estelledc · estelledc · Jul 14, 2026 · Jul 14, 2026
diff --git a/SESSION-HANDOFF.md b/SESSION-HANDOFF.md
@@ -2,24 +2,26 @@
 
 > 状态：当前接班入口。旧的批量生产 session 快照已失效，不得用于恢复自动循环；持续运行使用只读 supervisor + 有界 writer epoch。
 
-## 2026-07-14 论文生产与部署 Epoch Contract
+## 2026-07-14 新增 4 篇论文与部署 Epoch Contract
 
 - status：`running`
-- objective：在当前用户明确授权下，新研究并发布 5 篇公开 arXiv 论文笔记，覆盖 agent planning、SWE skills、context engineering、MCP tool use、computer-use benchmark 五条相邻主线。
-- scope：允许新增 `src/content/docs/papers/*.md`、`data/review-receipts/papers/*.json`，以及由构建/atlas/索引/部署门禁确定性更新的派生文件；不改既有笔记正文，不恢复旧批量循环，不修改 policy/threshold。
-- activated_by：`explicit-user-request-2026-07-14-new-5-papers-full-deploy`
+- objective：在当前用户明确授权下，新研究并发布 4 篇公开 arXiv 论文笔记：`OSWorld`、`ToolBench-X`、`MemGym`、`SWE-Bench-CL`，补齐 agent 环境、工具可靠性、长程记忆与 SWE 持续学习四条主线。
+- scope：允许新增 `src/content/docs/papers/*.md`、`data/review-receipts/papers/*.json`，以及由 atlas / note-index / 公开计数文案 / handoff / 部署门禁确定性更新的文件；不改候选队列，不改 policy/threshold，不改旧论文正文语义。
+- activated_by：`explicit-user-request-2026-07-14-new-4-papers-full-deploy`
 - review_after：`2026-07-14`
 - acceptance_checks：
   - `source "$HOME/.nvm/nvm.sh" && nvm use 22.23.1 >/dev/null && npm run status:supervisor`
-  - `source "$HOME/.nvm/nvm.sh" && nvm use 22.23.1 >/dev/null && STUDY_CHANGED_FROM=<base> npm run verify:ci`
+  - `source "$HOME/.nvm/nvm.sh" && nvm use 22.23.1 >/dev/null && node scripts/quality-gate.mjs src/content/docs/papers/{osworld,toolbench-x,memgym,swe-bench-cl}.md`
+  - `source "$HOME/.nvm/nvm.sh" && nvm use 22.23.1 >/dev/null && npm run audit:content-contract`
+  - `source "$HOME/.nvm/nvm.sh" && nvm use 22.23.1 >/dev/null && STUDY_CHANGED_FROM=f487efbcd135faf1e1de9fcd2ccf043437a244fe npm run verify:ci`
   - `git diff --check`
-  - GitHub PR/merge/deploy checks for the final pushed branch.
-- budget：最多 2 个内容小批次（按政策拆为 4 + 1）、5 篇新增 paper、1 个可写切片、1 个本地 writer、1 次部署窗口。
-- external_outcome：5 篇新增论文笔记进入公开 study 站点，并通过 GitHub Pages 线上部署验收；验证状态保持 `UNVERIFIED`，不声明人工审阅或真实执行复现。
-- stop_conditions：规范 Node/npm 不可用；内容契约或红线审计失败且无法在本 scope 内修复；需要修改 policy/threshold、旧正文、候选队列或隐私敏感内容；远端 CI/Pages 连续失败且需要新权限；用户停止。
+  - GitHub PR / merge / Pages deploy checks for the final pushed branch.
+- budget：1 个内容小批次、4 篇新增 paper、1 个可写切片、1 个本地 writer、1 次部署窗口。
+- external_outcome：4 篇新增论文笔记进入公开 study 站点，并通过 GitHub Pages 线上部署验收；验证状态保持 `UNVERIFIED`，不声明实际运行论文 benchmark。
+- stop_conditions：规范 Node/npm 不可用；内容契约或红线审计失败且无法在本 scope 内修复；需要修改 policy/threshold、候选队列或隐私敏感内容；远端 CI/Pages 连续失败且需要新权限；用户停止。
 - superseded_by：`none`
 
-## 当前接班点
+## 上一轮接班背景（保留历史）
 
 - supervisor 状态：`WAIT_HEALTHY`；`scale-budget-exceeded` 已通过批准的 legacy audit review 聚合迁移解除，当前无 hard blocker。
 - scope：launch scope 内的本地 workflow 文档、测试、审计、工具链和站点非内容代码质量维护。
@@ -48,6 +50,26 @@
 - 下一条命令：`source "$HOME/.nvm/nvm.sh" && nvm use 22.23.1 >/dev/null && npm run status:supervisor`；PR 状态用 GitHub API 或浏览器查看 `https://github.com/estelledc/study/pull/26`。
 - 下一位独立 agent 必须先读 `AGENTS.md`，建立 supervisor / epoch contract；不得自动恢复旧数量循环。
 
+## 当前接班点：2026-07-14 4 篇论文本地执行状态
+
+- 起始 ref：`f487efbcd135faf1e1de9fcd2ccf043437a244fe`（origin/main，PR #29 merge commit）。
+- 当前分支：`study/papers-20260714-four-more`。
+- dry-run 结果：`npm run round:dispatch -- --rewrite 0 --new 4 --dry-run` 因 `papers-new short: got 0, need 2` 被阻止；本轮未修改候选队列，改走显式授权的手工 Publication 路径。
+- 已完成切片：
+  1. 规范工具链下 `status:supervisor` 从 Node 版本 blocker 恢复到 `WAIT_HEALTHY`；
+  2. 新增 4 篇 `study-v2` paper note，均为 `STATIC_ANALYSIS` / `UNVERIFIED`；
+  3. 新增 4 个 `study-review-receipt-v1` 静态 review receipt，receipt digest 已通过 `verifyReceiptAgainstNote` 校验；
+  4. `npm run atlas` 刷新 `data/note-index.json`、`papers-atlas.md` 与 agent 主题 atlas chunk；
+  5. 同步公开规模文案：论文 1023、项目 961、总数 1984。
+- 本地已通过：
+  - `node scripts/quality-gate.mjs` 针对 `osworld`、`toolbench-x`、`memgym`、`swe-bench-cl` 四篇；
+  - `npm run audit:counts`；
+  - `npm run audit:content-contract`；
+  - `git ls-files -co --exclude-standard -z | node scripts/audit-public-redlines.mjs --stdin0`；
+  - `npm run build:strict -- --log /tmp/study-build-check.log`；
+  - `git diff --check`。
+- 剩余动作：提交当前变更，提交后重跑 `STUDY_CHANGED_FROM=f487efbcd135faf1e1de9fcd2ccf043437a244fe npm run verify:ci`，随后推送分支、创建 PR、合并并等待 Pages deploy。
+
 ## 当前政策
 
 - 不以内容总数作为本轮目标。

diff --git a/data/note-index.json b/data/note-index.json
@@ -3,16 +3,16 @@
   "taxonomy_version": "taxonomy-v1",
   "stats": {
     "summary": {
-      "total": 1980,
-      "classified": 1933,
+      "total": 1984,
+      "classified": 1937,
       "unclassified": 47,
       "unknown_difficulty": 1975,
       "empty_description": 1970
     },
     "by_area": {
       "papers": {
-        "total": 1019,
-        "classified": 1000,
+        "total": 1023,
+        "classified": 1004,
         "unclassified": 19,
         "unknown_difficulty": 1014,
         "empty_description": 1013
@@ -19237,6 +19237,38 @@
         "chunk_route": "/study/atlas/papers/topic-papers-agents-and-llm-systems-01/"
       }
     },
+    {
+      "id": "papers::memgym",
+      "area": "papers",
+      "slug": "memgym",
+      "title": "MemGym — 给长程 agent memory 做一间健身房",
+      "description": "用 MemGym 区分聊天记忆、执行记忆和可迁移的 agent 经验。",
+      "difficulty": "intermediate",
+      "canonical_topics": [
+        "papers-agents-and-llm-systems"
+      ],
+      "classification": {
+        "state": "classified",
+        "source": "frontmatter-category",
+        "topic_id": "papers-agents-and-llm-systems",
+        "matched_category": "agent",
+        "raw_category": "AI Agent / Memory"
+      },
+      "trust": {
+        "contract_state": "v2",
+        "verification_status": "UNVERIFIED"
+      },
+      "freshness": {
+        "state": "NOT_EVALUATED",
+        "reviewed_at": "2026-07-14",
+        "review_after": null
+      },
+      "route": "/study/papers/memgym/",
+      "atlas": {
+        "chunk_id": "topic-papers-agents-and-llm-systems-01",
+        "chunk_route": "/study/atlas/papers/topic-papers-agents-and-llm-systems-01/"
+      }
+    },
     {
       "id": "papers::mencius-2008",
       "area": "papers",
@@ -22074,6 +22106,38 @@
         "chunk_route": "/study/atlas/papers/topic-papers-agents-and-llm-systems-01/"
       }
     },
+    {
+      "id": "papers::osworld",
+      "area": "papers",
+      "slug": "osworld",
+      "title": "OSWorld — 把 GUI agent 放进真正的电脑里考试",
+      "description": "用 OSWorld 理解为什么电脑操作 agent 不能只在网页或脚本环境里评测。",
+      "difficulty": "intermediate",
+      "canonical_topics": [
+        "papers-agents-and-llm-systems"
+      ],
+      "classification": {
+        "state": "classified",
+        "source": "frontmatter-category",
+        "topic_id": "papers-agents-and-llm-systems",
+        "matched_category": "agent",
+        "raw_category": "AI Agent / Computer Use"
+      },
+      "trust": {
+        "contract_state": "v2",
+        "verification_status": "UNVERIFIED"
+      },
+      "freshness": {
+        "state": "NOT_EVALUATED",
+        "reviewed_at": "2026-07-14",
+        "review_after": null
+      },
+      "route": "/study/papers/osworld/",
+      "atlas": {
+        "chunk_id": "topic-papers-agents-and-llm-systems-01",
+        "chunk_route": "/study/atlas/papers/topic-papers-agents-and-llm-systems-01/"
+      }
+    },
     {
       "id": "papers::ot-1989",
       "area": "papers",
@@ -28747,6 +28811,38 @@
         "chunk_route": "/study/atlas/papers/topic-papers-agents-and-llm-systems-01/"
       }
     },
+    {
+      "id": "papers::swe-bench-cl",
+      "area": "papers",
+      "slug": "swe-bench-cl",
+      "title": "SWE-Bench-CL — coding agent 不能只刷静态题",
+      "description": "用 SWE-Bench-CL 理解软件工程 agent 的持续学习、迁移和灾难性遗忘。",
+      "difficulty": "intermediate",
+      "canonical_topics": [
+        "papers-agents-and-llm-systems"
+      ],
+      "classification": {
+        "state": "classified",
+        "source": "frontmatter-category",
+        "topic_id": "papers-agents-and-llm-systems",
+        "matched_category": "agent",
+        "raw_category": "AI Agent / Software Engineering"
+      },
+      "trust": {
+        "contract_state": "v2",
+        "verification_status": "UNVERIFIED"
+      },
+      "freshness": {
+        "state": "NOT_EVALUATED",
+        "reviewed_at": "2026-07-14",
+        "review_after": null
+      },
+      "route": "/study/papers/swe-bench-cl/",
+      "atlas": {
+        "chunk_id": "topic-papers-agents-and-llm-systems-01",
+        "chunk_route": "/study/atlas/papers/topic-papers-agents-and-llm-systems-01/"
+      }
+    },
     {
       "id": "papers::swe-skills-bench-2026",
       "area": "papers",
@@ -29925,6 +30021,38 @@
         "chunk_route": "/study/atlas/papers/topic-papers-compilers-and-programming-language-theory-02/"
       }
     },
+    {
+      "id": "papers::toolbench-x",
+      "area": "papers",
+      "slug": "toolbench-x",
+      "title": "ToolBench-X — 工具会坏时，agent 还能不能把事做完",
+      "description": "用 ToolBench-X 理解 tool-use benchmark 为什么要模拟规格漂移、调用错误、执行失败和结果冲突。",
+      "difficulty": "intermediate",
+      "canonical_topics": [
+        "papers-agents-and-llm-systems"
+      ],
+      "classification": {
+        "state": "classified",
+        "source": "frontmatter-category",
+        "topic_id": "papers-agents-and-llm-systems",
+        "matched_category": "agent",
+        "raw_category": "AI Agent / Tool Use"
+      },
+      "trust": {
+        "contract_state": "v2",
+        "verification_status": "UNVERIFIED"
+      },
+      "freshness": {
+        "state": "NOT_EVALUATED",
+        "reviewed_at": "2026-07-14",
+        "review_after": null
+      },
+      "route": "/study/papers/toolbench-x/",
+      "atlas": {
+        "chunk_id": "topic-papers-agents-and-llm-systems-01",
+        "chunk_route": "/study/atlas/papers/topic-papers-agents-and-llm-systems-01/"
+      }
+    },
     {
       "id": "papers::toolformer",
       "area": "papers",
@@ -62909,7 +63037,7 @@
       "page": 1,
       "pages": 1,
       "route": "/study/atlas/papers/topic-papers-agents-and-llm-systems-01/",
-      "entries": 64
+      "entries": 68
     },
     {
       "id": "topic-papers-ai-safety-and-interpretability-01",

diff --git a/data/review-receipts/papers/memgym.json b/data/review-receipts/papers/memgym.json
@@ -0,0 +1,55 @@
+{
+  "schema_version": "study-review-receipt-v1",
+  "generation": 1,
+  "predecessor_digest_sha256": null,
+  "note": {
+    "area": "papers",
+    "slug": "memgym",
+    "digest_sha256": "9939f45f2bd42198a5d78050f97acf1e28053d60bcaa7fafb3a4cfed290a9fb2"
+  },
+  "source_revision": "arXiv:2605.20833v1",
+  "research_input_sha256": "4d7fd8d84bb0e51701d1bf50f44abac1732cf8935594c9a647ea90b69e8b382f",
+  "reviewers": [
+    {
+      "role": "ZERO_BASE",
+      "reviewer_version": "study-static-review-20260714",
+      "decision": "PASS_WITH_NOTES",
+      "score": 89,
+      "warnings": [
+        "Memory examples are manual toy traces; MemGym evaluation tracks were not executed."
+      ],
+      "execution": {
+        "review_mode": "STATIC_REVIEW",
+        "code_mode": "MANUAL_SIMULATION"
+      }
+    },
+    {
+      "role": "ENGINEER",
+      "reviewer_version": "study-static-review-20260714",
+      "decision": "PASS_WITH_NOTES",
+      "score": 86,
+      "warnings": [
+        "MemRM and memory-isolated score discussion is abstract-level; no reward model run was performed."
+      ],
+      "execution": {
+        "review_mode": "STATIC_REVIEW",
+        "code_mode": "NOT_APPLICABLE"
+      }
+    },
+    {
+      "role": "ACADEMIC",
+      "reviewer_version": "study-static-review-20260714",
+      "decision": "PASS_WITH_NOTES",
+      "score": 85,
+      "warnings": [
+        "Claims are constrained to arXiv metadata and public abstract text."
+      ],
+      "execution": {
+        "review_mode": "STATIC_REVIEW",
+        "code_mode": "NOT_APPLICABLE"
+      }
+    }
+  ],
+  "waivers": [],
+  "created_at": "2026-07-14T11:39:34.000Z"
+}
diff --git a/data/review-receipts/papers/osworld.json b/data/review-receipts/papers/osworld.json
@@ -0,0 +1,55 @@
+{
+  "schema_version": "study-review-receipt-v1",
+  "generation": 1,
+  "predecessor_digest_sha256": null,
+  "note": {
+    "area": "papers",
+    "slug": "osworld",
+    "digest_sha256": "0520c74cc4628692071bad26e15a8206fd649b0a780d0de49787c81d8173cbdb"
+  },
+  "source_revision": "arXiv:2404.07972v2",
+  "research_input_sha256": "e155e4c3d6893198576ac46bc21a92f964f6fb49164692d0074161e0e057c385",
+  "reviewers": [
+    {
+      "role": "ZERO_BASE",
+      "reviewer_version": "study-static-review-20260714",
+      "decision": "PASS_WITH_NOTES",
+      "score": 88,
+      "warnings": [
+        "Computer-use explanation uses a toy file-copy task; OSWorld benchmark was not executed."
+      ],
+      "execution": {
+        "review_mode": "STATIC_REVIEW",
+        "code_mode": "MANUAL_SIMULATION"
+      }
+    },
+    {
+      "role": "ENGINEER",
+      "reviewer_version": "study-static-review-20260714",
+      "decision": "PASS_WITH_NOTES",
+      "score": 86,
+      "warnings": [
+        "Claims are limited to arXiv metadata, abstract, and static environment design reading."
+      ],
+      "execution": {
+        "review_mode": "STATIC_REVIEW",
+        "code_mode": "NOT_APPLICABLE"
+      }
+    },
+    {
+      "role": "ACADEMIC",
+      "reviewer_version": "study-static-review-20260714",
+      "decision": "PASS_WITH_NOTES",
+      "score": 85,
+      "warnings": [
+        "No independent reproduction of the reported success rates was performed."
+      ],
+      "execution": {
+        "review_mode": "STATIC_REVIEW",
+        "code_mode": "NOT_APPLICABLE"
+      }
+    }
+  ],
+  "waivers": [],
+  "created_at": "2026-07-14T11:39:34.000Z"
+}