diff --git a/src/main/java/com/sztzjy/linkCommerce/service/impl/AiTrainingEvaluationServiceImpl.java b/src/main/java/com/sztzjy/linkCommerce/service/impl/AiTrainingEvaluationServiceImpl.java
index c5845d2..7f87931 100644
--- a/src/main/java/com/sztzjy/linkCommerce/service/impl/AiTrainingEvaluationServiceImpl.java
+++ b/src/main/java/com/sztzjy/linkCommerce/service/impl/AiTrainingEvaluationServiceImpl.java
@@ -216,15 +216,27 @@ public class AiTrainingEvaluationServiceImpl implements AiTrainingEvaluationServ
     }
 
     private AiTrainingEvaluationView runAssessment(EvaluationContext context, JwtUser actor, AiTrainingEvaluation evaluation) {
+        long startedAt = System.currentTimeMillis();
         String answerSnapshot = answerSnapshot(context.answer);
         String taskSnapshot = taskSnapshot(context.effectiveContent);
         List<AssessmentCriterionPlan> criterionPlan = assessmentCriterionPlan(context.effectiveContent.getSteps());
         String rawResponse = null;
         try {
+            long knowledgeStartedAt = System.currentTimeMillis();
             String knowledgeContext = knowledgeContext(context, actor, "");
+            long knowledgeCost = System.currentTimeMillis() - knowledgeStartedAt;
             String choiceQuestionContext = choiceQuestionContext(context.effectiveContent, context.answer);
+            long modelStartedAt = System.currentTimeMillis();
             rawResponse = qwenChatClient.completeJson(assessmentSystemPrompt(criterionPlan),
                     assessmentUserPrompt(taskSnapshot, answerSnapshot, knowledgeContext, choiceQuestionContext));
+            long modelCost = System.currentTimeMillis() - modelStartedAt;
+            // AI 评分过慢排查:分段记录知识检索与模型调用耗时(prompt 长度一并输出)
+            log.info("AI assessment timing, evaluationId={}, knowledgeContext={} ms, modelCall={} ms, total={} ms, "
+                            + "knowledgeContextChars={}, answerSnapshotChars={}, promptChars={}",
+                    evaluation.getId(), knowledgeCost, modelCost, System.currentTimeMillis() - startedAt,
+                    knowledgeContext.length(), answerSnapshot.length(),
+                    assessmentSystemPrompt(criterionPlan).length()
+                            + assessmentUserPrompt(taskSnapshot, answerSnapshot, knowledgeContext, choiceQuestionContext).length());
             JsonNode assessmentReport = object(rawResponse);
             int score = validateAssessment(assessmentReport, criterionPlan);
             compactAssessmentRationales(assessmentReport);
@@ -302,20 +314,34 @@ public class AiTrainingEvaluationServiceImpl implements AiTrainingEvaluationServ
             // AI 评分报告是评语的主要依据;助学报告只做补充,没有也能生成
             throw new ServiceException(HttpStatus.BAD_REQUEST, "该学生还没有AI评分数据，请先执行AI评分");
         }
+        String comment = normalizeTeacherComment(generateCommentOnce(assessmentReport, helpReport, ""));
+        if (comment.codePointCount(0, comment.length()) < 30) {
+            // 验收要求 30-50 字;过短时带纠正指令重写一次,取更长的那条
+            String instruction = " 上一次生成的评语只有" + comment.codePointCount(0, comment.length())
+                    + "个字，太短；重新生成的评语必须不少于30个字、不超过50个字。";
+            String retried = normalizeTeacherComment(generateCommentOnce(assessmentReport, helpReport, instruction));
+            if (retried.codePointCount(0, retried.length()) > comment.codePointCount(0, comment.length())) {
+                comment = retried;
+            }
+        }
+        return Collections.singletonMap("comment", comment);
+    }
+
+    private String generateCommentOnce(String assessmentReport, String helpReport, String extraInstruction) {
         String rawResponse = qwenChatClient.completeJson(commentSystemPrompt(),
-                commentUserPrompt(assessmentReport, helpReport));
+                commentUserPrompt(assessmentReport, helpReport) + StringUtils.trimToEmpty(extraInstruction));
         JsonNode report = object(rawResponse);
         JsonNode commentNode = report.get("comment");
-        if (commentNode == null || !commentNode.isTextual() || StringUtils.isBlank(commentNode.textValue())) {
+        if (commentNode == null || !commentNode.isTextual()) {
             throw invalidReport();
         }
-        return Collections.singletonMap("comment", normalizeTeacherComment(commentNode.textValue()));
+        return commentNode.textValue();
     }
 
     private String commentSystemPrompt() {
         return "You are a vocational school teacher writing feedback for one student's training task. "
                 + "Based on the given AI scoring report (and AI tutoring report if present), write ONE Chinese teacher comment. "
-                + "Requirements: 10 to 50 Chinese characters; a single short paragraph; objective, specific and encouraging; "
+                + "Requirements: 30 to 50 Chinese characters; a single short paragraph; objective, specific and encouraging; "
                 + "comment on the performance and give one improvement direction; "
                 + "no bullet points, no quotes, no salutation, no prefix or suffix such as 评语. "
                 + "Return only JSON: {\"comment\":\"...\"}";