superlab 0.1.22 → 0.1.24

package/README.md

@@ -176,6 +176,7 @@ superlab auto stop
 - `L1` is a safe run envelope for `run`, `review`, and `report`
 - `L2` is the default bounded iteration envelope for `run`, `iterate`, `review`, and `report`
 - `L3` is an aggressive campaign envelope that may also include `write`
+- If you are unsure, choose `L2`
 
 - `run` and `iterate` must change persistent outputs under `results_root`
 - `review` must update canonical review context
@@ -189,10 +190,18 @@ It does not replace manual `idea`, `data`, `framing`, or `spec` decisions.
 
 Good `/lab:auto` input is explicit. Treat `Autonomy level L1/L2/L3` as execution privilege, and treat `paper layer`, `phase`, or `table` as experiment targets. If the workflow language is Chinese, summaries, checklist items, task labels, and progress updates should also stay in Chinese unless a literal identifier must remain unchanged.
 
+Level Guide for `/lab:auto`:
+
+- `L1`: use this when you want safe validation, one bounded real run, or a simple report refresh
+- `L2`: use this for normal bounded experiment iteration inside a frozen core
+- `L3`: use this only when you want a broader campaign with a larger search space and optional writing
+- If the request omits the level or mixes it with a paper layer, phase, or table target, stop and ask for an explicit level before starting
+- If you are unsure, choose `L2`
+
 Example:
 
 ```text
-/lab:auto Autonomy level L2. Objective: advance paper layer 3 organizer enforcement. Terminal goal: task-completion. Scope: bounded protocol, tests, minimal implementation, and one small run. Allowed modifications: evaluator prompt registry, ingestion, and parser only.
+/lab:auto Autonomy level L2. Objective: advance paper layer 3 through one bounded protocol improvement. Terminal goal: task-completion. Scope: bounded protocol, tests, one minimal implementation, and one small run. Allowed modifications: configuration, evaluation script, and data-loading logic only.
 ```
 
 ## Version

package/README.zh-CN.md

@@ -174,6 +174,7 @@ superlab auto stop
 - `L1` 是安全运行级别，只允许 `run`、`review`、`report`
 - `L2` 是默认推荐级别，允许 `run`、`iterate`、`review`、`report`
 - `L3` 是激进 campaign 级别，才允许额外编排 `write`
+- 如果不确定，默认推荐 `L2`
 
 - `run` 和 `iterate` 必须更新 `results_root` 下的持久输出
 - `review` 必须更新规范的审查上下文
@@ -187,10 +188,18 @@ superlab auto stop
 
 好的 `/lab:auto` 输入应该显式写清。把 `Autonomy level L1/L2/L3` 当成执行权限级别，把 `paper layer`、`phase`、`table` 当成实验目标，不要混用。如果 workflow language 是中文，摘要、清单条目、任务标签和进度更新也应保持中文，除非某个字面标识符必须保持原样。
 
+`/lab:auto` 层级指南：
+
+- `L1`：适合安全验证、一轮有边界真实运行，或简单的 report 刷新
+- `L2`：适合冻结核心边界内的常规实验迭代，也是默认推荐级别
+- `L3`：只在你明确想做更大范围 campaign、允许更广探索和可选写作时使用
+- 如果用户输入没写级别，或者把级别和 `paper layer`、`phase`、`table` 混用了，就应先停下来，给出更详细的层级说明，再要求用户明确选 `L1/L2/L3`
+- 如果不确定，默认推荐 `L2`
+
 示例：
 
 ```text
-/lab:auto 自治级别 L2。目标：推进 paper layer 3 的 organizer enforcement。终止条件：完成 bounded protocol、测试、最小实现和一轮小规模结果。允许修改：evaluator prompt registry、ingestion、parser。
+/lab:auto 自治级别 L2。目标：推进 paper layer 3 的一项有边界协议改进。终止条件：完成 bounded protocol、测试、一项最小实现和一轮小规模结果。允许修改：配置、评估脚本、数据加载逻辑。
 ```
 
 ## 版本查询

package/lib/auto_runner.cjs

@@ -278,6 +278,21 @@ async function startAutoMode({ targetDir, now = new Date() }) {
     comparisonSourcePapers: evalProtocol.comparisonSourcePapers,
     comparisonImplementationSource: evalProtocol.comparisonImplementationSource,
     deviationFromOriginalImplementation: evalProtocol.deviationFromOriginalImplementation,
+    evaluationSettingSemantics: evalProtocol.evaluationSettingSemantics,
+    visibilityAndLeakageRisks: evalProtocol.visibilityAndLeakageRisks,
+    anchorAndLabelPolicy: evalProtocol.anchorAndLabelPolicy,
+    scaleAndComparabilityPolicy: evalProtocol.scaleAndComparabilityPolicy,
+    metricValidityChecks: evalProtocol.metricValidityChecks,
+    comparisonValidityChecks: evalProtocol.comparisonValidityChecks,
+    statisticalValidityChecks: evalProtocol.statisticalValidityChecks,
+    claimBoundary: evalProtocol.claimBoundary,
+    integritySelfCheck: evalProtocol.integritySelfCheck,
+    anomalySignals: evalProtocol.anomalySignals,
+    implementationRealityChecks: evalProtocol.implementationRealityChecks,
+    alternativeExplanationsConsidered: evalProtocol.alternativeExplanationsConsidered,
+    crossCheckMethod: evalProtocol.crossCheckMethod,
+    bestSupportedInterpretation: evalProtocol.bestSupportedInterpretation,
+    escalationThreshold: evalProtocol.escalationThreshold,
   };
 
   const writeRunningStatus = (overrides = {}) => {
@@ -768,6 +783,21 @@ function stopAutoMode({ targetDir, now = new Date() }) {
     comparisonSourcePapers: evalProtocol.comparisonSourcePapers,
     comparisonImplementationSource: evalProtocol.comparisonImplementationSource,
     deviationFromOriginalImplementation: evalProtocol.deviationFromOriginalImplementation,
+    evaluationSettingSemantics: evalProtocol.evaluationSettingSemantics,
+    visibilityAndLeakageRisks: evalProtocol.visibilityAndLeakageRisks,
+    anchorAndLabelPolicy: evalProtocol.anchorAndLabelPolicy,
+    scaleAndComparabilityPolicy: evalProtocol.scaleAndComparabilityPolicy,
+    metricValidityChecks: evalProtocol.metricValidityChecks,
+    comparisonValidityChecks: evalProtocol.comparisonValidityChecks,
+    statisticalValidityChecks: evalProtocol.statisticalValidityChecks,
+    claimBoundary: evalProtocol.claimBoundary,
+    integritySelfCheck: evalProtocol.integritySelfCheck,
+    anomalySignals: evalProtocol.anomalySignals,
+    implementationRealityChecks: evalProtocol.implementationRealityChecks,
+    alternativeExplanationsConsidered: evalProtocol.alternativeExplanationsConsidered,
+    crossCheckMethod: evalProtocol.crossCheckMethod,
+    bestSupportedInterpretation: evalProtocol.bestSupportedInterpretation,
+    escalationThreshold: evalProtocol.escalationThreshold,
   };
   const status = {
     ...existing,

package/lib/auto_state.cjs

@@ -154,6 +154,21 @@ function renderAutoOutcome(outcome, { lang = "en" } = {}) {
 - 对比方法来源论文: ${outcome.comparisonSourcePapers || ""}
 - 对比方法实现来源: ${outcome.comparisonImplementationSource || ""}
 - 与原始实现的偏差: ${outcome.deviationFromOriginalImplementation || ""}
+- 评测设定语义: ${outcome.evaluationSettingSemantics || ""}
+- 可见性与泄漏风险: ${outcome.visibilityAndLeakageRisks || ""}
+- 锚点与标签策略: ${outcome.anchorAndLabelPolicy || ""}
+- 尺度与可比性策略: ${outcome.scaleAndComparabilityPolicy || ""}
+- 指标有效性检查: ${outcome.metricValidityChecks || ""}
+- 对比有效性检查: ${outcome.comparisonValidityChecks || ""}
+- 统计有效性检查: ${outcome.statisticalValidityChecks || ""}
+- 结论边界: ${outcome.claimBoundary || ""}
+- 完整性自检: ${outcome.integritySelfCheck || ""}
+- 异常信号: ${outcome.anomalySignals || ""}
+- 实现层现实检查: ${outcome.implementationRealityChecks || ""}
+- 已考虑的替代解释: ${outcome.alternativeExplanationsConsidered || ""}
+- 交叉验证方法: ${outcome.crossCheckMethod || ""}
+- 当前最站得住的解释: ${outcome.bestSupportedInterpretation || ""}
+- 升级阈值: ${outcome.escalationThreshold || ""}
 - 终止目标类型: ${outcome.terminalGoalType || ""}
 - 终止目标目标值: ${outcome.terminalGoalTarget || ""}
 - 必要终止工件: ${outcome.requiredTerminalArtifact || ""}
@@ -191,6 +206,21 @@ function renderAutoOutcome(outcome, { lang = "en" } = {}) {
 - Comparison source papers: ${outcome.comparisonSourcePapers || ""}
 - Comparison implementation source: ${outcome.comparisonImplementationSource || ""}
 - Deviation from original implementation: ${outcome.deviationFromOriginalImplementation || ""}
+- Evaluation setting semantics: ${outcome.evaluationSettingSemantics || ""}
+- Visibility and leakage risks: ${outcome.visibilityAndLeakageRisks || ""}
+- Anchor and label policy: ${outcome.anchorAndLabelPolicy || ""}
+- Scale and comparability policy: ${outcome.scaleAndComparabilityPolicy || ""}
+- Metric validity checks: ${outcome.metricValidityChecks || ""}
+- Comparison validity checks: ${outcome.comparisonValidityChecks || ""}
+- Statistical validity checks: ${outcome.statisticalValidityChecks || ""}
+- Claim boundary: ${outcome.claimBoundary || ""}
+- Integrity self-check: ${outcome.integritySelfCheck || ""}
+- Anomaly signals: ${outcome.anomalySignals || ""}
+- Implementation reality checks: ${outcome.implementationRealityChecks || ""}
+- Alternative explanations considered: ${outcome.alternativeExplanationsConsidered || ""}
+- Cross-check method: ${outcome.crossCheckMethod || ""}
+- Best-supported interpretation: ${outcome.bestSupportedInterpretation || ""}
+- Escalation threshold: ${outcome.escalationThreshold || ""}
 - Terminal goal type: ${outcome.terminalGoalType || ""}
 - Terminal goal target: ${outcome.terminalGoalTarget || ""}
 - Required terminal artifact: ${outcome.requiredTerminalArtifact || ""}

package/lib/context.cjs

@@ -27,6 +27,24 @@ const EVAL_COLLABORATOR_FIELDS = [
     labels: ["Method and baseline implementation source", "方法与基线实现来源"],
   },
   { name: "Metric source papers", labels: ["Metric source papers", "指标来源论文"] },
+  { name: "Evaluation setting semantics", labels: ["Evaluation setting semantics", "评测设定语义"] },
+  { name: "Visibility and leakage risks", labels: ["Visibility and leakage risks", "可见性与泄漏风险"] },
+  { name: "Anchor and label policy", labels: ["Anchor and label policy", "锚点与标签策略"] },
+  { name: "Scale and comparability policy", labels: ["Scale and comparability policy", "尺度与可比性策略"] },
+  { name: "Metric validity checks", labels: ["Metric validity checks", "指标有效性检查"] },
+  { name: "Comparison validity checks", labels: ["Comparison validity checks", "对比方法有效性检查"] },
+  { name: "Statistical validity checks", labels: ["Statistical validity checks", "统计有效性检查"] },
+  { name: "Claim boundary", labels: ["Claim boundary", "结论边界"] },
+  { name: "Integrity self-check", labels: ["Integrity self-check", "完整性自检"] },
+  { name: "Anomaly signals", labels: ["Anomaly signals", "异常信号"] },
+  { name: "Implementation reality checks", labels: ["Implementation reality checks", "实现层现实检查"] },
+  {
+    name: "Alternative explanations considered",
+    labels: ["Alternative explanations considered", "已考虑的替代解释"],
+  },
+  { name: "Cross-check method", labels: ["Cross-check method", "交叉验证方法"] },
+  { name: "Best-supported interpretation", labels: ["Best-supported interpretation", "当前最站得住的解释"] },
+  { name: "Escalation threshold", labels: ["Escalation threshold", "升级阈值"] },
   { name: "Required output artifacts", labels: ["Required output artifacts", "必要输出工件"] },
 ];
 const REPORT_REQUIRED_SECTIONS = [
@@ -44,7 +62,34 @@ const REPORT_REQUIRED_SECTIONS = [
     patterns: [/^##\s+Method and Baseline Sources\s*$/m, /^##\s+方法与基线来源\s*$/m],
   },
   { name: "Metric Sources", patterns: [/^##\s+Metric Sources\s*$/m, /^##\s+指标来源\s*$/m] },
+  {
+    name: "Sanity and Alternative Explanations",
+    patterns: [/^##\s+Sanity and Alternative Explanations\s*$/m, /^##\s+异常与替代解释\s*$/m],
+  },
+];
+const REPORT_SOURCE_SECTION_NAMES = new Set([
+  "Background Sources",
+  "Method and Baseline Sources",
+  "Metric Sources",
+]);
+const REPORT_SOURCE_PATH_PATTERNS = [
+  /\/Users\//,
+  /\/home\//,
+  /\/tmp\//,
+  /\/private\/tmp\//,
+  /\.lab\//,
+  /outputs\//,
+  /docs\/research\//,
+];
+const REPORT_SOURCE_CITATION_MARKERS = ["Citation:", "引用："];
+const REPORT_SOURCE_ROLE_MARKERS = [
+  "What it established:",
+  "What it does:",
+  "What it measures:",
+  "做了什么：",
+  "衡量什么：",
 ];
+const REPORT_SOURCE_LIMITATION_MARKERS = ["Limitation", "局限"];
 const MAIN_TABLES_REQUIRED_SECTIONS = [
   { name: "Reader Summary", patterns: [/^##\s+Reader Summary\s*$/m, /^##\s+给用户看的总结\s*$/m] },
   { name: "Selected Metrics", patterns: [/^##\s+Selected Metrics\s*$/m, /^##\s+选定指标\s*$/m] },
@@ -82,6 +127,13 @@ const REPORT_FIELDS = {
   metricSourcePapers: ["Metric source papers", "指标来源论文"],
   metricImplementationSource: ["Metric implementation source", "指标实现来源"],
   metricDeviation: ["Deviation from original implementation", "与原始实现的偏差"],
+  claimBoundary: ["Claim boundary", "结论边界"],
+  anomalySignals: ["Anomaly signals observed", "观察到的异常信号"],
+  implementationChecks: ["Implementation checks performed", "做过的实现层检查"],
+  alternativeExplanations: ["Alternative explanations ruled out", "已排除的更简单解释"],
+  crossChecks: ["Cross-checks that strengthen the current interpretation", "支撑当前解释的交叉验证"],
+  bestSupportedInterpretation: ["Best-supported interpretation", "当前最站得住的解释"],
+  escalationThreshold: ["Escalation threshold if future anomalies appear", "未来异常出现时的升级阈值"],
   datasets: ["Datasets", "数据集"],
   baselines: ["Baselines", "基线"],
   metrics: ["Metrics", "指标"],
@@ -229,6 +281,48 @@ function missingRequiredSections(text, sections) {
     .map((section) => section.name);
 }
 
+function extractSectionBody(text, section) {
+  if (!text) {
+    return "";
+  }
+  for (const pattern of section.patterns) {
+    const match = pattern.exec(text);
+    if (!match) {
+      continue;
+    }
+    const start = match.index + match[0].length;
+    const nextHeading = text.slice(start).search(/^##\s+/m);
+    const end = nextHeading >= 0 ? start + nextHeading : text.length;
+    return text.slice(start, end).trim();
+  }
+  return "";
+}
+
+function sourceSectionIssues(reportText) {
+  const issues = [];
+  for (const section of REPORT_REQUIRED_SECTIONS) {
+    if (!REPORT_SOURCE_SECTION_NAMES.has(section.name)) {
+      continue;
+    }
+    const body = extractSectionBody(reportText, section);
+    if (!body) {
+      continue;
+    }
+    if (REPORT_SOURCE_PATH_PATTERNS.some((pattern) => pattern.test(body))) {
+      issues.push(`report.md section '${section.name}' must not rely on local file paths or internal provenance`);
+    }
+    if (!REPORT_SOURCE_CITATION_MARKERS.some((marker) => body.includes(marker))) {
+      issues.push(`report.md section '${section.name}' must include at least one citation anchor`);
+    }
+    const hasRole = REPORT_SOURCE_ROLE_MARKERS.some((marker) => body.includes(marker));
+    const hasLimitation = REPORT_SOURCE_LIMITATION_MARKERS.some((marker) => body.includes(marker));
+    if (!hasRole || !hasLimitation) {
+      issues.push(`report.md section '${section.name}' must explain what the anchor does and one limitation`);
+    }
+  }
+  return issues;
+}
+
 function collaboratorReportIssues(targetDir) {
   if (!hasCollaboratorFacingDeliverables(targetDir)) {
     return [];
@@ -236,10 +330,12 @@ function collaboratorReportIssues(targetDir) {
   const { reportPath, mainTablesPath } = getCollaboratorDeliverablePaths(targetDir);
   const issues = [];
   if (fs.existsSync(reportPath)) {
-    const missing = missingRequiredSections(readFileIfExists(reportPath), REPORT_REQUIRED_SECTIONS);
+    const reportText = readFileIfExists(reportPath);
+    const missing = missingRequiredSections(reportText, REPORT_REQUIRED_SECTIONS);
     if (missing.length > 0) {
       issues.push(`report.md is missing required collaborator-facing sections: ${missing.join(", ")}`);
     }
+    issues.push(...sourceSectionIssues(reportText));
   }
   if (fs.existsSync(mainTablesPath)) {
     const missing = missingRequiredSections(readFileIfExists(mainTablesPath), MAIN_TABLES_REQUIRED_SECTIONS);
@@ -397,6 +493,27 @@ function buildEvalProtocolText(lang, fields, rungs) {
 - 对比方法实现来源: ${fields.comparisonImplementationSource || "待补充"}
 - 与原始实现的偏差: ${fields.deviationFromOriginalImplementation || "待补充"}
 
+## 学术有效性检查
+
+- 评测设定语义: ${fields.evaluationSettingSemantics || "待补充"}
+- 可见性与泄漏风险: ${fields.visibilityAndLeakageRisks || "待补充"}
+- 锚点与标签策略: ${fields.anchorAndLabelPolicy || "待补充"}
+- 尺度与可比性策略: ${fields.scaleAndComparabilityPolicy || "待补充"}
+- 指标有效性检查: ${fields.metricValidityChecks || "待补充"}
+- 对比方法有效性检查: ${fields.comparisonValidityChecks || "待补充"}
+- 统计有效性检查: ${fields.statisticalValidityChecks || "待补充"}
+- 结论边界: ${fields.claimBoundary || "待补充"}
+- 完整性自检: ${fields.integritySelfCheck || "待补充"}
+
+## 异常与替代解释检查
+
+- 异常信号: ${fields.anomalySignals || "待补充"}
+- 实现层现实检查: ${fields.implementationRealityChecks || "待补充"}
+- 已考虑的替代解释: ${fields.alternativeExplanationsConsidered || "待补充"}
+- 交叉验证方法: ${fields.crossCheckMethod || "待补充"}
+- 当前最站得住的解释: ${fields.bestSupportedInterpretation || "待补充"}
+- 升级阈值: ${fields.escalationThreshold || "待补充"}
+
 ## Gate Ladder
 
 - 实验阶梯: ${fields.experimentLadder || "待补充"}
@@ -439,6 +556,27 @@ Use this file to define the paper-facing evaluation target, table plan, gates, a
 - Comparison implementation source: ${fields.comparisonImplementationSource || "TBD"}
 - Deviation from original implementation: ${fields.deviationFromOriginalImplementation || "TBD"}
 
+## Academic Validity Checks
+
+- Evaluation setting semantics: ${fields.evaluationSettingSemantics || "TBD"}
+- Visibility and leakage risks: ${fields.visibilityAndLeakageRisks || "TBD"}
+- Anchor and label policy: ${fields.anchorAndLabelPolicy || "TBD"}
+- Scale and comparability policy: ${fields.scaleAndComparabilityPolicy || "TBD"}
+- Metric validity checks: ${fields.metricValidityChecks || "TBD"}
+- Comparison validity checks: ${fields.comparisonValidityChecks || "TBD"}
+- Statistical validity checks: ${fields.statisticalValidityChecks || "TBD"}
+- Claim boundary: ${fields.claimBoundary || "TBD"}
+- Integrity self-check: ${fields.integritySelfCheck || "TBD"}
+
+## Sanity and Alternative-Explanation Checks
+
+- Anomaly signals: ${fields.anomalySignals || "TBD"}
+- Implementation reality checks: ${fields.implementationRealityChecks || "TBD"}
+- Alternative explanations considered: ${fields.alternativeExplanationsConsidered || "TBD"}
+- Cross-check method: ${fields.crossCheckMethod || "TBD"}
+- Best-supported interpretation: ${fields.bestSupportedInterpretation || "TBD"}
+- Escalation threshold: ${fields.escalationThreshold || "TBD"}
+
 ## Gate Ladder
 
 - Experiment ladder: ${fields.experimentLadder || "TBD"}
@@ -638,6 +776,74 @@ function hydrateEvalProtocol(targetDir) {
       protocol.metricImplementationSource,
       extractReportValue(reportText, "metricImplementationSource")
     ),
+    evaluationSettingSemantics: mergePreferred(
+      protocol.evaluationSettingSemantics,
+      extractReportValue(reportText, "setting"),
+      extractValue(missionText, ["Approved direction", "已批准方向"])
+    ),
+    visibilityAndLeakageRisks: mergePreferred(
+      protocol.visibilityAndLeakageRisks,
+      extractValue(dataDecisions, ["Remaining preprocessing or leakage risks", "剩余预处理或 leakage 风险"]),
+      "stay within the approved visible inputs and avoid unavailable information or leakage-prone artifacts"
+    ),
+    anchorAndLabelPolicy: mergePreferred(
+      protocol.anchorAndLabelPolicy,
+      "use only anchors and labels that are legitimate and visible inside the approved evaluation setting"
+    ),
+    scaleAndComparabilityPolicy: mergePreferred(
+      protocol.scaleAndComparabilityPolicy,
+      "compare only natively aligned scales or explicitly justified mappings recorded in the protocol"
+    ),
+    metricValidityChecks: mergePreferred(
+      protocol.metricValidityChecks,
+      "promote only metrics that directly support the active claim; keep health metrics as support checks"
+    ),
+    comparisonValidityChecks: mergePreferred(
+      protocol.comparisonValidityChecks,
+      "keep canonical, historical, recent, and closest-prior comparisons fair, source-backed, and justified"
+    ),
+    statisticalValidityChecks: mergePreferred(
+      protocol.statisticalValidityChecks,
+      "report minimum sample sizes, seed variance, and uncertainty before promotion"
+    ),
+    claimBoundary: mergePreferred(
+      protocol.claimBoundary,
+      extractValue(missionText, ["Approved direction", "已批准方向"])
+    ),
+    integritySelfCheck: mergePreferred(
+      protocol.integritySelfCheck,
+      "do not use unavailable information, do not promote health metrics into main claims, and do not treat workflow status as scientific evidence"
+    ),
+    anomalySignals: mergePreferred(
+      protocol.anomalySignals,
+      extractReportValue(reportText, "anomalySignals"),
+      "treat all-null outputs, suspiciously identical runs, no-op deltas, or impl/result mismatches as diagnostic triggers instead of findings"
+    ),
+    implementationRealityChecks: mergePreferred(
+      protocol.implementationRealityChecks,
+      extractReportValue(reportText, "implementationChecks"),
+      "inspect the concrete code path, parser or metric wiring, input chain, split logic, and output artifacts before interpreting an anomaly"
+    ),
+    alternativeExplanationsConsidered: mergePreferred(
+      protocol.alternativeExplanationsConsidered,
+      extractReportValue(reportText, "alternativeExplanations"),
+      "rule out simpler explanations such as parser bugs, stale checkpoints, split mistakes, leakage, or unfair comparisons before promoting a claim"
+    ),
+    crossCheckMethod: mergePreferred(
+      protocol.crossCheckMethod,
+      extractReportValue(reportText, "crossChecks"),
+      "use at least one independent cross-check such as direct artifact inspection, alternate parsing, tiny-slice rerun, or code-path verification"
+    ),
+    bestSupportedInterpretation: mergePreferred(
+      protocol.bestSupportedInterpretation,
+      extractReportValue(reportText, "bestSupportedInterpretation"),
+      "prefer the narrowest interpretation that still survives the anomaly checks and alternative explanations"
+    ),
+    escalationThreshold: mergePreferred(
+      protocol.escalationThreshold,
+      extractReportValue(reportText, "escalationThreshold"),
+      "switch to diagnostic mode or reviewer escalation when anomalies remain unresolved or risk increases across two rounds"
+    ),
     comparisonSourcePapers: mergePreferred(
       protocol.comparisonSourcePapers,
       extractReportValue(reportText, "baselineSourcePapers")
@@ -767,6 +973,21 @@ function renderSummary(lang, data) {
 - Comparison source papers: ${data.evalComparisonSourcePapers || "待补充"}
 - Comparison implementation source: ${data.evalComparisonImplementationSource || "待补充"}
 - Deviation from original implementation: ${data.evalDeviationFromOriginalImplementation || "待补充"}
+- Evaluation setting semantics: ${data.evalEvaluationSettingSemantics || "待补充"}
+- Visibility and leakage risks: ${data.evalVisibilityAndLeakageRisks || "待补充"}
+- Anchor and label policy: ${data.evalAnchorAndLabelPolicy || "待补充"}
+- Scale and comparability policy: ${data.evalScaleAndComparabilityPolicy || "待补充"}
+- Metric validity checks: ${data.evalMetricValidityChecks || "待补充"}
+- Comparison validity checks: ${data.evalComparisonValidityChecks || "待补充"}
+- Statistical validity checks: ${data.evalStatisticalValidityChecks || "待补充"}
+- Claim boundary: ${data.evalClaimBoundary || "待补充"}
+- Integrity self-check: ${data.evalIntegritySelfCheck || "待补充"}
+- Anomaly signals: ${data.evalAnomalySignals || "待补充"}
+- Implementation reality checks: ${data.evalImplementationRealityChecks || "待补充"}
+- Alternative explanations considered: ${data.evalAlternativeExplanationsConsidered || "待补充"}
+- Cross-check method: ${data.evalCrossCheckMethod || "待补充"}
+- Best-supported interpretation: ${data.evalBestSupportedInterpretation || "待补充"}
+- Escalation threshold: ${data.evalEscalationThreshold || "待补充"}
 - Experiment ladder: ${data.evalExperimentLadder || "待补充"}
 - Benchmark ladder: ${data.evalBenchmarkLadder || "待补充"}
 - Promotion gate: ${data.evalPromotionGate || "待补充"}
@@ -830,6 +1051,21 @@ function renderSummary(lang, data) {
 - Comparison source papers: ${data.evalComparisonSourcePapers || "TBD"}
 - Comparison implementation source: ${data.evalComparisonImplementationSource || "TBD"}
 - Deviation from original implementation: ${data.evalDeviationFromOriginalImplementation || "TBD"}
+- Evaluation setting semantics: ${data.evalEvaluationSettingSemantics || "TBD"}
+- Visibility and leakage risks: ${data.evalVisibilityAndLeakageRisks || "TBD"}
+- Anchor and label policy: ${data.evalAnchorAndLabelPolicy || "TBD"}
+- Scale and comparability policy: ${data.evalScaleAndComparabilityPolicy || "TBD"}
+- Metric validity checks: ${data.evalMetricValidityChecks || "TBD"}
+- Comparison validity checks: ${data.evalComparisonValidityChecks || "TBD"}
+- Statistical validity checks: ${data.evalStatisticalValidityChecks || "TBD"}
+- Claim boundary: ${data.evalClaimBoundary || "TBD"}
+- Integrity self-check: ${data.evalIntegritySelfCheck || "TBD"}
+- Anomaly signals: ${data.evalAnomalySignals || "TBD"}
+- Implementation reality checks: ${data.evalImplementationRealityChecks || "TBD"}
+- Alternative explanations considered: ${data.evalAlternativeExplanationsConsidered || "TBD"}
+- Cross-check method: ${data.evalCrossCheckMethod || "TBD"}
+- Best-supported interpretation: ${data.evalBestSupportedInterpretation || "TBD"}
+- Escalation threshold: ${data.evalEscalationThreshold || "TBD"}
 - Experiment ladder: ${data.evalExperimentLadder || "TBD"}
 - Benchmark ladder: ${data.evalBenchmarkLadder || "TBD"}
 - Promotion gate: ${data.evalPromotionGate || "TBD"}
@@ -948,6 +1184,21 @@ ${data.problem || "待补充"}
 - Comparison source papers: ${data.evalComparisonSourcePapers || "待补充"}
 - Comparison implementation source: ${data.evalComparisonImplementationSource || "待补充"}
 - Deviation from original implementation: ${data.evalDeviationFromOriginalImplementation || "待补充"}
+- Evaluation setting semantics: ${data.evalEvaluationSettingSemantics || "待补充"}
+- Visibility and leakage risks: ${data.evalVisibilityAndLeakageRisks || "待补充"}
+- Anchor and label policy: ${data.evalAnchorAndLabelPolicy || "待补充"}
+- Scale and comparability policy: ${data.evalScaleAndComparabilityPolicy || "待补充"}
+- Metric validity checks: ${data.evalMetricValidityChecks || "待补充"}
+- Comparison validity checks: ${data.evalComparisonValidityChecks || "待补充"}
+- Statistical validity checks: ${data.evalStatisticalValidityChecks || "待补充"}
+- Claim boundary: ${data.evalClaimBoundary || "待补充"}
+- Integrity self-check: ${data.evalIntegritySelfCheck || "待补充"}
+- Anomaly signals: ${data.evalAnomalySignals || "待补充"}
+- Implementation reality checks: ${data.evalImplementationRealityChecks || "待补充"}
+- Alternative explanations considered: ${data.evalAlternativeExplanationsConsidered || "待补充"}
+- Cross-check method: ${data.evalCrossCheckMethod || "待补充"}
+- Best-supported interpretation: ${data.evalBestSupportedInterpretation || "待补充"}
+- Escalation threshold: ${data.evalEscalationThreshold || "待补充"}
 - Experiment ladder: ${data.evalExperimentLadder || "待补充"}
 - Benchmark ladder: ${data.evalBenchmarkLadder || "待补充"}
 - Promotion gate: ${data.evalPromotionGate || "待补充"}
@@ -1022,6 +1273,21 @@ ${data.problem || "TBD"}
 - Comparison source papers: ${data.evalComparisonSourcePapers || "TBD"}
 - Comparison implementation source: ${data.evalComparisonImplementationSource || "TBD"}
 - Deviation from original implementation: ${data.evalDeviationFromOriginalImplementation || "TBD"}
+- Evaluation setting semantics: ${data.evalEvaluationSettingSemantics || "TBD"}
+- Visibility and leakage risks: ${data.evalVisibilityAndLeakageRisks || "TBD"}
+- Anchor and label policy: ${data.evalAnchorAndLabelPolicy || "TBD"}
+- Scale and comparability policy: ${data.evalScaleAndComparabilityPolicy || "TBD"}
+- Metric validity checks: ${data.evalMetricValidityChecks || "TBD"}
+- Comparison validity checks: ${data.evalComparisonValidityChecks || "TBD"}
+- Statistical validity checks: ${data.evalStatisticalValidityChecks || "TBD"}
+- Claim boundary: ${data.evalClaimBoundary || "TBD"}
+- Integrity self-check: ${data.evalIntegritySelfCheck || "TBD"}
+- Anomaly signals: ${data.evalAnomalySignals || "TBD"}
+- Implementation reality checks: ${data.evalImplementationRealityChecks || "TBD"}
+- Alternative explanations considered: ${data.evalAlternativeExplanationsConsidered || "TBD"}
+- Cross-check method: ${data.evalCrossCheckMethod || "TBD"}
+- Best-supported interpretation: ${data.evalBestSupportedInterpretation || "TBD"}
+- Escalation threshold: ${data.evalEscalationThreshold || "TBD"}
 - Experiment ladder: ${data.evalExperimentLadder || "TBD"}
 - Benchmark ladder: ${data.evalBenchmarkLadder || "TBD"}
 - Promotion gate: ${data.evalPromotionGate || "TBD"}
@@ -1286,6 +1552,21 @@ function buildContextSnapshot(targetDir) {
     evalComparisonSourcePapers: evalProtocol.comparisonSourcePapers,
     evalComparisonImplementationSource: evalProtocol.comparisonImplementationSource,
     evalDeviationFromOriginalImplementation: evalProtocol.deviationFromOriginalImplementation,
+    evalEvaluationSettingSemantics: evalProtocol.evaluationSettingSemantics,
+    evalVisibilityAndLeakageRisks: evalProtocol.visibilityAndLeakageRisks,
+    evalAnchorAndLabelPolicy: evalProtocol.anchorAndLabelPolicy,
+    evalScaleAndComparabilityPolicy: evalProtocol.scaleAndComparabilityPolicy,
+    evalMetricValidityChecks: evalProtocol.metricValidityChecks,
+    evalComparisonValidityChecks: evalProtocol.comparisonValidityChecks,
+    evalStatisticalValidityChecks: evalProtocol.statisticalValidityChecks,
+    evalClaimBoundary: evalProtocol.claimBoundary,
+    evalIntegritySelfCheck: evalProtocol.integritySelfCheck,
+    evalAnomalySignals: evalProtocol.anomalySignals,
+    evalImplementationRealityChecks: evalProtocol.implementationRealityChecks,
+    evalAlternativeExplanationsConsidered: evalProtocol.alternativeExplanationsConsidered,
+    evalCrossCheckMethod: evalProtocol.crossCheckMethod,
+    evalBestSupportedInterpretation: evalProtocol.bestSupportedInterpretation,
+    evalEscalationThreshold: evalProtocol.escalationThreshold,
     evalExperimentLadder: evalProtocol.experimentLadder,
     evalBenchmarkLadder: evalProtocol.benchmarkLadder,
     evalPromotionGate: evalProtocol.promotionGate,

package/lib/eval_protocol.cjs

@@ -79,6 +79,81 @@ const EVAL_PROTOCOL_FIELDS = [
     key: "deviationFromOriginalImplementation",
     labels: ["Deviation from original implementation", "与原始实现的偏差"],
   },
+  {
+    name: "Evaluation setting semantics",
+    key: "evaluationSettingSemantics",
+    labels: ["Evaluation setting semantics", "评测设定语义"],
+  },
+  {
+    name: "Visibility and leakage risks",
+    key: "visibilityAndLeakageRisks",
+    labels: ["Visibility and leakage risks", "可见性与泄漏风险"],
+  },
+  {
+    name: "Anchor and label policy",
+    key: "anchorAndLabelPolicy",
+    labels: ["Anchor and label policy", "锚点与标签策略"],
+  },
+  {
+    name: "Scale and comparability policy",
+    key: "scaleAndComparabilityPolicy",
+    labels: ["Scale and comparability policy", "尺度与可比性策略"],
+  },
+  {
+    name: "Metric validity checks",
+    key: "metricValidityChecks",
+    labels: ["Metric validity checks", "指标有效性检查"],
+  },
+  {
+    name: "Comparison validity checks",
+    key: "comparisonValidityChecks",
+    labels: ["Comparison validity checks", "对比方法有效性检查"],
+  },
+  {
+    name: "Statistical validity checks",
+    key: "statisticalValidityChecks",
+    labels: ["Statistical validity checks", "统计有效性检查"],
+  },
+  {
+    name: "Claim boundary",
+    key: "claimBoundary",
+    labels: ["Claim boundary", "结论边界"],
+  },
+  {
+    name: "Integrity self-check",
+    key: "integritySelfCheck",
+    labels: ["Integrity self-check", "完整性自检"],
+  },
+  {
+    name: "Anomaly signals",
+    key: "anomalySignals",
+    labels: ["Anomaly signals", "异常信号"],
+  },
+  {
+    name: "Implementation reality checks",
+    key: "implementationRealityChecks",
+    labels: ["Implementation reality checks", "实现层现实检查"],
+  },
+  {
+    name: "Alternative explanations considered",
+    key: "alternativeExplanationsConsidered",
+    labels: ["Alternative explanations considered", "已考虑的替代解释"],
+  },
+  {
+    name: "Cross-check method",
+    key: "crossCheckMethod",
+    labels: ["Cross-check method", "交叉验证方法"],
+  },
+  {
+    name: "Best-supported interpretation",
+    key: "bestSupportedInterpretation",
+    labels: ["Best-supported interpretation", "当前最站得住的解释"],
+  },
+  {
+    name: "Escalation threshold",
+    key: "escalationThreshold",
+    labels: ["Escalation threshold", "升级阈值"],
+  },
   {
     name: "Benchmark ladder",
     key: "benchmarkLadder",

package/lib/i18n.cjs

@@ -332,9 +332,12 @@ const ZH_SKILL_FILES = {
 - 必须用白话解释选定的主指标和次级指标：每个指标在衡量什么、越高还是越低更好、它是主结果指标还是健康度/支持性指标。
 - 如果出现 coverage、completeness、confidence 或类似健康度指标，必须明确说明这类指标回答的是“实验是否跑稳、证据是否完整”，而不是主要科学效应本身。
 - 要把最关键的背景来源、方法/基线来源和指标来源直接写进报告，不要把它们藏在 \`.lab/context/*\` 里。
+- 把 \`report.md\` 当作给外部评审或合作者看的研究 memo；来源章节必须给出人类可读的 anchor references，不能拿本地路径或内部 provenance 充数。
 - 如果 \`.lab/context/terminology-lock.md\` 里已经冻结了方法名和 contribution bullets，就必须把它们带进报告。
 - 方法概述必须用协作者能读懂的话说明：我们的方法大致怎么做、相对 closest prior work 或 strongest baseline 改了什么、这些 prior 方法各自做了什么，以及它们为什么在当前 claim 下仍然不够。
 - 只保留少量最关键的 prior work/baseline 锚点；每个锚点都要用一句话交代它做了什么和它的局限。
+- 在“背景来源”“方法与基线来源”“指标来源”里，每个锚点都必须包含：引用、它做了什么或衡量什么、以及至少一个局限或 caveat。
+- 内部 provenance 只能放到 \`工件状态\` 或 \`.lab/context/evidence-index.md\`，不能塞进来源章节。
 - 在起草报告前，先检查 \`.lab/context/mission.md\` 和 \`.lab/context/eval-protocol.md\` 是否仍是模板空壳。
 - 如果 canonical context 还是空壳，要先根据 frozen result artifacts、data-decisions、evidence-index 和已批准上下文回填“最小可信版本”，再写报告。
 - 如果回填后仍缺少协作者可读所需的关键字段，就必须把输出降级成 \`artifact-anchored interim report\`，不能冒充最终协作者报告。
@@ -699,6 +702,10 @@ const ZH_SKILL_FILES = {
 
 ## Checklist
 
+- 学术有效性检查是否已经填写，并且和实际实验设置保持一致？
+- 完整性自检是否排除了不可见输入、不合理指标使用和把工作流状态当成科学证据的做法？
+- 异常信号是否先被当成 diagnostic trigger，而不是被直接合理化成结果？
+- 在升格当前解释前，是否已经记录更简单的替代解释和至少一种交叉验证？
 - 是否把 claims 和 evidence 分开写清楚？
 - baseline 是否公平且足够强？
 - 数据集、切分和指标是否合理？
@@ -763,20 +770,36 @@ const ZH_SKILL_FILES = {
 
 ## 背景来源
 
-- 最关键的背景论文或 benchmark 参考：
-- 为什么这些来源足以锚定当前问题：
+- 参考 1：
+  - 引用：
+  - 做了什么：
+  - 为什么和当前问题相关：
+  - 对当前项目的局限：
 
 ## 方法与基线来源
 
-- 我们的方法来源或实现基础：
-- baseline 与 comparison 的来源论文：
-- baseline 与 comparison 的实现来源：
+- 参考 1：
+  - 引用：
+  - 做了什么：
+  - 为什么是这里的关键对照：
+  - 相对我们目标的局限：
 
 ## 指标来源
 
-- 指标来源论文：
-- 指标实现来源：
-- 与原始实现的偏差：
+- 参考 1：
+  - 引用：
+  - 衡量什么：
+  - 为什么适合这里：
+  - 局限或注意事项：
+
+## 异常与替代解释
+
+- 观察到的异常信号：
+- 做过的实现层检查：
+- 已排除的更简单解释：
+- 支撑当前解释的交叉验证：
+- 当前最站得住的解释：
+- 未来异常出现时的升级阈值：
 
 ## 怎么看主表
 
@@ -1107,7 +1130,13 @@ const ZH_SKILL_FILES = {
 - Objective:
 - Autonomy level: L2
 - Autonomy level 只表示执行权限级别，不表示论文 layer 或 table 编号。
+- 层级指南：
+  - \`L1\` = 一轮安全验证或简单 report 刷新
+  - \`L2\` = 默认推荐级别，适合冻结核心边界内的实验迭代
+  - \`L3\` = 激进 campaign，适合更大范围探索和可选写作
+- 如果不确定，默认推荐 \`L2\`。
 - 如果你想表达论文层、实验 phase 或主表，请明确写成 \`paper layer\`、\`phase\` 或 \`table\`。
+- 如果你的请求提到了论文层、实验 phase 或主表，但没写自治级别，就不要启动循环，先补明确的 \`Autonomy level\`。
 - Approval status: draft
 - Allowed stages: run, iterate, review, report
 - Success criteria:
@@ -1115,7 +1144,7 @@ const ZH_SKILL_FILES = {
 - Terminal goal target:
 - Required terminal artifact:
 - 如果 workflow language 是中文，摘要、清单条目、任务标签和进度更新都应使用中文。
-- 示例 Objective: 推进 paper layer 3 的 organizer enforcement，完成一轮 bounded protocol、测试、最小实现和一轮小规模结果。
+- 示例 Objective: 推进 paper layer 3，完成一轮 bounded protocol、测试、最小实现和一轮小规模结果。
 
 ## 循环预算
 
@@ -1178,6 +1207,21 @@ const ZH_SKILL_FILES = {
 - 对比方法来源论文:
 - 对比方法实现来源:
 - 与原始实现的偏差:
+- 评测设定语义:
+- 可见性与泄漏风险:
+- 锚点与标签策略:
+- 尺度与可比性策略:
+- 指标有效性检查:
+- 对比有效性检查:
+- 统计有效性检查:
+- 结论边界:
+- 完整性自检:
+- 异常信号:
+- 实现层现实检查:
+- 已考虑的替代解释:
+- 交叉验证方法:
+- 当前最站得住的解释:
+- 升级阈值:
 - 终止目标类型:
 - 终止目标目标值:
 - 必要终止工件:
@@ -1572,7 +1616,7 @@ 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 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\n## 如何输入 `/lab:auto`\n\n- 把 `Autonomy level L1/L2/L3` 视为执行权限级别，不要和论文里的 layer、phase、table 编号混用。\n- 把 `paper layer`、`phase`、`table` 视为实验目标。例如 `paper layer 3` 或 `Phase 1 reviewer fidelity` 不是 `Autonomy level L3`。\n- 一条好的 `/lab:auto` 输入应至少说清：objective、自治级别、terminal goal、scope、allowed modifications。\n- 如果 workflow language 是中文，摘要、清单条目、任务标签和进度更新都应使用中文，除非文件路径、代码标识符或字面指标名必须保持原样。\n- 示例：`/lab:auto 自治级别 L2。目标：推进 paper layer 3 的 organizer enforcement。终止条件：完成 bounded protocol、测试、最小实现和一轮小规模结果。允许修改：evaluator prompt registry、ingestion、parser。`\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\n## 如何输入 `/lab:auto`\n\n## `/lab:auto` 层级指南\n\n- `L1`：适合安全验证、一轮 bounded 真实运行，或简单 report 刷新。\n- `L2`：默认推荐级别，适合冻结核心边界内的常规实验迭代。\n- `L3`：激进 campaign 级别，只在你明确想做更大范围探索和可选写作时使用。\n- 如果不确定，默认推荐 `L2`。\n- 如果用户输入没写级别，或者把级别和 `paper layer`、`phase`、`table` 混用了，就应先停下来，要求用户明确选 `L1/L2/L3`。\n\n- 把 `Autonomy level L1/L2/L3` 视为执行权限级别，不要和论文里的 layer、phase、table 编号混用。\n- 把 `paper layer`、`phase`、`table` 视为实验目标。例如 `paper layer 3` 或 `Phase 1` 不是 `Autonomy level L3`。\n- 一条好的 `/lab:auto` 输入应至少说清：objective、自治级别、terminal goal、scope、allowed modifications。\n- 如果 workflow language 是中文，摘要、清单条目、任务标签和进度更新都应使用中文，除非文件路径、代码标识符或字面指标名必须保持原样。\n- 示例：`/lab:auto 自治级别 L2。目标：推进 paper layer 3。终止条件：完成 bounded protocol、测试、最小实现和一轮小规模结果。允许修改：配置、数据接入、评估脚本。`\n"
 );
 
 ZH_CONTENT[path.join(".codex", "prompts", "lab-data.md")] = codexPrompt(
@@ -1591,7 +1635,7 @@ 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",
   "[stage] [target]",
-  "# `/lab` for Claude\n\n`/lab` 是 Claude Code 里的 lab 工作流分发入口。调用方式有两种：\n\n- `/lab <stage> ...`\n- `/lab-idea`、`/lab-data`、`/lab-auto`、`/lab-framing`、`/lab-spec`、`/lab-run`、`/lab-iterate`、`/lab-review`、`/lab-report`、`/lab-write`\n\n## 阶段别名\n\n- `/lab idea ...` 或 `/lab-idea`\n- `/lab data ...` 或 `/lab-data`\n- `/lab auto ...` 或 `/lab-auto`\n- `/lab framing ...` 或 `/lab-framing`\n- `/lab spec ...` 或 `/lab-spec`\n- `/lab run ...` 或 `/lab-run`\n- `/lab iterate ...` 或 `/lab-iterate`\n- `/lab review ...` 或 `/lab-review`\n- `/lab report ...` 或 `/lab-report`\n- `/lab write ...` 或 `/lab-write`\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> ...` 或 `/lab-<stage>` 时，要立刻执行该 stage，而不是只推荐别的阶段。\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- `spec` 前应已有经批准的数据集与 benchmark 方案。\n- `run`、`iterate`、`auto`、`report` 都应遵循 `.lab/context/eval-protocol.md`。\n- `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- `write` 前必须已有经批准的 `framing` 工件。\n\n## 如何输入 `/lab auto`\n\n- 把 `Autonomy level L1/L2/L3` 视为执行权限级别，不要和论文里的 layer、phase、table 编号混用。\n- 把 `paper layer`、`phase`、`table` 视为实验目标。例如 `paper layer 3` 或 `Phase 1 reviewer fidelity` 不是 `Autonomy level L3`。\n- 一条好的 `/lab auto` 输入应至少说清：objective、自治级别、terminal goal、scope、allowed modifications。\n- 如果 workflow language 是中文，摘要、清单条目、任务标签和进度更新都应使用中文，除非文件路径、代码标识符或字面指标名必须保持原样。\n- 示例：`/lab auto 自治级别 L2。目标：推进 paper layer 3 的 organizer enforcement。终止条件：完成 bounded protocol、测试、最小实现和一轮小规模结果。允许修改：evaluator prompt registry、ingestion、parser。`\n"
+  "# `/lab` for Claude\n\n`/lab` 是 Claude Code 里的 lab 工作流分发入口。调用方式有两种：\n\n- `/lab <stage> ...`\n- `/lab-idea`、`/lab-data`、`/lab-auto`、`/lab-framing`、`/lab-spec`、`/lab-run`、`/lab-iterate`、`/lab-review`、`/lab-report`、`/lab-write`\n\n## 阶段别名\n\n- `/lab idea ...` 或 `/lab-idea`\n- `/lab data ...` 或 `/lab-data`\n- `/lab auto ...` 或 `/lab-auto`\n- `/lab framing ...` 或 `/lab-framing`\n- `/lab spec ...` 或 `/lab-spec`\n- `/lab run ...` 或 `/lab-run`\n- `/lab iterate ...` 或 `/lab-iterate`\n- `/lab review ...` 或 `/lab-review`\n- `/lab report ...` 或 `/lab-report`\n- `/lab write ...` 或 `/lab-write`\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> ...` 或 `/lab-<stage>` 时，要立刻执行该 stage，而不是只推荐别的阶段。\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- `spec` 前应已有经批准的数据集与 benchmark 方案。\n- `run`、`iterate`、`auto`、`report` 都应遵循 `.lab/context/eval-protocol.md`。\n- `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- `write` 前必须已有经批准的 `framing` 工件。\n\n## 如何输入 `/lab auto`\n\n## `/lab auto` 层级指南\n\n- `L1`：适合安全验证、一轮 bounded 真实运行，或简单 report 刷新。\n- `L2`：默认推荐级别，适合冻结核心边界内的常规实验迭代。\n- `L3`：激进 campaign 级别，只在你明确想做更大范围探索和可选写作时使用。\n- 如果不确定，默认推荐 `L2`。\n- 如果用户输入没写级别，或者把级别和 `paper layer`、`phase`、`table` 混用了，就应先停下来，要求用户明确选 `L1/L2/L3`。\n\n- 把 `Autonomy level L1/L2/L3` 视为执行权限级别，不要和论文里的 layer、phase、table 编号混用。\n- 把 `paper layer`、`phase`、`table` 视为实验目标。例如 `paper layer 3` 或 `Phase 1` 不是 `Autonomy level L3`。\n- 一条好的 `/lab auto` 输入应至少说清：objective、自治级别、terminal goal、scope、allowed modifications。\n- 如果 workflow language 是中文，摘要、清单条目、任务标签和进度更新都应使用中文，除非文件路径、代码标识符或字面指标名必须保持原样。\n- 示例：`/lab auto 自治级别 L2。目标：推进 paper layer 3。终止条件：完成 bounded protocol、测试、最小实现和一轮小规模结果。允许修改：配置、数据接入、评估脚本。`\n"
 );
 
 ZH_CONTENT[path.join(".claude", "commands", "lab-data.md")] = claudeCommand(
@@ -2034,6 +2078,27 @@ ZH_CONTENT[path.join(".lab", "context", "eval-protocol.md")] = `# 评估协议
 - 对比方法实现来源：
 - 与原始实现的偏差：
 
+## 学术有效性检查
+
+- 评测设定语义：
+- 可见性与泄漏风险：
+- 锚点与标签策略：
+- 尺度与可比性策略：
+- 指标有效性检查：
+- 对比有效性检查：
+- 统计有效性检查：
+- 结论边界：
+- 完整性自检：
+
+## 异常与替代解释检查
+
+- 异常信号：
+- 实现层现实检查：
+- 已考虑的替代解释：
+- 交叉验证方法：
+- 当前最站得住的解释：
+- 升级阈值：
+
 ## Gate Ladder
 
 - 实验阶梯：
@@ -2132,11 +2197,25 @@ ZH_CONTENT[path.join(".codex", "skills", "lab", "stages", "auto.md")] = `# \`/la
 
 ## 交互约束
 
+## 层级指南
+
+- \`L1\` = safe validation
+- \`L2\` = 默认推荐的 bounded iteration
+- \`L3\` = aggressive campaign
+- 如果不确定，默认推荐 \`L2\`。
+
 - 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。
+- 每次进入 \`/lab:auto\` 都要先给出这份层级指南。
 - 先做输入归一化：把 \`Autonomy level L1/L2/L3\` 视为执行权限级别，把 \`Layer 3\`、\`Phase 1\`、\`Table 2\` 视为论文范围目标。
+- 如果用户没有写自治级别，或者把自治级别和论文层、phase、table 混用了，就必须先给一版更详细的层级说明，至少解释：
+  - \`L1/L2/L3\` 的典型适用场景
+  - 每个级别允许改什么
+  - 每个级别通常在什么 stop boundary 停下
+  - 如果不确定，默认推荐 \`L2\`
+- 给完这版详细说明后，再追问一个明确的 \`L1/L2/L3\` 选择；在用户明确选级别前不要启动循环。
 - 如果用户同时提了论文层、实验 phase 和自治级别，先用一句话重述：objective、自治级别、terminal goal、scope、allowed modifications。
 - 如果 workflow language 是中文，摘要、清单条目、任务标签和进度更新都应使用中文，除非文件路径、代码标识符或字面指标名必须保持原样。
 - 当循环进入 \`report\` 时，要主动给出用户可读的白话总结，解释主指标、次级指标和主表作用；不要等用户额外发一句“解释这些指标”。

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

@@ -62,8 +62,16 @@ Use the same repository artifacts and stage boundaries every time.
 
 ## How to Ask for `/lab auto`
 
+## Level Guide for `/lab auto`
+
+- `L1` is the safe validation level. Use it for a smoke run, one bounded real run, or a simple review/report refresh when you do not want automatic iteration.
+- `L2` is the default recommended level. Use it for bounded experiment iteration inside a frozen core when you want auto to keep running until a gate, stop condition, or terminal goal is hit.
+- `L3` is the aggressive campaign level. Use it only when you explicitly want broad exploration, larger search space changes, and optional manuscript-writing work.
+- If you are unsure, choose `L2`.
+- If the request omits the level or mixes it with a paper layer, phase, or table target, `/lab auto` should stop and ask for an explicit autonomy level before arming the loop.
+
 - Treat `Autonomy level L1/L2/L3` as the execution privilege level, not as a paper layer, phase, or table number.
-- Treat `paper layer`, `phase`, and `table` as experiment targets. For example, `paper layer 3` or `Phase 1 reviewer fidelity` should not be interpreted as `Autonomy level L3`.
+- Treat `paper layer`, `phase`, and `table` as experiment targets. For example, `paper layer 3` or `Phase 1` should not be interpreted as `Autonomy level L3`.
 - A good `/lab auto` request should name:
   - the objective
   - the autonomy level
@@ -72,4 +80,4 @@ Use the same repository artifacts and stage boundaries every time.
   - the allowed modifications
 - If the repository workflow language is Chinese, summaries, checklist items, task labels, and progress updates should be written in Chinese unless a code identifier or file path must stay literal.
 - Good example:
-  - `/lab auto Autonomy level L2. Objective: advance paper layer 3 organizer enforcement. Terminal goal: task-completion. Scope: bounded protocol, tests, minimal implementation, and one small run. Allowed modifications: evaluator prompt registry, ingestion, and parser only.`
+  - `/lab auto Autonomy level L2. Objective: advance paper layer 3 through one bounded protocol improvement. Terminal goal: task-completion. Scope: bounded protocol, tests, one minimal implementation, and one small run. Allowed modifications: configuration, evaluation script, and data-loading logic only.`

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

@@ -56,8 +56,16 @@ argument-hint: workflow question or stage choice
 
 ## How to Ask for `/lab:auto`
 
+## Level Guide for `/lab:auto`
+
+- `L1` is the safe validation level. Use it for a smoke run, one bounded real run, or a simple review/report refresh when you do not want automatic iteration.
+- `L2` is the default recommended level. Use it for bounded experiment iteration inside a frozen core when you want auto to keep running until a gate, stop condition, or terminal goal is hit.
+- `L3` is the aggressive campaign level. Use it only when you explicitly want broad exploration, larger search space changes, and optional manuscript-writing work.
+- If you are unsure, choose `L2`.
+- If the request omits the level or mixes it with a paper layer, phase, or table target, `/lab:auto` should stop and ask for an explicit autonomy level before arming the loop.
+
 - Treat `Autonomy level L1/L2/L3` as the execution privilege level, not as a paper layer, phase, or table number.
-- Treat `paper layer`, `phase`, and `table` as experiment targets. For example, `paper layer 3` or `Phase 1 reviewer fidelity` should not be interpreted as `Autonomy level L3`.
+- Treat `paper layer`, `phase`, and `table` as experiment targets. For example, `paper layer 3` or `Phase 1` should not be interpreted as `Autonomy level L3`.
 - A good `/lab:auto` request should name:
   - the objective
   - the autonomy level
@@ -66,4 +74,4 @@ argument-hint: workflow question or stage choice
   - the allowed modifications
 - If the repository workflow language is Chinese, summaries, checklist items, task labels, and progress updates should be written in Chinese unless a code identifier or file path must stay literal.
 - Good example:
-  - `/lab:auto Autonomy level L2. Objective: advance paper layer 3 organizer enforcement. Terminal goal: task-completion. Scope: bounded protocol, tests, minimal implementation, and one small run. Allowed modifications: evaluator prompt registry, ingestion, and parser only.`
+  - `/lab:auto Autonomy level L2. Objective: advance paper layer 3 through one bounded protocol improvement. Terminal goal: task-completion. Scope: bounded protocol, tests, one minimal implementation, and one small run. Allowed modifications: configuration, evaluation script, and data-loading logic only.`

package/package-assets/shared/lab/.managed/scripts/validate_collaborator_report.py

@@ -20,6 +20,10 @@ REPORT_REQUIRED_SECTIONS = {
         r"^##\s+方法与基线来源\s*$",
     ],
     "Metric Sources": [r"^##\s+Metric Sources\s*$", r"^##\s+指标来源\s*$"],
+    "Sanity and Alternative Explanations": [
+        r"^##\s+Sanity and Alternative Explanations\s*$",
+        r"^##\s+异常与替代解释\s*$",
+    ],
 }
 
 MAIN_TABLES_REQUIRED_SECTIONS = {
@@ -30,6 +34,24 @@ MAIN_TABLES_REQUIRED_SECTIONS = {
     "How to Read These Tables": [r"^##\s+How to Read These Tables\s*$", r"^##\s+怎么读这些表\s*$"],
 }
 
+SOURCE_SECTION_NAMES = (
+    "Background Sources",
+    "Method and Baseline Sources",
+    "Metric Sources",
+)
+SOURCE_SECTION_PATH_MARKERS = (
+    "/Users/",
+    "/home/",
+    "/tmp/",
+    "/private/tmp/",
+    ".lab/",
+    "outputs/",
+    "docs/research/",
+)
+SOURCE_SECTION_CITATION_MARKERS = ("Citation:", "引用：")
+SOURCE_SECTION_ROLE_MARKERS = ("What it established:", "What it does:", "What it measures:", "做了什么：", "衡量什么：")
+SOURCE_SECTION_LIMITATION_MARKERS = ("Limitation", "局限")
+
 
 def parse_args():
     parser = argparse.ArgumentParser(
@@ -48,6 +70,35 @@ def missing_sections(text: str, required_sections: dict[str, list[str]]) -> list
     return missing
 
 
+def extract_section_body(text: str, patterns: list[str]) -> str:
+    for pattern in patterns:
+        match = re.search(pattern, text, flags=re.MULTILINE)
+        if not match:
+            continue
+        start = match.end()
+        next_heading = re.search(r"^##\s+", text[start:], flags=re.MULTILINE)
+        end = start + next_heading.start() if next_heading else len(text)
+        return text[start:end].strip()
+    return ""
+
+
+def validate_source_sections(text: str, label: str) -> list[str]:
+    issues = []
+    for section_name in SOURCE_SECTION_NAMES:
+        body = extract_section_body(text, REPORT_REQUIRED_SECTIONS[section_name])
+        if not body:
+            continue
+        if any(marker in body for marker in SOURCE_SECTION_PATH_MARKERS):
+            issues.append(f"{label} section '{section_name}' must not rely on local file paths or internal provenance")
+        if not any(marker in body for marker in SOURCE_SECTION_CITATION_MARKERS):
+            issues.append(f"{label} section '{section_name}' must include at least one citation anchor")
+        has_role = any(marker in body for marker in SOURCE_SECTION_ROLE_MARKERS)
+        has_limitation = any(marker in body for marker in SOURCE_SECTION_LIMITATION_MARKERS)
+        if not has_role or not has_limitation:
+            issues.append(f"{label} section '{section_name}' must explain what the anchor does and one limitation")
+    return issues
+
+
 def validate(path_str: str, required_sections: dict[str, list[str]], label: str) -> list[str]:
     path = Path(path_str)
     if not path.exists():
@@ -56,6 +107,8 @@ def validate(path_str: str, required_sections: dict[str, list[str]], label: str)
     missing = missing_sections(text, required_sections)
     if missing:
         return [f"{label} is missing required sections: {', '.join(missing)}"]
+    if label == "report.md":
+        return validate_source_sections(text, label)
     return []
 
 

package/package-assets/shared/lab/.managed/templates/final-report.md

@@ -54,20 +54,36 @@
 
 ## Background Sources
 
-- Most important background papers or benchmark references:
-- Why these are the right background anchors:
+- Anchor reference 1:
+  - Citation:
+  - What it established:
+  - Why it matters here:
+  - Limitation for the current project:
 
 ## Method and Baseline Sources
 
-- Our method source or implementation basis:
-- Baseline and comparison source papers:
-- Baseline and comparison implementation sources:
+- Anchor reference 1:
+  - Citation:
+  - What it does:
+  - Why it is the right anchor here:
+  - Limitation relative to our goal:
 
 ## Metric Sources
 
-- Metric source papers:
-- Metric implementation source:
-- Deviation from original implementation:
+- Anchor reference 1:
+  - Citation:
+  - What it measures:
+  - Why it is appropriate here:
+  - Limitation or caveat:
+
+## Sanity and Alternative Explanations
+
+- Anomaly signals observed:
+- Implementation checks performed:
+- Alternative explanations ruled out:
+- Cross-checks that strengthen the current interpretation:
+- Best-supported interpretation:
+- Escalation threshold if future anomalies appear:
 
 ## Experiment Setup
 

package/package-assets/shared/lab/.managed/templates/review-checklist.md

@@ -23,6 +23,10 @@
 
 ## Checklist
 
+- Are the academic validity checks filled and still consistent with the actual setup?
+- Does the integrity self-check rule out unavailable inputs, invalid metric use, and workflow-status-as-evidence?
+- Have anomaly signals been treated as diagnostic triggers instead of being rationalized into findings?
+- Are simpler alternative explanations and at least one cross-check recorded before promoting the current interpretation?
 - Are the dataset and split choices stated clearly?
 - Is the baseline fair, current, and reproducible?
 - Are the primary and secondary metrics justified?

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

@@ -9,7 +9,13 @@ If `eval-protocol.md` declares structured rung entries, auto mode follows those
 - Objective: 
 - Autonomy level: L2
 - Autonomy level controls execution privilege, not paper layer or table number.
+- Level guide:
+  - `L1` = safe validation over one bounded run/review/report cycle
+  - `L2` = default recommended bounded iteration inside a frozen core
+  - `L3` = aggressive campaign with broader exploration and optional writing
+- If you are unsure, choose `L2`.
 - If you mean a paper layer, phase, or table, spell it explicitly as `paper layer`, `phase`, or `table`.
+- If your request mentions a paper layer, phase, or table but omits the autonomy level, do not arm the loop until the level is explicit.
 - Approval status: draft
 - Allowed stages: run, iterate, review, report
 - Success criteria:
@@ -17,7 +23,7 @@ If `eval-protocol.md` declares structured rung entries, auto mode follows those
 - Terminal goal target:
 - Required terminal artifact:
 - If the workflow language is Chinese, keep summaries, checklist items, task labels, and progress updates in Chinese.
-- Example objective: advance paper layer 3 organizer enforcement through one bounded protocol, tests, minimal implementation, and one small run.
+- Example objective: advance paper layer 3 through one bounded protocol, tests, minimal implementation, and one small run.
 
 ## Loop Budget
 

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

@@ -13,6 +13,21 @@
 - Comparison source papers:
 - Comparison implementation source:
 - Deviation from original implementation:
+- Evaluation setting semantics:
+- Visibility and leakage risks:
+- Anchor and label policy:
+- Scale and comparability policy:
+- Metric validity checks:
+- Comparison validity checks:
+- Statistical validity checks:
+- Claim boundary:
+- Integrity self-check:
+- Anomaly signals:
+- Implementation reality checks:
+- Alternative explanations considered:
+- Cross-check method:
+- Best-supported interpretation:
+- Escalation threshold:
 - Terminal goal type:
 - Terminal goal target:
 - Required terminal artifact:

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

@@ -29,6 +29,27 @@ Use this file to define the paper-facing evaluation objective, table plan, gates
 
 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.
 
+## Academic Validity Checks
+
+- Evaluation setting semantics:
+- Visibility and leakage risks:
+- Anchor and label policy:
+- Scale and comparability policy:
+- Metric validity checks:
+- Comparison validity checks:
+- Statistical validity checks:
+- Claim boundary:
+- Integrity self-check:
+
+## Sanity and Alternative-Explanation Checks
+
+- Anomaly signals:
+- Implementation reality checks:
+- Alternative explanations considered:
+- Cross-check method:
+- Best-supported interpretation:
+- Escalation threshold:
+
 ## Gate Ladder
 
 - Experiment ladder:

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

@@ -87,6 +87,7 @@ Use this skill when the user invokes `/lab:*` or asks for the structured researc
 - 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.
+- Surface the level guide every time `/lab:auto` starts, and make the detailed guide mandatory when the user omits the level or mixes it with a paper layer, phase, or table target.
 - 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.

package/package-assets/shared/skills/lab/stages/auto.md

@@ -40,6 +40,8 @@
 - Treat `/lab:auto` as an orchestration layer, not a replacement for existing `/lab:*` stages.
 - Treat `.lab/context/eval-protocol.md` as the source of truth for paper-facing metrics, metric glossary, table plan, gates, and structured experiment ladders.
 - Treat the evaluation protocol as source-backed, not imagination-backed: metric definitions, baseline behavior, comparison implementations, and deviations must come from recorded sources before they are used in gates or promotions.
+- Treat `Academic Validity Checks` and `Integrity self-check` as mandatory automation gates. Auto mode should not proceed, promote, or declare success while those fields are missing, stale, or contradicted by the current rung.
+- Treat `Sanity and Alternative-Explanation Checks` as the anomaly gate for automation. When a rung yields all-null outputs, suspiciously identical runs, no-op deltas, or impl/result mismatches, pause promotion logic until implementation reality checks, alternative explanations, and at least one cross-check are recorded.
 - Treat paper-template selection as an explicit write-time gate, not as a silent fallback, when the loop is about to create `.tex` deliverables for the first time.
 - The contract must declare `Autonomy level` and `Approval status`, and execution starts only when approval is explicitly set to `approved`.
 - The contract must also declare a concrete terminal goal:
@@ -74,6 +76,8 @@
   - `write` must produce LaTeX output under `<deliverables_root>/paper/`
 - Treat promotion as incomplete unless it writes back to `data-decisions.md`, `decisions.md`, `state.md`, and `session-brief.md`.
 - Do not stop or promote on the basis of a metric or comparison claim whose source-backed definition is missing from the approved evaluation protocol.
+- Before each rung and before each success, stop, or promotion decision, re-check the generic academic-risk questions: setting semantics, visibility/leakage, anchor or label policy, scale comparability, metric validity, comparison validity, statistical validity, claim boundary, and integrity self-check.
+- Before each success, stop, or promotion decision, also re-check the anomaly policy: whether anomaly signals fired, whether simpler explanations were ruled out, whether a cross-check was performed, and whether the current interpretation is still the narrowest supported one.
 
 ## Minimum Procedure
 
@@ -90,7 +94,15 @@
 
 ## Interaction Contract
 
+## Level Guide
+
+- `L1` = safe validation
+- `L2` = default recommended bounded iteration
+- `L3` = aggressive campaign
+- If you are unsure, choose `L2`.
+
 - Start with a concise summary of the objective, the frozen core, and the next automatic stage.
+- Always surface the level guide before execution.
 - If the contract is incomplete, ask one clarifying question at a time.
 - If multiple next actions are credible, present 2-3 bounded options with trade-offs before arming a long run.
 - Only ask for approval when the next step would leave the approved exploration envelope, exceed the chosen autonomy level, or materially change the frozen core.
@@ -100,8 +112,14 @@
 - Normalize ambiguous user requests before arming the loop.
 - Treat `Autonomy level L1/L2/L3` as execution privilege only.
 - Treat `Layer`, `Phase`, and `Table` references as paper-structure or experiment-scope targets, not as autonomy levels.
+- If the user does not name an autonomy level, or mixes it with a paper layer, phase, or table target, stop and deliver a detailed level guide before execution. That detailed guide should explain:
+  - the typical use case for `L1`, `L2`, and `L3`
+  - what kinds of modifications each level allows
+  - what kind of stop boundary each level is meant for
+  - that `L2` is the default recommendation when the user is unsure
+- Ask for one explicit level choice before arming the loop after that detailed guide.
 - Example:
-  - `Layer 3 organizer enforcement` means a paper layer or experiment target.
+  - `Layer 3` means a paper layer or experiment target.
   - `Autonomy level L3` means the aggressive campaign permission envelope.
 - If the user mixes framework work and experiment work in one request, restate a normalized contract with:
   - objective

package/package-assets/shared/skills/lab/stages/iterate.md

@@ -64,6 +64,10 @@ If the loop stops without success, record:
 - Do not change metric definitions, baseline semantics, or comparison implementations unless the approved evaluation protocol records both their sources and any deviations.
 - When you change ladders, sample sizes, or promotion gates, keep the resulting logic anchored to the source-backed evaluation protocol instead of ad-hoc chat reasoning.
 - Keep `.lab/context/eval-protocol.md` synchronized with the active benchmark scope, ladder gates, source-backed metric definitions, and any accepted implementation deviations instead of leaving it as a stale template.
+- Re-run the `Academic Validity Checks` and `Integrity self-check` whenever you change inputs, anchors, labels, metrics, comparisons, or promotion logic.
+- Re-run the `Sanity and Alternative-Explanation Checks` whenever a round produces anomaly signals, suspiciously unchanged results, impl/result mismatches, or other outcomes that could still have simpler explanations.
+- If a round reveals leakage risk, invalid scale comparisons, unsupported metric semantics, or overstated claim boundaries, treat that as a fatal methodological flaw instead of a normal failed iteration.
+- If anomaly signals remain unresolved after implementation reality checks and at least one cross-check, switch to diagnostic mode instead of continuing as if the interpretation were settled.
 
 ## Interaction Contract
 

package/package-assets/shared/skills/lab/stages/report.md

@@ -50,11 +50,18 @@
 - Explain the selected primary and secondary metrics in plain language for the user: what each metric measures, whether higher or lower is better, and whether it is a main result metric or only a health/support metric.
 - If coverage, completeness, confidence, or similar health metrics appear, explicitly say that they describe experimental reliability rather than the main scientific effect.
 - Pull the core background references, method or baseline references, and metric references out of the approved evaluation protocol instead of hiding them in `.lab/context/*`.
+- Treat `report.md` as an external-review-ready memo. Source sections must not rely on local file paths or internal provenance notes; they must give a few human-readable anchor references instead.
 - Pull the approved method name and contribution bullets out of `.lab/context/terminology-lock.md` when that framing context exists; do not silently drop them from the collaborator-facing report.
 - Explain the method overview in collaborator language: what the method roughly does, what changed relative to the closest prior work or strongest baseline, what those prior methods do, and why they remain insufficient for the approved claim.
 - When citing prior work or baselines in the method overview, include only the few anchor references a collaborator needs, and summarize their role and limitation in one short line each.
 - Report only the few references a collaborator needs to orient themselves quickly; do not turn `report.md` into a full bibliography dump.
+- In `Background Sources`, `Method and Baseline Sources`, and `Metric Sources`, every anchor must include a citation line, one short line about what it established or measures, and one limitation or caveat.
+- Internal provenance belongs in `Artifact Status` or `.lab/context/evidence-index.md`, not in the external-review-ready source sections.
 - If the report depends on a deviation from an original metric or implementation, state that deviation explicitly instead of smoothing it over.
+- Carry the approved `Claim boundary` into the collaborator-facing report instead of implying broader validity than the protocol allows.
+- If the `Academic Validity Checks` or `Integrity self-check` sections are incomplete, contradictory, or obviously violated by the evidence, degrade the report instead of presenting it as collaborator-ready.
+- Carry the protocol's anomaly handling into the report: summarize any anomaly signals, what implementation checks were performed, which simpler explanations were ruled out, what cross-check was used, and why the current interpretation is still the best-supported one.
+- If anomaly signals remain unresolved, or the report cannot explain why the current interpretation beat the simpler alternatives, degrade the report instead of presenting it as collaborator-ready.
 - Before drafting the report, inspect `.lab/context/mission.md` and `.lab/context/eval-protocol.md` for skeletal template fields.
 - If either canonical context file is still skeletal, hydrate the smallest trustworthy version from frozen result artifacts, dataset decisions, evidence-index, and prior approved context, and write that back before finalizing the report.
 - If collaborator-critical fields still remain missing after hydration, downgrade the output to an `artifact-anchored interim report` instead of presenting it as a final collaborator-ready report.

package/package-assets/shared/skills/lab/stages/review.md

@@ -24,6 +24,10 @@
 
 ## Reviewer Priorities
 
+- academic validity checks are missing, shallow, or contradicted by the actual setup
+- integrity self-check is missing or obviously ignored
+- anomaly signals are being rationalized away without concrete code, input-chain, or artifact checks
+- simpler alternative explanations were not examined before elevating the preferred interpretation
 - unfair or weak baselines
 - missing canonical baselines, strong historical baselines, recent strong public methods, or closest prior work without a justified omission
 - unrepresentative benchmark mix or missing classic-public versus recent-strong-public coverage

package/package-assets/shared/skills/lab/stages/run.md

@@ -28,7 +28,11 @@
 - Fail fast on data, environment, or metric wiring problems.
 - Tie the run to the approved evaluation protocol, not just an ad-hoc chat goal.
 - Do not invent metric definitions, baseline behavior, or comparison implementations from memory; anchor them to the approved evaluation protocol and its recorded sources.
+- Treat `Academic Validity Checks` and `Integrity self-check` as preflight gates, not optional notes. Do not bless a run as protocol-valid until those fields are filled and still match the current experiment.
+- Treat `Sanity and Alternative-Explanation Checks` as a second preflight gate. If anomaly signals have fired and the implementation reality checks, alternative explanations, cross-check method, best-supported interpretation, or escalation threshold are still blank, do not bless the run as valid evidence.
 - If `.lab/context/eval-protocol.md` is still skeletal, write the smallest trustworthy version of the current evaluation objective, metric set, ladder, and source-backed implementation notes before treating the run as the new protocol anchor.
+- Refuse to treat a run as scientifically valid if the protocol has not answered the generic academic-risk questions: setting semantics, visibility/leakage, anchor or label policy, scale comparability, metric validity, comparison validity, statistical validity, and claim boundary.
+- Treat all-null outputs, suspiciously identical reruns, no-op deltas, and impl/result mismatches as diagnostic triggers first; check code paths and rule out simpler explanations before interpreting them as findings.
 - Record the exact launch command and output location.
 - Write durable run outputs, logs, and checkpoints under `results_root`.
 - Write figures or plots under `figures_root`.

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "superlab",
-  "version": "0.1.22",
+  "version": "0.1.24",
   "description": "Strict /lab research workflow installer for Codex and Claude",
   "keywords": [
     "codex",