon比较环境domain shift。",
        "decision_zh": "scope终态；不重采集59h人类交互，也不以数据分析可CPU运行冒充原语料复现。",
        "basis": "reviewed primary-paper configuration, directly reported resource/runtime where available, otherwise qualitative bounded estimate",
        "local_cost_measurement_performed": false,
        "training_or_model_download_performed": false
      },
      "source_url": "https://arxiv.org/html/2607.13614v1",
      "source_sha256": "37be407ed96b5b57907daf687209446b46da9094132ece85df830a4f3b285206",
      "source_hash_kind": "sha256_of_public_metadata_locations_and_chinese_paraphrase_receipt_not_raw_capture_or_article",
      "source_retrieval_reference": "evidence/google39/2607.13614.json",
      "retrieval_utc": {
        "start": "2026-10-20T18:48:00Z",
        "end": "2026-10-20T18:56:04Z",
        "precision": "bounded_session_capture_window_not_per_request_timestamp"
      },
      "official_identity_sources": [
        "https://research.google/pubs/vip-mingle-a-corpus-for-videoconference-and-in-person-multimodal-interaction-in-group-language-engagement/"
      ],
      "original_capture_sha256": "81daccb2a655ad7d0de298c7cbbe9aa90a14b1137d550e91634137aeb954bba9"
    },
    {
      "id": "2607.25537",
      "title": "Visual prompt engineering for video models",
      "track": "cross-domain",
      "status": "deferred",
      "review_decision": "deferred",
      "review_done": true,
      "resolution": "not_reproduced_high_cost",
      "implementation_status": "not_implemented",
      "implementation_authorized_this_run": false,
      "automatic_implementation_queue": false,
      "automatic_retry": false,
      "evolve_integrated": false,
      "promotion_eligible": false,
      "review_decision_scope": "full_paper_reproduction_closed_without_implementation",
      "full_experiment_resolution": "not_reproduced_high_cost",
      "full_experiment_review_done": true,
      "arxiv_v1_date": "2026-07-28",
      "date_evidence": "current_official_arxiv_html_v1_header",
      "earlier_public_release_exhaustively_checked": false,
      "authorship": {
        "first_author_institution_or_role": "Google DeepMind",
        "google_priority_basis": "current_arxiv_author_block_has_google_affiliation_or_collaboration",
        "first_author_google_required": false
      },
      "method_summary_zh": "VIPE ideator生成视觉改写idea，image editor修改任务图，可选filter保持语义；freeform open-loop或ACE edits，与video test-time scaling比较。",
      "experiment_summary_zh": "6视觉任务；20variants、每edit best5；单实验18,160个720p/8s视频，Veo3.1；论文API成本$3.20/video与约$0.40/VIPE。",
      "public_code": {
        "status": "project_or_data_link_only_no_algorithm_code_confirmed",
        "url": "https://visual-prompt-engineering.github.io/",
        "notes_zh": "项目主页公开；已读原文未核到作者执行代码。",
        "search_scope": "reviewed arXiv v1 article text, code/repository mentions and linked primary repository when explicitly provided; no exhaustive web-wide absence claim"
      },
      "industrial_online_evidence": {
        "strict_online_ab_confirmed": false,
        "explicit_full_traffic_confirmed": false,
        "notes_zh": "离线video reasoning benchmark，非生产A/B。"
      },
      "full_text_review": {
        "scope": "bounded_primary_fulltext_method_experiment_resource_and_release_review",
        "source_url": "https://arxiv.org/html/2607.25537v1",
        "read_sections": [
          "§4 automated VIPE配置",
          "Appendix B预算",
          "§5 test-time scaling"
        ],
        "source_locations": "L151-L162;L177-L180;L335-L337",
        "full_article_read": false,
        "proofs_read_in_full": false,
        "all_tables_read": false,
        "latest_revision_reviewed": false,
        "version": "v1",
        "limitations": "只审读列明关键段与检索返回内容，不声称逐段通读、穷尽全部附录或确认最新修订。"
      },
      "cost_assessment": {
        "available_resource": "one NVIDIA A100 or A30, or local Mac; no multi-GPU assumption",
        "paper_configuration_and_scale": "6视觉任务；20variants、每edit best5；单实验18,160个720p/8s视频，Veo3.1；论文API成本$3.20/video与约$0.40/VIPE。",
        "decision_zh": "仅18,160×$3.20=$58,112视频API费（按论文当时价格估算），尚未含编辑、filter、其他实验；明确高成本，未验证当前价格也未调用付费生成。",
        "basis": "reviewed primary-paper configuration, directly reported resource/runtime where available, otherwise qualitative bounded estimate",
        "local_cost_measurement_performed": false,
        "training_or_model_download_performed": false
      },
      "source_url": "https://arxiv.org/html/2607.25537v1",
      "source_sha256": "22c92799af3dea509393aec62c87ab209b72c7513a039cef8a115f52aff8d6e0",
      "source_hash_kind": "sha256_of_public_metadata_locations_and_chinese_paraphrase_receipt_not_raw_capture_or_article",
      "source_retrieval_reference": "evidence/google39/2607.25537.json",
      "retrieval_utc": {
        "start": "2026-10-20T18:48:00Z",
        "end": "2026-10-20T18:56:04Z",
        "precision": "bounded_session_capture_window_not_per_request_timestamp"
      },
      "official_identity_sources": [
        "https://deepmind.google/research/publications/264392/"
      ],
      "original_capture_sha256": "2a29a954c747d5c62ed63321012c1efbd34524013c49d9a121a08c2ba22c87b0"
    },
    {
      "id": "2608.05578",
      "title": "Detecting Safety Training Modification in Language Models via Activation Analysis",
      "track": "foundation-model",
      "status": "accepted",
      "review_decision": "accepted",
      "review_done": true,
      "resolution": "accepted_core_not_implemented_this_run",
      "implementation_status": "not_implemented",
      "implementation_authorized_this_run": false,
      "automatic_implementation_queue": false,
      "automatic_retry": false,
      "evolve_integrated": false,
      "promotion_eligible": false,
      "review_decision_scope": "bounded_core_or_evaluation_candidate_only_not_industrial_admission",
      "full_experiment_resolution": "not_reproduced_high_cost",
      "full_experiment_review_done": true,
      "arxiv_v1_date": "2026-08-06",
      "date_evidence": "current_official_arxiv_html_v1_header",
      "earlier_public_release_exhaustively_checked": false,
      "authorship": {
        "first_author_institution_or_role": "Google Cloud",
        "google_priority_basis": "current_arxiv_author_block_has_google_affiliation_or_collaboration",
        "first_author_google_required": false
      },
      "method_summary_zh": "AMS三个安全概念各16 contrastive pairs，扫40–80%深度activation取最大separation层，direction/sigma；Tier1阈值与Tier2 baseline cos/drift验证。",
      "experiment_summary_zh": "14model configs/4families；LOO threshold71%(10/14)、20JailbreakBench prompts/model；A10040GB扫描10–40秒但不含模型加载；bootstrap区间宽且部分跨PASS阈值。",
      "public_code": {
        "status": "released_primary_repository_reachable",
        "url": "https://github.com/GoogleCloudPlatform/activation-model-scanner",
        "notes_zh": "作者正式公开Apache2 GoogleCloudPlatform/activation-model-scanner和Zenodo归档。",
        "search_scope": "reviewed arXiv v1 article text, code/repository mentions and linked primary repository when explicitly provided; no exhaustive web-wide absence claim",
        "primary_repository_receipt": "evidence/google39-code-primary-receipts.json"
      },
      "industrial_online_evidence": {
        "strict_online_ab_confirmed": false,
        "explicit_full_traffic_confirmed": false,
        "notes_zh": "离线模型tamper检测，无生产A/B。"
      },
      "full_text_review": {
        "scope": "bounded_primary_fulltext_method_experiment_resource_and_release_review",
        "source_url": "https://arxiv.org/html/2608.05578v1",
        "read_sections": [
          "§IV-F algorithm",
          "Resource requirements",
          "Data availability/摘要统计"
        ],
        "source_locations": "L242-L277;L583-L602;L631-L646",
        "full_article_read": false,
        "proofs_read_in_full": false,
        "all_tables_read": false,
        "latest_revision_reviewed": false,
        "version": "v1",
        "limitations": "只审读列明关键段与检索返回内容，不声称逐段通读、穷尽全部附录或确认最新修订。"
      },
      "cost_assessment": {
        "available_resource": "one NVIDIA A100 or A30, or local Mac; no multi-GPU assumption",
        "paper_configuration_and_scale": "14model configs/4families；LOO threshold71%(10/14)、20JailbreakBench prompts/model；A10040GB扫描10–40秒但不含模型加载；bootstrap区间宽且部分跨PASS阈值。",
        "decision_zh": "单已缓存3B/7B模型扫描本身低成本，无需训练；14config权重总下载较大，完整矩阵不本轮下载。可负担core候选not_implemented；不能据71%判生产安全扫描已validated。",
        "basis": "reviewed primary-paper configuration, directly reported resource/runtime where available, otherwise qualitative bounded estimate",
        "local_cost_measurement_performed": false,
        "training_or_model_download_performed": false
      },
      "source_url": "https://arxiv.org/html/2608.05578v1",
      "source_sha256": "724cd5f758a0b87229282126e12cddcc3f3e0116203ff551a5fbf874bf06fd53",
      "source_hash_kind": "sha256_of_public_metadata_locations_and_chinese_paraphrase_receipt_not_raw_capture_or_article",
      "source_retrieval_reference": "evidence/google39/2608.05578.json",
      "retrieval_utc": {
        "start": "2026-10-20T18:48:00Z",
        "end": "2026-10-20T18:56:04Z",
        "precision": "bounded_session_capture_window_not_per_request_timestamp"
      },
      "official_identity_sources": [
        "https://research.google/pubs/ams-detecting-unsafe-and-tampered-language-models-via-activation-analysis/"
      ],
      "original_capture_sha256": "507e88fac3f841a6b67a172f5785459006178421a5024321026130de6cad17cb"
    },
    {
      "id": "2608.11367",
      "title": "Gaze Target Estimation Anywhere with Concepts",
      "track": "multimodal",
      "status": "deferred",
      "review_decision": "deferred",
      "review_done": true,
      "resolution": "not_reproduced_high_cost",
      "implementation_status": "not_implemented",
      "implementation_authorized_this_run": false,
      "automatic_implementation_queue": false,
      "automatic_retry": false,
      "evolve_integrated": false,
      "promotion_eligible": false,
      "review_decision_scope": "full_paper_reproduction_closed_without_implementation",
      "full_experiment_resolution": "not_reproduced_high_cost",
      "full_experiment_review_done": true,
      "arxiv_v1_date": "2026-08-11",
      "date_evidence": "current_official_arxiv_html_v1_header",
      "earlier_public_release_exhaustively_checked": false,
      "authorship": {
        "first_author_institution_or_role": "University of Illinois Urbana-Champaign / Google",
        "google_priority_basis": "current_arxiv_author_block_has_google_affiliation_or_collaboration",
        "first_author_google_required": false
      },
      "method_summary_zh": "PGE promptable subject识别与gaze统一；frozen图文encoders＋detector transformer同时localization/inframe/heatmap；120K Gaze-Co由MLLM概念生成与human verification。",
      "experiment_summary_zh": "119,525train samples/3公开test及私有临床OOD；25+5epochs、batch128、4×H100；L40S推理与DigiLens ARGO smartglass。",
      "public_code": {
        "status": "released_primary_repository_reachable",
        "url": "https://github.com/IrohXu/GazeAnywhere",
        "notes_zh": "作者明确公开IrohXu/GazeAnywhere。",
        "search_scope": "reviewed arXiv v1 article text, code/repository mentions and linked primary repository when explicitly provided; no exhaustive web-wide absence claim",
        "primary_repository_receipt": "evidence/google39-code-primary-receipts.json"
      },
      "industrial_online_evidence": {
        "strict_online_ab_confirmed": false,
        "explicit_full_traffic_confirmed": false,
        "notes_zh": "AR实地demo和临床OOD评价非生产A/B/全流量。"
      },
      "full_text_review": {
        "scope": "bounded_primary_fulltext_method_experiment_resource_and_release_review",
        "source_url": "https://arxiv.org/html/2608.11367v1",
        "read_sections": [
          "§5.1 implementation",
          "Appendix数据与model细节"
        ],
        "source_locations": "L242-L246;L427-L473",
        "full_article_read": false,
        "proofs_read_in_full": false,
        "all_tables_read": false,
        "latest_revision_reviewed": false,
        "version": "v1",
        "limitations": "只审读列明关键段与检索返回内容，不声称逐段通读、穷尽全部附录或确认最新修订。"
      },
      "cost_assessment": {
        "available_resource": "one NVIDIA A100 or A30, or local Mac; no multi-GPU assumption",
        "paper_configuration_and_scale": "119,525train samples/3公开test及私有临床OOD；25+5epochs、batch128、4×H100；L40S推理与DigiLens ARGO smartglass。",
        "decision_zh": "4H100×30epoch、多backbone/baseline与120K multimodal数据/MLLM构建，完整实验明显超单A100/A30短预算；私有临床数据不可替代作论文级结论。",
        "basis": "reviewed primary-paper configuration, directly reported resource/runtime where available, otherwise qualitative bounded estimate",
        "local_cost_measurement_performed": false,
        "training_or_model_download_performed": false
      },
      "source_url": "https://arxiv.org/html/2608.11367v1",
      "source_sha256": "4ca88f4973bfbd30b7b993cc917566f0a0e815d2f4bce90e53e0fffa25b32375",
      "source_hash_kind": "sha256_of_public_metadata_locations_and_chinese_paraphrase_receipt_not_raw_capture_or_article",
      "source_retrieval_reference": "evidence/google39/2608.11367.json",
      "retrieval_utc": {
        "start": "2026-10-20T18:48:00Z",
        "end": "2026-10-20T18:56:04Z",
        "precision": "bounded_session_capture_window_not_per_request_timestamp"
      },
      "official_identity_sources": [
        "https://research.google/pubs/gaze-target-estimation-anywhere-with-concepts/"
      ],
      "original_capture_sha256": "1c271c22067c37bcb0e6ab50b9c4e834c4c04588675f7a36c48021c7d6d8729e"
    },
    {
      "id": "2609.01588",
      "title": "Designing Proactive Thought Partners for Writing",
      "track": "cross-domain",
      "status": "rejected_scope",
      "review_decision": "rejected_scope",
      "review_done": true,
      "resolution": "out_of_scope",
      "implementation_status": "not_implemented",
      "implementation_authorized_this_run": false,
      "automatic_implementation_queue": false,
      "automatic_retry": false,
      "evolve_integrated": false,
      "promotion_eligible": false,
      "review_decision_scope": "paper_method_scope",
      "full_experiment_resolution": "not_applicable_scope",
      "full_experiment_review_done": true,
      "arxiv_v1_date": "2026-09-01",
      "date_evidence": "current_official_arxiv_html_v1_header",
      "earlier_public_release_exhaustively_checked": false,
      "authorship": {
        "first_author_institution_or_role": "Google DeepMind collaboration",
        "google_priority_basis": "current_arxiv_author_block_has_google_affiliation_or_collaboration",
        "first_author_google_required": false
      },
      "method_summary_zh": "可配置role/timing的proactive writing thought partners，NextJS/BlockNote＋Gemini decision engine；核心贡献为technology probe交互design space。",
      "experiment_summary_zh": "16人一周diary/interview；FlashLite选择partner<1s，Flash suggestion约8s；作者明确此study不是产品efficacy评测。",
      "public_code": {
        "status": "not_found_in_reviewed_primary_sources",
        "url": null,
        "notes_zh": "已读正文未发现作者完整probe实现仓库，附录仅prompts。",
        "search_scope": "reviewed arXiv v1 article text, code/repository mentions and linked primary repository when explicitly provided; no exhaustive web-wide absence claim"
      },
      "industrial_online_evidence": {
        "strict_online_ab_confirmed": false,
        "explicit_full_traffic_confirmed": false,
        "notes_zh": "探索性一周部署study非生产A/B/全流量。"
      },
      "full_text_review": {
        "scope": "bounded_primary_fulltext_method_experiment_resource_and_release_review",
        "source_url": "https://arxiv.org/html/2609.01588v1",
        "read_sections": [
          "§3设计目标",
          "§4.5 implementation",
          "§5 user study"
        ],
        "source_locations": "L114-L119;L189-L207",
        "full_article_read": false,
        "proofs_read_in_full": false,
        "all_tables_read": false,
        "latest_revision_reviewed": false,
        "version": "v1",
        "limitations": "只审读列明关键段与检索返回内容，不声称逐段通读、穷尽全部附录或确认最新修订。"
      },
      "cost_assessment": {
        "available_resource": "one NVIDIA A100 or A30, or local Mac; no multi-GPU assumption",
        "paper_configuration_and_scale": "16人一周diary/interview；FlashLite选择partner<1s，Flash suggestion约8s；作者明确此study不是产品efficacy评测。",
        "decision_zh": "scope终态；不重复受试者研究，不把廉价API原型冒充能力算法。",
        "basis": "reviewed primary-paper configuration, directly reported resource/runtime where available, otherwise qualitative bounded estimate",
        "local_cost_measurement_performed": false,
        "training_or_model_download_performed": false
      },
      "source_url": "https://arxiv.org/html/2609.01588v1",
      "source_sha256": "e3aa3647a99f4bc82cf95f45f189343c5fa662729ff994dc09c61ee67f770757",
      "source_hash_kind": "sha256_of_public_metadata_locations_and_chinese_paraphrase_receipt_not_raw_capture_or_article",
      "source_retrieval_reference": "evidence/google39/2609.01588.json",
      "retrieval_utc": {
        "start": "2026-10-20T18:48:00Z",
        "end": "2026-10-20T18:56:04Z",
        "precision": "bounded_session_capture_window_not_per_request_timestamp"
      },
      "official_identity_sources": [
        "https://deepmind.google/research/publications/265605/"
      ],
      "original_capture_sha256": "a2472deb884f47f65347b1a36c02f79ef64c37231679a6d43b36f571e217abf0"
    },
    {
      "id": "2609.01834",
      "title": "Architecting Conversational Data Systems for Stateless LLM APIs",
      "track": "agent",
      "status": "rejected_scope",
      "review_decision": "rejected_scope",
      "review_done": true,
      "resolution": "out_of_scope",
      "implementation_status": "not_implemented",
      "implementation_authorized_this_run": false,
      "automatic_implementation_queue": false,
      "automatic_retry": false,
      "evolve_integrated": false,
      "promotion_eligible": false,
      "review_decision_scope": "paper_method_scope",
      "full_experiment_resolution": "not_applicable_scope",
      "full_experiment_review_done": true,
      "arxiv_v1_date": "2026-09-01",
      "date_evidence": "current_official_arxiv_html_v1_header",
      "earlier_public_release_exhaustively_checked": false,
      "authorship": {
        "first_author_institution_or_role": "Google官方收录；本文作者机构未另核",
        "google_priority_basis": "official_google_or_deepmind_listing_inherited_not_current_author_block_confirmation",
        "first_author_google_required": false
      },
      "method_summary_zh": "Hydration Proxy三层架构pattern：UI/secure gateway/relational ledger＋object storage；按stable prefix/full-history hydration设计context治理。",
      "experiment_summary_zh": "原文153行，§2架构、§3cache strategy/§4benefits后结论/参考；Evaluation检索无匹配。没有受控量化实验或具体待复现算法。",
      "public_code": {
        "status": "not_found_in_reviewed_primary_sources",
        "url": null,
        "notes_zh": "已读正文未发现作者公开实现仓库。",
        "search_scope": "reviewed arXiv v1 article text, code/repository mentions and linked primary repository when explicitly provided; no exhaustive web-wide absence claim"
      },
      "industrial_online_evidence": {
        "strict_online_ab_confirmed": false,
        "explicit_full_traffic_confirmed": false,
        "notes_zh": "production-grade是需求措辞，无真实生产A/B/全流量量化证据。"
      },
      "full_text_review": {
        "scope": "bounded_primary_fulltext_method_experiment_resource_and_release_review",
        "source_url": "https://arxiv.org/html/2609.01834v1",
        "read_sections": [
          "§2架构",
          "§3 cache-aware context",
          "结论及全文章节/Evaluation检索"
        ],
        "source_locations": "L53-L124;L127-L138",
        "full_article_read": false,
        "proofs_read_in_full": false,
        "all_tables_read": false,
        "latest_revision_reviewed": false,
        "version": "v1",
        "limitations": "只审读列明关键段与检索返回内容，不声称逐段通读、穷尽全部附录或确认最新修订。"
      },
      "cost_assessment": {
        "available_resource": "one NVIDIA A100 or A30, or local Mac; no multi-GPU assumption",
        "paper_configuration_and_scale": "原文153行，§2架构、§3cache strategy/§4benefits后结论/参考；Evaluation检索无匹配。没有受控量化实验或具体待复现算法。",
        "decision_zh": "概念架构而非可验收论文算法；scope终态，不用成本理由。",
        "basis": "reviewed primary-paper configuration, directly reported resource/runtime where available, otherwise qualitative bounded estimate",
        "local_cost_measurement_performed": false,
        "training_or_model_download_performed": false
      },
      "source_url": "https://arxiv.org/html/2609.01834v1",
      "source_sha256": "2e0a42d1801fd8d1e21c610ca4cdfc8c6e8383cf90d5b1d85d5ae085a4d9de2a",
      "source_hash_kind": "sha256_of_public_metadata_locations_and_chinese_paraphrase_receipt_not_raw_capture_or_article",
      "source_retrieval_reference": "evidence/google39/2609.01834.json",
      "retrieval_utc": {
        "start": "2026-10-20T18:48:00Z",
        "end": "2026-10-20T18:56:04Z",
        "precision": "bounded_session_capture_window_not_per_request_timestamp"
      },
      "official_identity_sources": [
        "https://research.google/pubs/architecting-conversational-data-systems-for-stateless-llm-apis/"
      ],
      "original_capture_sha256": "ba0074c89960cd7d477e33265b0a6010be107f35ce33bd71e226441c9a8cc2f4"
    },
    {
      "id": "2609.12464",
      "title": "Beyond Vector Similarity: Hierarchical Context-Aware Graph RAG vs Standard RAG in Enterprise Code Migration",
      "track": "agent",
      "status": "accepted",
      "review_decision": "accepted",
      "review_done": true,
      "resolution": "accepted_core_not_implemented_this_run",
      "implementation_status": "not_implemented",
      "implementation_authorized_this_run": false,
      "automatic_implementation_queue": false,
      "automatic_retry": false,
      "evolve_integrated": false,
      "promotion_eligible": false,
      "review_decision_scope": "bounded_core_or_evaluation_candidate_only_not_industrial_admission",
      "full_experiment_resolution": "not_implemented",
      "full_experiment_review_done": true,
      "arxiv_v1_date": "2026-09-11",
      "date_evidence": "current_official_arxiv_html_v1_header",
      "earlier_public_release_exhaustively_checked": false,
      "authorship": {
        "first_author_institution_or_role": "Google官方收录；本文作者机构未另核",
        "google_priority_basis": "official_google_or_deepmind_listing_inherited_not_current_author_block_confirmation",
        "first_author_google_required": false
      },
      "method_summary_zh": "HCRG tree-sitter AST入graph、architecture skeleton缓存、拓扑顺序迁移，3hop检索；对照800char/150overlap/top5向量RAG；AST+pyflakes七结构指标。",
      "experiment_summary_zh": "spring-petclinic Java→Python单repo；Gemini2.5Pro T0、Spanner/Vertex1h context cache；DRQ34.8→65.9%、PCC26.7→45.5%，但cyclomatic一致性下降。无多seed/多repo泛化证据。",
      "public_code": {
        "status": "not_found_in_reviewed_primary_sources",
        "url": null,
        "notes_zh": "原文链接输入benchmark repo而非作者pipeline代码；未核到HCRG实现公开。",
        "search_scope": "reviewed arXiv v1 article text, code/repository mentions and linked primary repository when explicitly provided; no exhaustive web-wide absence claim"
      },
      "industrial_online_evidence": {
        "strict_online_ab_confirmed": false,
        "explicit_full_traffic_confirmed": false,
        "notes_zh": "离线代码迁移，非企业生产A/B或全流量。"
      },
      "full_text_review": {
        "scope": "bounded_primary_fulltext_method_experiment_resource_and_release_review",
        "source_url": "https://arxiv.org/html/2609.12464v1",
        "read_sections": [
          "§4.2实验参数",
          "§5.2指标",
          "§6结果"
        ],
        "source_locations": "L128-L175;L212-L225",
        "full_article_read": false,
        "proofs_read_in_full": false,
        "all_tables_read": false,
        "latest_revision_reviewed": false,
        "version": "v1",
        "limitations": "只审读列明关键段与检索返回内容，不声称逐段通读、穷尽全部附录或确认最新修订。"
      },
      "cost_assessment": {
        "available_resource": "one NVIDIA A100 or A30, or local Mac; no multi-GPU assumption",
        "paper_configuration_and_scale": "spring-petclinic Java→Python单repo；Gemini2.5Pro T0、Spanner/Vertex1h context cache；DRQ34.8→65.9%、PCC26.7→45.5%，但cyclomatic一致性下降。无多seed/多repo泛化证据。",
        "decision_zh": "小public repo的AST/拓扑检索与静态指标core可负担，付费云服务可替为同语义本地graph但须实际LLM迁移；未获本轮实施授权，未测总代价。",
        "basis": "reviewed primary-paper configuration, directly reported resource/runtime where available, otherwise qualitative bounded estimate",
        "local_cost_measurement_performed": false,
        "training_or_model_download_performed": false
      },
      "source_url": "https://arxiv.org/html/2609.12464v1",
      "source_sha256": "1aaf750fa1321128cc804923bf9021b1da88aa8789c97f97c365279ec1151ed6",
      "source_hash_kind": "sha256_of_public_metadata_locations_and_chinese_paraphrase_receipt_not_raw_capture_or_article",
      "source_retrieval_reference": "evidence/google39/2609.12464.json",
      "retrieval_utc": {
        "start": "2026-10-20T18:48:00Z",
        "end": "2026-10-20T18:56:04Z",
        "precision": "bounded_session_capture_window_not_per_request_timestamp"
      },
      "official_identity_sources": [
        "https://research.google/pubs/beyond-vector-similarity-hierarchical-context-aware-graph-rag-vs-standard-rag-in-enterprise-code-migration/",
        "https://research.google/pubs/beyond-vector-similarity-hierarchical-context-aware-graph-rag-vs-standard-rag-in-enterprise-code-migration-2/"
      ],
      "original_capture_sha256": "2cdc153fc13b53e23b3f5f872d48212bf270364bafa6bc0a400e93ce67bea7f4"
    }
  ],
  "public_evidence_refined_at_utc": "2026-10-20T19:02:56Z",
  "public_evidence_policy": {
    "raw_capture_archived_only_in_local_private_temporary_storage": true,
    "raw_capture_committed": false,
    "public_receipts_contain_only": "primary URLs, reviewed section/line locations, UTC windows, original capture hashes, short Chinese method/experiment paraphrases",
    "internal_browser_references_retained_publicly": false
  }
}