diff --git a/index.html b/index.html
index ebc156a..3dc40b9 100644
--- a/index.html
+++ b/index.html
@@ -770,18 +770,11 @@ function handleAction(action){
     renderEntryEssay();
   }
 }
-function applyUnderstandingCorrection(text){
-  const idx = parseParagraphIndex(text);
-  if(idx===null || !analysis || !analysis.paragraphs[idx]){
-    globalConstraints.push(text);
-    return {scope:'global', label:'文章理解的补充'};
-  }
-  analysis.paragraphs[idx].natural_meaning_zh = text;
-  analysis.paragraphs[idx].semantic_anchor = text;
-  const pid = pidOf(idx);
-  (paragraphConstraints[pid]=paragraphConstraints[pid]||[]).push(text);
-  return {scope:pid, label:`${paragraphLabel(idx)} 的理解`};
-}
+// 原先这里是 applyUnderstandingCorrection：顾问在理解态说「第三段重点不是……」时，
+// 它把这句话同时写进 analysis.paragraphs[i] 的 natural_meaning_zh / semantic_anchor 和
+// paragraphConstraints。Agent 化之后这条输入走 record_advisor_note，只落成分段约束，
+// **锚点不再跟着改**（纠正内容仍会进诊断/复检的 paragraph_constraints）。
+// 这是已知偏差，不是遗漏：要不要补一个带 anchor 更新的「纠正理解」工具待定。
 function applyDiagnosisConstraint(text){
   const raw = text.trim();
   const idx = parseParagraphIndex(raw);