superlab 0.1.11 → 0.1.13

package/lib/i18n.cjs

@@ -129,6 +129,7 @@ const ZH_SKILL_FILES = {
 - \`.lab/context/mission.md\`
 - \`.lab/context/state.md\`
 - \`.lab/context/data-decisions.md\`
+- \`.lab/context/eval-protocol.md\`
 - \`.lab/config/workflow.json\`
 
 ## 上下文写回
@@ -140,6 +141,7 @@ const ZH_SKILL_FILES = {
 
 - 优先选择能打通全链路的最小实验。
 - 数据、环境或 metric 接线有问题时要尽快失败。
+- run 目标必须对齐已批准的评估协议，而不是只跟随聊天里的临时目标。
 - 记录精确启动命令和输出位置。
 - 持久 run 输出、日志和 checkpoint 写到 \`results_root\`。
 - 图表和可视化写到 \`figures_root\`。
@@ -151,7 +153,11 @@ const ZH_SKILL_FILES = {
 2. 登记 run
 3. 执行最小有意义实验
 4. 标准化原始指标
-5. 校验标准化摘要
+5. 按当前评估协议校验标准化摘要
+
+## 约束
+
+- 不要凭记忆现想指标定义、baseline 行为或对比方法实现；它们必须锚定到已批准评估协议里记录的来源。
 
 ## 交互约束
 
@@ -171,6 +177,7 @@ const ZH_SKILL_FILES = {
 - baseline
 - 主指标
 - 成功阈值
+- evaluation ladder 与 benchmark 扩量 gate
 - verification commands
 - completion_promise
 - 最大迭代轮次
@@ -182,6 +189,7 @@ const ZH_SKILL_FILES = {
 - \`.lab/context/decisions.md\`
 - \`.lab/context/evidence-index.md\`
 - \`.lab/context/data-decisions.md\`
+- \`.lab/context/eval-protocol.md\`
 - \`.lab/config/workflow.json\`
 
 ## 上下文写回
@@ -221,6 +229,8 @@ const ZH_SKILL_FILES = {
 - 持久 run 输出、日志和 checkpoint 放在 \`results_root\`。
 - 图表和可视化放在 \`figures_root\`。
 - 不要把长期结果堆在 \`.lab/changes/<change-id>/runs\` 里。
+- 不要修改指标定义、baseline 语义或对比方法实现，除非评估协议已经记录了来源和与原始实现的偏差。
+- 如果要调整 ladder、样本量或升格 gate，必须继续锚定到带来源的评估协议，而不是靠聊天临时判断。
 
 ## 交互约束
 
@@ -294,6 +304,7 @@ const ZH_SKILL_FILES = {
 - \`.lab/context/state.md\`
 - \`.lab/context/decisions.md\`
 - \`.lab/context/evidence-index.md\`
+- \`.lab/context/eval-protocol.md\`
 
 ## 上下文写回
 
@@ -304,6 +315,9 @@ const ZH_SKILL_FILES = {
 
 - 不能隐藏失败迭代。
 - 每个主要 claim 都要指向已记录的 summary 或 iteration artifact。
+- 主表结构、gate 和最终结果 framing 必须对齐已批准的评估协议。
+- 不要凭记忆重述指标定义、baseline 行为或对比方法实现；直接引用评估协议里记录的来源。
+- 如果报告依赖了对原始指标或原始实现的偏差，必须明确写出这个偏差。
 - 解释优先保守，不要写成营销文案。
 - 要给 \`/lab:write\` 留下清晰 handoff，尤其是 section draft 可以直接引用的证据链接。
 
@@ -919,8 +933,13 @@ const ZH_SKILL_FILES = {
 ## 目标
 
 - Objective:
+- Autonomy level: L2
+- Approval status: draft
 - Allowed stages: run, iterate, review, report
 - Success criteria:
+- Terminal goal type:
+- Terminal goal target:
+- Required terminal artifact:
 
 ## 循环预算
 
@@ -941,6 +960,14 @@ const ZH_SKILL_FILES = {
 - Promotion check command:
 - Promotion command:
 
+## 阶段产物约束
+
+- Run stage contract: write persistent outputs under \`results_root\`.
+- Iterate stage contract: update persistent outputs under \`results_root\`.
+- Review stage contract: update canonical review context such as \`.lab/context/decisions.md\`、\`state.md\`、\`open-questions.md\` or \`evidence-index.md\`.
+- Report stage contract: write the final report to \`<deliverables_root>/report.md\`.
+- Write stage contract: write LaTeX output under \`<deliverables_root>/paper/\`.
+
 ## 升格策略
 
 - Promotion policy:
@@ -954,6 +981,37 @@ const ZH_SKILL_FILES = {
 
 - Stop conditions:
 - Escalation conditions:
+- Canonical promotion writeback: update \`.lab/context/data-decisions.md\`、\`.lab/context/decisions.md\`、\`.lab/context/state.md\` and \`.lab/context/session-brief.md\`.
+`,
+  [path.join(".lab", "context", "auto-outcome.md")]:
+`# 自动结果
+
+## 目标
+
+- Objective:
+- Experiment ladder:
+- Metric glossary:
+- Metric source papers:
+- Metric implementation source:
+- Comparison source papers:
+- Comparison implementation source:
+- Deviation from original implementation:
+- Terminal goal type:
+- Terminal goal target:
+- Required terminal artifact:
+
+## 结果
+
+- Status: idle
+- Goal reached: no
+- Stop reason:
+- Promotion applied: no
+- Final artifact:
+- Final rung:
+- Executed stages:
+- Iterations completed: 0
+- Started at:
+- Finished at:
 `,
   [path.join(".lab", "context", "auto-status.md")]:
 `# 自动模式状态
@@ -965,6 +1023,9 @@ const ZH_SKILL_FILES = {
 - Current command:
 - Active run id:
 - Iteration count: 0
+- Current rung:
+- Watch target:
+- Next rung:
 
 ## 时间
 
@@ -1322,7 +1383,7 @@ ZH_CONTENT[path.join(".lab", ".managed", "templates", "framing.md")] = `# 论文
 ZH_CONTENT[path.join(".codex", "prompts", "lab.md")] = codexPrompt(
   "查看 /lab 研究工作流总览并选择合适阶段",
   "workflow question 或 stage choice",
-  "# `/lab` for Codex\n\n`/lab` 是严格的研究工作流命令族。每次都使用同一套仓库工件和阶段边界。\n\n## 子命令\n\n- `/lab:idea`\n  调研 idea，定义问题与 failure case，归类 contribution 与 breakthrough level，对比现有方法，收束三个一眼就有意义的点，并在实现前保留 approval gate。\n\n- `/lab:data`\n  把已批准的 idea 转成数据集与 benchmark 方案，记录数据集年份、使用过该数据集的论文、下载来源、许可或访问限制，以及 classic-public、recent-strong-public、claim-specific 三类 benchmark 的纳入理由，和 canonical baselines、strong historical baselines、recent strong public methods、closest prior work 四类对比方法的纳入理由。\n\n- `/lab:auto`\n  在不改变 mission、framing 和核心 claims 的前提下，读取 auto-mode 契约并自动编排 `run`、`iterate`、`review`、`report`，必要时扩展数据集、benchmark 和 comparison methods，并在满足升格策略时自动升级 primary package。\n\n- `/lab:framing`\n  通过审计当前领域与相邻领域的术语，锁定 paper-facing 的方法名、模块名、论文题目和 contribution bullets，并在 section 起草前保留 approval gate。\n\n- `/lab:spec`\n  把已批准的 idea 转成 `.lab/changes/<change-id>/` 下的一个 lab change 目录，并在其中写出 `proposal`、`design`、`spec`、`tasks`。\n\n- `/lab:run`\n  执行最小有意义验证运行，登记 run，并生成第一版标准化评估摘要。\n\n- `/lab:iterate`\n  在冻结 mission、阈值、verification commands 与 `completion_promise` 的前提下执行有边界的实验迭代。\n\n- `/lab:review`\n  以 reviewer mode 审查文档或结果，先给短摘要，再输出 findings、fatal flaws、fix priority 和 residual risks。\n\n- `/lab:report`\n  从 runs 和 iterations 工件生成最终研究报告。\n\n- `/lab:write`\n  使用已安装 `lab` skill 下 vendored 的 paper-writing references，把稳定 report 工件转成论文 section。\n\n## 调度规则\n\n- 始终使用 `skills/lab/SKILL.md` 作为工作流合同。\n- 用户显式调用 `/lab:<stage>` 时，要立刻执行该 stage，而不是只推荐别的 `/lab` stage。\n- 先给简洁摘要，再决定是否写工件，最后回报输出路径和下一步。\n- 如果歧义会影响结论，一次只问一个问题；如果有多条可行路径，先给 2-3 个方案再收敛。\n- `/lab:spec` 前应已有经批准的数据集与 benchmark 方案。\n- `/lab:auto` 只编排已批准边界内的执行阶段，不替代手动的 idea/data/framing/spec 决策。\n- `/lab:write` 前必须已有经批准的 `/lab:framing` 工件。\n"
+  "# `/lab` for Codex\n\n`/lab` 是严格的研究工作流命令族。每次都使用同一套仓库工件和阶段边界。\n\n## 子命令\n\n- `/lab:idea`\n  调研 idea，定义问题与 failure case，归类 contribution 与 breakthrough level，对比现有方法，收束三个一眼就有意义的点，并在实现前保留 approval gate。\n\n- `/lab:data`\n  把已批准的 idea 转成数据集与 benchmark 方案，记录数据集年份、使用过该数据集的论文、下载来源、许可或访问限制，以及 classic-public、recent-strong-public、claim-specific 三类 benchmark 的纳入理由，和 canonical baselines、strong historical baselines、recent strong public methods、closest prior work 四类对比方法的纳入理由。\n\n- `/lab:auto`\n  在不改变 mission、framing 和核心 claims 的前提下，读取 eval-protocol 与 auto-mode 契约并自动编排 `run`、`iterate`、`review`、`report`，必要时扩展数据集、benchmark 和 comparison methods，并在满足升格策略时自动升级 primary package。启动前必须选定 autonomy level、声明 terminal goal，并显式批准契约。\n\n- `/lab:framing`\n  通过审计当前领域与相邻领域的术语，锁定 paper-facing 的方法名、模块名、论文题目和 contribution bullets，并在 section 起草前保留 approval gate。\n\n- `/lab:spec`\n  把已批准的 idea 转成 `.lab/changes/<change-id>/` 下的一个 lab change 目录，并在其中写出 `proposal`、`design`、`spec`、`tasks`。\n\n- `/lab:run`\n  执行最小有意义验证运行，登记 run，并生成第一版标准化评估摘要。\n\n- `/lab:iterate`\n  在冻结 mission、阈值、verification commands 与 `completion_promise` 的前提下执行有边界的实验迭代。\n\n- `/lab:review`\n  以 reviewer mode 审查文档或结果，先给短摘要，再输出 findings、fatal flaws、fix priority 和 residual risks。\n\n- `/lab:report`\n  从 runs 和 iterations 工件生成最终研究报告。\n\n- `/lab:write`\n  使用已安装 `lab` skill 下 vendored 的 paper-writing references，把稳定 report 工件转成论文 section。\n\n## 调度规则\n\n- 始终使用 `skills/lab/SKILL.md` 作为工作流合同。\n- 用户显式调用 `/lab:<stage>` 时，要立刻执行该 stage，而不是只推荐别的 `/lab` stage。\n- 先给简洁摘要，再决定是否写工件，最后回报输出路径和下一步。\n- 如果歧义会影响结论，一次只问一个问题；如果有多条可行路径，先给 2-3 个方案再收敛。\n- `/lab:spec` 前应已有经批准的数据集与 benchmark 方案。\n- `/lab:run`、`/lab:iterate`、`/lab:auto`、`/lab:report` 都应遵循 `.lab/context/eval-protocol.md`。\n- `.lab/context/eval-protocol.md` 不只定义主指标和主表，也应定义指标释义、实验阶梯，以及指标和对比实现的来源。\n- `/lab:auto` 只编排已批准边界内的执行阶段，不替代手动的 idea/data/framing/spec 决策。\n- `/lab:write` 前必须已有经批准的 `/lab:framing` 工件。\n"
 );
 
 ZH_CONTENT[path.join(".codex", "prompts", "lab-data.md")] = codexPrompt(
@@ -1334,14 +1395,14 @@ ZH_CONTENT[path.join(".codex", "prompts", "lab-data.md")] = codexPrompt(
 ZH_CONTENT[path.join(".codex", "prompts", "lab-auto.md")] = codexPrompt(
   "在已批准边界内编排自动实验循环",
   "auto mode objective",
-  "使用已安装的 `lab` 技能：`.codex/skills/lab/SKILL.md`。\n\n立刻针对用户当前给出的参数执行 `/lab:auto`，不要只推荐别的 `/lab` 阶段。只有在缺少阻塞性前提时，才明确指出缺什么，并且一次最多追问一个问题。\n\n本命令运行 `/lab:auto` 阶段。它必须读取 `.lab/context/auto-mode.md` 与 `.lab/context/auto-status.md`，在不修改 mission、framing 和核心 claims 的前提下，编排已批准的 `run`、`iterate`、`review`、`report`，轮询长任务完成情况，并在满足升格策略时自动升级 primary package。"
+  "使用已安装的 `lab` 技能：`.codex/skills/lab/SKILL.md`。\n\n立刻针对用户当前给出的参数执行 `/lab:auto`，不要只推荐别的 `/lab` 阶段。只有在缺少阻塞性前提时，才明确指出缺什么，并且一次最多追问一个问题。\n\n本命令运行 `/lab:auto` 阶段。它必须读取 `.lab/context/eval-protocol.md`、`.lab/context/auto-mode.md`、`.lab/context/auto-status.md` 与 `.lab/context/auto-outcome.md`，先确认 autonomy level、approval status 与 terminal goal schema，再把 eval-protocol 里的指标释义、主表计划、来源约束与结构化实验阶梯当作执行依据，在不修改 mission、framing 和核心 claims 的前提下编排已批准的 `run`、`iterate`、`review`、`report`，轮询长任务完成情况；如果声明了 rung，就保持会话活着并按 rung 转移继续推进。"
 );
 
 ZH_CONTENT[path.join(".claude", "commands", "lab.md")] = claudeCommand(
   "LAB",
   "查看 /lab 研究工作流总览并选择合适阶段",
   "workflow, research, overview",
-  "# `/lab` for Claude\n\n`/lab` 是严格的研究工作流命令族。每次都使用同一套仓库工件和阶段边界。\n\n## 子命令\n\n- `/lab:idea`\n  调研 idea，定义问题与 failure case，归类 contribution 与 breakthrough level，对比现有方法，收束三个一眼就有意义的点，并在实现前保留 approval gate。\n\n- `/lab:data`\n  把已批准的 idea 转成数据集与 benchmark 方案，记录数据集年份、使用过该数据集的论文、下载来源、许可或访问限制，以及 classic-public、recent-strong-public、claim-specific 三类 benchmark 的纳入理由，和 canonical baselines、strong historical baselines、recent strong public methods、closest prior work 四类对比方法的纳入理由。\n\n- `/lab:auto`\n  在不改变 mission、framing 和核心 claims 的前提下，读取 auto-mode 契约并自动编排 `run`、`iterate`、`review`、`report`，必要时扩展数据集、benchmark 和 comparison methods，并在满足升格策略时自动升级 primary package。\n\n- `/lab:framing`\n  通过审计当前领域与相邻领域的术语，锁定 paper-facing 的方法名、模块名、论文题目和 contribution bullets，并在 section 起草前保留 approval gate。\n\n- `/lab:spec`\n  把已批准的 idea 转成 `.lab/changes/<change-id>/` 下的一个 lab change 目录，并在其中写出 `proposal`、`design`、`spec`、`tasks`。\n\n- `/lab:run`\n  执行最小有意义验证运行，登记 run，并生成第一版标准化评估摘要。\n\n- `/lab:iterate`\n  在冻结 mission、阈值、verification commands 与 `completion_promise` 的前提下执行有边界的实验迭代。\n\n- `/lab:review`\n  以 reviewer mode 审查文档或结果，先给短摘要，再输出 findings、fatal flaws、fix priority 和 residual risks。\n\n- `/lab:report`\n  从 runs 和 iterations 工件生成最终研究报告。\n\n- `/lab:write`\n  使用已安装 `lab` skill 下 vendored 的 paper-writing references，把稳定 report 工件转成论文 section。\n\n## 调度规则\n\n- 始终使用 `skills/lab/SKILL.md` 作为工作流合同。\n- 用户显式调用 `/lab:<stage>` 时，要立刻执行该 stage，而不是只推荐别的 `/lab` stage。\n- 先给简洁摘要，再决定是否写工件，最后回报输出路径和下一步。\n- 如果歧义会影响结论，一次只问一个问题；如果有多条可行路径，先给 2-3 个方案再收敛。\n- `/lab:spec` 前应已有经批准的数据集与 benchmark 方案。\n- `/lab:auto` 只编排已批准边界内的执行阶段，不替代手动的 idea/data/framing/spec 决策。\n- `/lab:write` 前必须已有经批准的 `/lab:framing` 工件。\n"
+  "# `/lab` for Claude\n\n`/lab` 是严格的研究工作流命令族。每次都使用同一套仓库工件和阶段边界。\n\n## 子命令\n\n- `/lab:idea`\n  调研 idea，定义问题与 failure case，归类 contribution 与 breakthrough level，对比现有方法，收束三个一眼就有意义的点，并在实现前保留 approval gate。\n\n- `/lab:data`\n  把已批准的 idea 转成数据集与 benchmark 方案，记录数据集年份、使用过该数据集的论文、下载来源、许可或访问限制，以及 classic-public、recent-strong-public、claim-specific 三类 benchmark 的纳入理由，和 canonical baselines、strong historical baselines、recent strong public methods、closest prior work 四类对比方法的纳入理由。\n\n- `/lab:auto`\n  在不改变 mission、framing 和核心 claims 的前提下，读取 eval-protocol 与 auto-mode 契约并自动编排 `run`、`iterate`、`review`、`report`，必要时扩展数据集、benchmark 和 comparison methods，并在满足升格策略时自动升级 primary package。启动前必须选定 autonomy level、声明 terminal goal，并显式批准契约。\n\n- `/lab:framing`\n  通过审计当前领域与相邻领域的术语，锁定 paper-facing 的方法名、模块名、论文题目和 contribution bullets，并在 section 起草前保留 approval gate。\n\n- `/lab:spec`\n  把已批准的 idea 转成 `.lab/changes/<change-id>/` 下的一个 lab change 目录，并在其中写出 `proposal`、`design`、`spec`、`tasks`。\n\n- `/lab:run`\n  执行最小有意义验证运行，登记 run，并生成第一版标准化评估摘要。\n\n- `/lab:iterate`\n  在冻结 mission、阈值、verification commands 与 `completion_promise` 的前提下执行有边界的实验迭代。\n\n- `/lab:review`\n  以 reviewer mode 审查文档或结果，先给短摘要，再输出 findings、fatal flaws、fix priority 和 residual risks。\n\n- `/lab:report`\n  从 runs 和 iterations 工件生成最终研究报告。\n\n- `/lab:write`\n  使用已安装 `lab` skill 下 vendored 的 paper-writing references，把稳定 report 工件转成论文 section。\n\n## 调度规则\n\n- 始终使用 `skills/lab/SKILL.md` 作为工作流合同。\n- 用户显式调用 `/lab:<stage>` 时，要立刻执行该 stage，而不是只推荐别的 `/lab` stage。\n- 先给简洁摘要，再决定是否写工件，最后回报输出路径和下一步。\n- 如果歧义会影响结论，一次只问一个问题；如果有多条可行路径，先给 2-3 个方案再收敛。\n- `/lab:spec` 前应已有经批准的数据集与 benchmark 方案。\n- `/lab:run`、`/lab:iterate`、`/lab:auto`、`/lab:report` 都应遵循 `.lab/context/eval-protocol.md`。\n- `.lab/context/eval-protocol.md` 不只定义主指标和主表，也应定义指标释义、实验阶梯，以及指标和对比实现的来源。\n- `/lab:auto` 只编排已批准边界内的执行阶段，不替代手动的 idea/data/framing/spec 决策。\n- `/lab:write` 前必须已有经批准的 `/lab:framing` 工件。\n"
 );
 
 ZH_CONTENT[path.join(".claude", "commands", "lab", "data.md")] = claudeCommand(
@@ -1355,7 +1416,7 @@ ZH_CONTENT[path.join(".claude", "commands", "lab", "auto.md")] = claudeCommand(
   "LAB: Auto",
   "在已批准边界内编排自动实验循环",
   "workflow, research, auto",
-  "使用已安装的 `lab` 技能：`.claude/skills/lab/SKILL.md`。\n\n立刻针对用户当前给出的参数执行 `/lab:auto`，不要只推荐别的 `/lab` 阶段。只有在缺少阻塞性前提时，才明确指出缺什么，并且一次最多追问一个问题。\n\n本命令运行 `/lab:auto` 阶段。它必须读取 `.lab/context/auto-mode.md` 与 `.lab/context/auto-status.md`，在不修改 mission、framing 和核心 claims 的前提下，编排已批准的 `run`、`iterate`、`review`、`report`，轮询长任务完成情况，并在满足升格策略时自动升级 primary package。"
+  "使用已安装的 `lab` 技能：`.claude/skills/lab/SKILL.md`。\n\n立刻针对用户当前给出的参数执行 `/lab:auto`，不要只推荐别的 `/lab` 阶段。只有在缺少阻塞性前提时，才明确指出缺什么，并且一次最多追问一个问题。\n\n本命令运行 `/lab:auto` 阶段。它必须读取 `.lab/context/eval-protocol.md`、`.lab/context/auto-mode.md`、`.lab/context/auto-status.md` 与 `.lab/context/auto-outcome.md`，先确认 autonomy level、approval status 与 terminal goal schema，再把 eval-protocol 里的指标释义、主表计划、来源约束与结构化实验阶梯当作执行依据，在不修改 mission、framing 和核心 claims 的前提下编排已批准的 `run`、`iterate`、`review`、`report`，轮询长任务完成情况；如果声明了 rung，就保持会话活着并按 rung 转移继续推进。"
 );
 
 ZH_CONTENT[path.join(".codex", "skills", "lab", "SKILL.md")] = `---
@@ -1755,6 +1816,52 @@ ZH_CONTENT[path.join(".lab", "context", "data-decisions.md")] = `# 已批准数
 - 剩余预处理或 leakage 风险：
 `;
 
+ZH_CONTENT[path.join(".lab", "context", "eval-protocol.md")] = `# 评估协议
+
+用这份文件定义 \`/lab:run\`、\`/lab:iterate\`、\`/lab:auto\` 和 \`/lab:report\` 共用的论文导向评估目标、主表计划、gate 与 benchmark ladder。
+
+## 主评估目标
+
+- 主评估目标：
+- 主指标：
+- 次级指标：
+- 必要终局证据：
+
+## 主表计划
+
+- 主表计划：
+- 每张表必须支撑的 claims：
+
+## 指标释义
+
+- 指标释义：
+- 指标来源论文：
+- 指标实现来源：
+- 对比方法来源论文：
+- 对比方法实现来源：
+- 与原始实现的偏差：
+
+## Gate Ladder
+
+- 实验阶梯：
+- benchmark 阶梯：
+- 对比方法 gate：
+- 升格 gate：
+- 最小样本量：
+- 必要输出工件：
+
+### Rung: <rung-id>
+
+- 阶段：
+- 目标：
+- 命令：
+- 监视目标：
+- gate 命令：
+- 通过后：
+- 失败后：
+- 停止后：
+`;
+
 ZH_CONTENT[path.join(".codex", "skills", "lab", "stages", "auto.md")] = `# \`/lab:auto\` 阶段指南
 
 ## 必要输出
@@ -1772,9 +1879,11 @@ ZH_CONTENT[path.join(".codex", "skills", "lab", "stages", "auto.md")] = `# \`/la
 - \`.lab/context/decisions.md\`
 - \`.lab/context/data-decisions.md\`
 - \`.lab/context/evidence-index.md\`
+- \`.lab/context/eval-protocol.md\`
 - \`.lab/context/terminology-lock.md\`
 - \`.lab/context/auto-mode.md\`
 - \`.lab/context/auto-status.md\`
+- \`.lab/context/auto-outcome.md\`
 
 ## 上下文写回
 
@@ -1785,31 +1894,51 @@ ZH_CONTENT[path.join(".codex", "skills", "lab", "stages", "auto.md")] = `# \`/la
 - \`.lab/context/summary.md\`
 - \`.lab/context/session-brief.md\`
 - \`.lab/context/auto-status.md\`
+- \`.lab/context/auto-outcome.md\`
 
 ## 边界规则
 
 - 把 \`/lab:auto\` 当作编排层，不要再发明第二套 workflow。
+- 把 \`.lab/context/eval-protocol.md\` 当作论文导向指标、指标释义、主表、gate 与结构化实验阶梯的唯一来源。
+- 把评估协议当作“带来源的协议”，不是“临场想出来的说明”：指标定义、baseline 行为、对比实现和偏差都必须先写明来源，再用于 gate 或 promotion。
+- 契约里必须声明 \`Autonomy level\` 和 \`Approval status\`，只有显式写成 \`approved\` 才能启动。
+- 契约里还必须声明具体的 terminal goal：\`rounds\`、\`metric-threshold\` 或 \`task-completion\`，并补齐 \`Terminal goal target\` 与 \`Required terminal artifact\`。
+- 级别含义固定为：
+  - \`L1\`：safe run，只允许 \`run\`、\`review\`、\`report\`
+  - \`L2\`：bounded iteration，允许 \`run\`、\`iterate\`、\`review\`、\`report\`
+  - \`L3\`：aggressive campaign，才允许额外编排 \`write\`
 - 默认只编排 \`run\`、\`iterate\`、\`review\`、\`report\`；只有 framing 已批准时才可选 \`write\`。
 - 不要自动修改 mission、paper-facing framing 或核心 claims。
 - 可以在 exploration envelope 内增加数据集、benchmark 和 comparison methods。
 - 只有在 auto-mode 契约中的升格策略满足时，才允许把 exploratory addition 自动升格为 primary package。
 - 长任务必须通过轮询推进，直到完成、超时或命中停止条件。
+- 每次结束都必须写出规范的 \`.lab/context/auto-outcome.md\`。
+- 如果评估协议声明了结构化 rung，就按前台 rung 状态机执行：每个 rung 都要声明阶段、目标、命令、监视目标、gate、通过后/失败后/停止后的转移，并把当前 rung、监视目标和下一 rung 写进 \`.lab/context/auto-status.md\`。
+- 不要只看命令退出码；必须检查阶段产物约束：
+  - \`run\` 和 \`iterate\` 更新 \`results_root\`
+  - \`review\` 更新规范审查上下文
+  - \`report\` 写出 \`<deliverables_root>/report.md\`
+  - \`write\` 写出 \`<deliverables_root>/paper/\` 下的 LaTeX 产物
+- promotion 成功后，必须写回 \`data-decisions.md\`、\`decisions.md\`、\`state.md\` 和 \`session-brief.md\`。
+- 如果某个指标或对比 claim 在评估协议里没有带来源的定义，就不能拿它做 stop 或 promotion 判断。
 
 ## 最小流程
 
 1. 校验自动模式契约
-2. 设置或刷新自动模式状态
-3. 选择下一个允许的 \`/lab\` 子阶段
-4. 发起有边界动作
-5. 轮询进程、checkpoint 或 summary 的变化
-6. 记录结果并决定 continue、promote、stop 或 escalate
+2. 确认已批准的 autonomy level 与允许阶段一致
+3. 设置或刷新自动模式状态
+4. 选择下一个允许的 \`/lab\` 子阶段
+5. 发起有边界动作
+6. 轮询进程、checkpoint 或 summary 的变化
+7. 评估声明过的 terminal goal 是否已经达成
+8. 记录结果并决定 continue、promote、stop 或 escalate
 
 ## 交互约束
 
 - 开始前先简洁说明：objective、frozen core 和下一自动阶段。
 - 如果契约本身不完整，一次只追问一个问题。
 - 如果存在多个可信的下一动作，先给 2-3 个 bounded 方案和推荐项，再启动长任务。
-- 只有当下一步会离开 frozen core 时，才保留人工 approval gate。
+- 只有当下一步会离开已批准的 exploration envelope、超出选定 autonomy level，或实质改变 frozen core 时，才保留人工 approval gate。
 `;
 
 ZH_CONTENT[path.join(".claude", "skills", "lab", "stages", "auto.md")] =

package/lib/install.cjs

@@ -40,8 +40,10 @@ const PROJECT_OWNED_LOCALIZED_PATHS = [
   path.join(".lab", "context", "evidence-index.md"),
   path.join(".lab", "context", "open-questions.md"),
   path.join(".lab", "context", "data-decisions.md"),
+  path.join(".lab", "context", "eval-protocol.md"),
   path.join(".lab", "context", "auto-mode.md"),
   path.join(".lab", "context", "auto-status.md"),
+  path.join(".lab", "context", "auto-outcome.md"),
   path.join(".lab", "context", "terminology-lock.md"),
   path.join(".lab", "context", "summary.md"),
   path.join(".lab", "context", "next-action.md"),
@@ -431,13 +433,31 @@ function registerProjectInstall(targetDir, metadata, { env = process.env } = {})
 }
 
 function isTemporaryTestPath(targetDir) {
-  const resolvedTarget = path.resolve(targetDir);
-  const tmpRoot = path.resolve(os.tmpdir());
-  const relativeToTmp = path.relative(tmpRoot, resolvedTarget);
-  if (relativeToTmp.startsWith("..") || path.isAbsolute(relativeToTmp)) {
-    return false;
+  const normalizedCandidates = new Set([path.resolve(targetDir)]);
+  try {
+    normalizedCandidates.add(fs.realpathSync(targetDir));
+  } catch {}
+
+  const tempRoots = new Set([path.resolve(os.tmpdir()), path.resolve("/tmp"), path.resolve("/private/tmp")]);
+  for (const root of Array.from(tempRoots)) {
+    try {
+      tempRoots.add(fs.realpathSync(root));
+    } catch {}
+  }
+
+  for (const candidate of normalizedCandidates) {
+    if (!path.basename(candidate).startsWith("superlab-")) {
+      continue;
+    }
+    for (const tempRoot of tempRoots) {
+      const relativeToTmp = path.relative(tempRoot, candidate);
+      if (!relativeToTmp.startsWith("..") && !path.isAbsolute(relativeToTmp)) {
+        return true;
+      }
+    }
   }
-  return path.basename(resolvedTarget).startsWith("superlab-");
+
+  return false;
 }
 
 function detectLanguage({ explicitLang, env = process.env } = {}) {

package/package-assets/claude/commands/lab/auto.md

@@ -8,4 +8,4 @@ tags: [workflow, research, auto]
 Use the installed `lab` skill at `.claude/skills/lab/SKILL.md`.
 
 Execute the requested `/lab:auto` stage against the user's argument now. Do not only recommend another lab stage. If a blocking prerequisite is missing, say exactly what is missing and ask at most one clarifying question.
-This command runs the `/lab:auto` stage. It must read `.lab/context/auto-mode.md` and `.lab/context/auto-status.md`, orchestrate approved run, iterate, review, and report stages inside that contract, poll long-running work until completion or stop conditions, and write progress back into `.lab/context/auto-status.md`.
+This command runs the `/lab:auto` stage. It must read `.lab/context/eval-protocol.md`, `.lab/context/auto-mode.md`, `.lab/context/auto-status.md`, and `.lab/context/auto-outcome.md`, enforce the declared terminal goal schema, orchestrate approved run, iterate, review, and report stages inside that contract, poll long-running work until completion or stop conditions, and write progress plus the final outcome back into `.lab/context/auto-status.md` and `.lab/context/auto-outcome.md`.

package/package-assets/claude/commands/lab.md

@@ -18,7 +18,7 @@ tags: [workflow, research, overview]
   Turn the approved idea into an approved dataset and benchmark package with dataset years, papers that used each dataset, source audit, download plan, classic-public versus recent-strong-public versus claim-specific benchmark roles, and explicit rationale for canonical baselines, strong historical baselines, recent strong public methods, and closest prior work.
 
 - `/lab:auto`
-  Run a bounded orchestration loop over approved execution stages. Use an auto-mode contract plus live auto-status to drive `run`, `iterate`, `review`, `report`, and optionally `write` without changing the frozen mission or framing.
+  Run a bounded orchestration loop over approved execution stages. Use an auto-mode contract plus live auto-status to drive `run`, `iterate`, `review`, `report`, and optionally `write` without changing the frozen mission or framing. Choose an autonomy level, declare a concrete terminal goal, explicitly approve the contract before starting, and treat `.lab/context/eval-protocol.md` as the source of truth for metrics, metric glossary, source-backed comparison semantics, tables, and structured experiment-ladder rungs.
 
 - `/lab:framing`
   Lock paper-facing method name, module names, paper title, and contribution bullets by auditing current-field and adjacent-field terminology, then keep an approval gate before any section drafting.
@@ -51,5 +51,6 @@ tags: [workflow, research, overview]
 - `/lab:spec` should inherit the approved dataset package from `.lab/context/data-decisions.md`.
 - Never skip directly from `/lab:idea` to code.
 - `/lab:iterate` requires a normalized summary from `scripts/eval_report.py`.
+- `/lab:run`, `/lab:iterate`, `/lab:auto`, and `/lab:report` should all follow `.lab/context/eval-protocol.md`, including its recorded sources for metrics and comparison implementations.
 - `/lab:write` requires an approved framing artifact from `/lab:framing`.
 - `/lab:write` requires stable report artifacts, a mini-outline, the active section guide, `paper-review.md`, and `does-my-writing-flow-source.md`, and should only change one section per round.

package/package-assets/codex/prompts/lab-auto.md

@@ -6,4 +6,4 @@ argument-hint: autonomous campaign target
 Use the installed `lab` skill at `.codex/skills/lab/SKILL.md`.
 
 Execute the requested `/lab:auto` stage against the user's argument now. Do not only recommend another lab stage. If a blocking prerequisite is missing, say exactly what is missing and ask at most one clarifying question.
-This command runs the `/lab:auto` stage. It must read `.lab/context/auto-mode.md` and `.lab/context/auto-status.md`, orchestrate approved run, iterate, review, and report stages inside that contract, poll long-running work until completion or stop conditions, and write progress back into `.lab/context/auto-status.md`.
+This command runs the `/lab:auto` stage. It must read `.lab/context/eval-protocol.md`, `.lab/context/auto-mode.md`, `.lab/context/auto-status.md`, and `.lab/context/auto-outcome.md`, enforce the declared terminal goal schema, orchestrate approved run, iterate, review, and report stages inside that contract, poll long-running work until completion or stop conditions, and write progress plus the final outcome back into `.lab/context/auto-status.md` and `.lab/context/auto-outcome.md`.

package/package-assets/codex/prompts/lab.md

@@ -16,7 +16,7 @@ argument-hint: workflow question or stage choice
   Turn the approved idea into an approved dataset and benchmark package with dataset years, papers that used each dataset, source audit, download plan, classic-public versus recent-strong-public versus claim-specific benchmark roles, and explicit rationale for canonical baselines, strong historical baselines, recent strong public methods, and closest prior work.
 
 - `/lab:auto`
-  Run a bounded orchestration loop over approved execution stages. Use an auto-mode contract plus live auto-status to drive `run`, `iterate`, `review`, `report`, and optionally `write` without changing the frozen mission or framing.
+  Run a bounded orchestration loop over approved execution stages. Use an auto-mode contract plus live auto-status to drive `run`, `iterate`, `review`, `report`, and optionally `write` without changing the frozen mission or framing. Choose an autonomy level, declare a concrete terminal goal, explicitly approve the contract before starting, and treat `.lab/context/eval-protocol.md` as the source of truth for metrics, metric glossary, source-backed comparison semantics, tables, and structured experiment-ladder rungs.
 
 - `/lab:framing`
   Lock paper-facing method name, module names, paper title, and contribution bullets by auditing current-field and adjacent-field terminology, then keep an approval gate before any section drafting.
@@ -49,5 +49,6 @@ argument-hint: workflow question or stage choice
 - `/lab:spec` should inherit the approved dataset package from `.lab/context/data-decisions.md`.
 - Never skip directly from `/lab:idea` to code.
 - `/lab:iterate` requires a normalized summary from `scripts/eval_report.py`.
+- `/lab:run`, `/lab:iterate`, `/lab:auto`, and `/lab:report` should all follow `.lab/context/eval-protocol.md`, including its recorded sources for metrics and comparison implementations.
 - `/lab:write` requires an approved framing artifact from `/lab:framing`.
 - `/lab:write` requires stable report artifacts, a mini-outline, the active section guide, `paper-review.md`, and `does-my-writing-flow-source.md`, and should only change one section per round.

package/package-assets/shared/lab/context/auto-mode.md

@@ -1,12 +1,19 @@
 # Auto Mode Contract
 
 Use this file to define the bounded autonomous execution envelope for `/lab:auto`.
+Pair it with `.lab/context/eval-protocol.md`, which defines the paper-facing metrics, tables, gates, and benchmark ladder that auto mode should optimize against.
+If `eval-protocol.md` declares structured rung entries, auto mode follows those rung transitions first and uses the stage commands here as per-stage fallbacks.
 
 ## Objective
 
 - Objective:
+- Autonomy level: L2
+- Approval status: draft
 - Allowed stages: run, iterate, review, report
 - Success criteria:
+- Terminal goal type:
+- Terminal goal target:
+- Required terminal artifact:
 
 ## Loop Budget
 
@@ -27,6 +34,14 @@ Use this file to define the bounded autonomous execution envelope for `/lab:auto
 - Promotion check command:
 - Promotion command:
 
+## Stage Output Contracts
+
+- Run stage contract: write persistent outputs under `results_root`.
+- Iterate stage contract: update persistent outputs under `results_root`.
+- Review stage contract: update canonical review context such as `.lab/context/decisions.md`, `state.md`, `open-questions.md`, or `evidence-index.md`.
+- Report stage contract: write the final report to `<deliverables_root>/report.md`.
+- Write stage contract: write LaTeX output under `<deliverables_root>/paper/`.
+
 ## Promotion Policy
 
 - Promotion policy:
@@ -40,3 +55,4 @@ Use this file to define the bounded autonomous execution envelope for `/lab:auto
 
 - Stop conditions:
 - Escalation conditions:
+- Canonical promotion writeback: update `.lab/context/data-decisions.md`, `.lab/context/decisions.md`, `.lab/context/state.md`, and `.lab/context/session-brief.md`.

package/package-assets/shared/lab/context/auto-outcome.md

@@ -0,0 +1,28 @@
+# Auto Outcome
+
+## Goal
+
+- Objective:
+- Experiment ladder:
+- Metric glossary:
+- Metric source papers:
+- Metric implementation source:
+- Comparison source papers:
+- Comparison implementation source:
+- Deviation from original implementation:
+- Terminal goal type:
+- Terminal goal target:
+- Required terminal artifact:
+
+## Outcome
+
+- Status: idle
+- Goal reached: no
+- Stop reason:
+- Promotion applied: no
+- Final artifact:
+- Final rung:
+- Executed stages:
+- Iterations completed: 0
+- Started at:
+- Finished at:

package/package-assets/shared/lab/context/auto-status.md

@@ -7,6 +7,9 @@
 - Current command:
 - Active run id:
 - Iteration count: 0
+- Current rung:
+- Watch target:
+- Next rung:
 
 ## Timing
 

package/package-assets/shared/lab/context/eval-protocol.md

@@ -0,0 +1,46 @@
+# Evaluation Protocol
+
+Use this file to define the paper-facing evaluation objective, table plan, gates, and benchmark ladder for `/lab:run`, `/lab:iterate`, `/lab:auto`, and `/lab:report`.
+
+## Primary Evaluation Objective
+
+- Primary evaluation objective:
+- Primary metrics:
+- Secondary metrics:
+- Required terminal evidence:
+
+## Table Plan
+
+- Table plan:
+- Required claims per table:
+
+## Metric Glossary
+
+- Metric glossary:
+- Metric source papers:
+- Metric implementation source:
+- Comparison source papers:
+- Comparison implementation source:
+- Deviation from original implementation:
+
+Record enough source detail here that later `run`, `iterate`, `auto`, and `report` stages do not have to guess what a metric means, which baseline implementation is canonical, or where a comparison method came from.
+
+## Gate Ladder
+
+- Experiment ladder:
+- Benchmark ladder:
+- Comparison gate:
+- Promotion gate:
+- Minimum sample sizes:
+- Required output artifacts:
+
+### Rung: <rung-id>
+
+- Stage:
+- Goal:
+- Command:
+- Watch:
+- Gate:
+- On pass:
+- On fail:
+- On stop:

package/package-assets/shared/skills/lab/SKILL.md

@@ -22,6 +22,8 @@ Use this skill when the user invokes `/lab:*` or asks for the structured researc
 - Write durable artifacts to disk instead of leaving key decisions only in chat.
 - Use `.lab/config/workflow.json` as the global contract for workflow language, paper language, and paper format.
 - Use `.lab/context/` as the shared project state for both Codex and Claude entrypoints.
+- Use `.lab/context/eval-protocol.md` as the shared evaluation contract for run, iterate, auto, and report stages, including metric glossary and experiment ladder semantics.
+- Treat evaluation semantics as source-backed once evaluation planning starts: metrics, benchmark gates, baseline behavior, comparison implementations, and deviations should come from recorded sources, not memory.
 - Workflow artifacts should follow the installed workflow language.
 - Final paper output should default to LaTeX, and its manuscript language should be decided separately from the workflow language.
 - Separate sourced facts from model-generated hypotheses.
@@ -82,6 +84,8 @@ Use this skill when the user invokes `/lab:*` or asks for the structured researc
 - Use this stage to orchestrate approved execution stages with bounded autonomy.
 - Read `.lab/config/workflow.json`, `.lab/context/mission.md`, `.lab/context/state.md`, `.lab/context/decisions.md`, `.lab/context/data-decisions.md`, `.lab/context/evidence-index.md`, `.lab/context/terminology-lock.md`, `.lab/context/auto-mode.md`, and `.lab/context/auto-status.md` before acting.
 - Treat `.lab/context/auto-mode.md` as the control contract and `.lab/context/auto-status.md` as the live state file.
+- Require `Autonomy level` and `Approval status` in `.lab/context/auto-mode.md` before execution.
+- Treat `L1` as safe-run validation, `L2` as bounded iteration, and `L3` as aggressive campaign mode.
 - Reuse `/lab:run`, `/lab:iterate`, `/lab:review`, `/lab:report`, and optional `/lab:write` instead of inventing a second workflow.
 - Do not automatically change the research mission, paper-facing framing, or core claims.
 - You may add exploratory datasets, benchmarks, and comparison methods inside the approved exploration envelope.
@@ -106,6 +110,7 @@ Use this skill when the user invokes `/lab:*` or asks for the structured researc
 - Register the run with `.lab/.managed/scripts/register_run.py`.
 - Normalize the result with `.lab/.managed/scripts/eval_report.py`.
 - Validate normalized output with `.lab/.managed/scripts/validate_results.py`.
+- Read `.lab/context/eval-protocol.md` before choosing the smallest run so the first experiment already targets the approved tables, metrics, and gates.
 - Update `.lab/context/state.md` and `.lab/context/evidence-index.md` after the run.
 
 ### `/lab:iterate`
@@ -122,6 +127,8 @@ Use this skill when the user invokes `/lab:*` or asks for the structured researc
 - Require a normalized evaluation report each round.
 - Read `.lab/context/mission.md`, `.lab/context/state.md`, `.lab/context/decisions.md`, and `.lab/context/evidence-index.md` at the start of each round.
 - Read `.lab/context/data-decisions.md` before changing benchmark-facing experiments.
+- Read `.lab/context/eval-protocol.md` before changing evaluation ladders, sample sizes, or promotion gates.
+- Keep metric definitions, baseline behavior, and comparison implementations anchored to the source-backed evaluation protocol before changing thresholds, gates, or ladder transitions.
 - Switch to diagnostic mode if risk increases for two consecutive rounds.
 - Write round reports with `.lab/.managed/templates/iteration-report.md`.
 - Update `.lab/context/state.md`, `.lab/context/decisions.md`, `.lab/context/evidence-index.md`, and `.lab/context/open-questions.md` each round as needed.
@@ -141,6 +148,8 @@ Use this skill when the user invokes `/lab:*` or asks for the structured researc
 
 - Summarize all validated iteration summaries.
 - Read `.lab/context/mission.md`, `.lab/context/state.md`, `.lab/context/decisions.md`, `.lab/context/evidence-index.md`, and `.lab/context/data-decisions.md` before drafting.
+- Read `.lab/context/eval-protocol.md` before choosing tables, thresholds, or final result framing.
+- Keep metric definitions, comparison semantics, and implementation references anchored to the approved evaluation protocol instead of re-deriving them during reporting.
 - Aggregate them with `.lab/.managed/scripts/summarize_iterations.py`.
 - Write the final document with `.lab/.managed/templates/final-report.md`.
 - Keep failed attempts and limitations visible.
@@ -172,7 +181,9 @@ Use this skill when the user invokes `/lab:*` or asks for the structured researc
 - No implementation before `/lab:spec`.
 - No frozen spec without an approved dataset package or an explicit defer reason recorded in `.lab/context/data-decisions.md`.
 - No unconstrained iteration. Every `/lab:iterate` campaign must declare done criteria and `max_iterations`.
+- No execution or reporting campaign without an evaluation protocol or an explicit defer reason recorded in `.lab/context/eval-protocol.md`.
 - No unconstrained auto mode. Every `/lab:auto` campaign must declare allowed stages, stop conditions, and a promotion policy in `.lab/context/auto-mode.md`.
+- No auto start without an explicit autonomy level and `Approval status: approved`.
 - No final report without validated normalized results.
 - No paper-writing round without stable report artifacts, an approved framing artifact, evidence links, and LaTeX manuscript output.
 
@@ -194,6 +205,6 @@ Use this skill when the user invokes `/lab:*` or asks for the structured researc
 - Vendored paper-writing references: `.codex/skills/lab/references/paper-writing/{abstract,introduction,related-work,method,experiments,conclusion,paper-review,does-my-writing-flow-source}.md` or `.claude/skills/lab/references/paper-writing/{abstract,introduction,related-work,method,experiments,conclusion,paper-review,does-my-writing-flow-source}.md`
 - Command adapters: the installed `/lab:*` command assets
 - Shared workflow config: `.lab/config/workflow.json`
-- Shared project context: `.lab/context/{mission,state,decisions,evidence-index,open-questions,data-decisions,auto-mode,auto-status}.md`
+- Shared project context: `.lab/context/{mission,state,decisions,evidence-index,open-questions,data-decisions,eval-protocol,auto-mode,auto-status}.md`
 - Templates: `.lab/.managed/templates/`
 - Scripts: `.lab/.managed/scripts/`