fix: resolve legacy evidence targets without guessing and guide retries

backend/wisegrade/llm.py

@@ -238,11 +238,19 @@ def validate_score(result: ScoreResult, criterion: dict, chunks: list[dict], *,
             # Empty identifiers are accepted only for pre-object legacy rubrics.
             if any(isinstance(value, dict) for value in criterion.get("required_evidence", [])):
                 raise ModelError("模型证据缺少 target_id")
-            target = default_content or (targets[0] if targets else None)
+            associated = [by_target[key] for key in chunk.get("target_ids", []) if key in by_target]
+            if associated:
+                if len(associated) != 1:
+                    raise ModelError(f"模型证据缺少 target_id：chunk_id={chunk['chunk_id']} 对应多个检索目标，请明确选择")
+                target = associated[0]
+            else:
+                target = default_content or (targets[0] if targets else None)
         if target is None:
             raise ModelError("模型证据未关联任何必备证据目标")
         if chunk.get("target_ids") and target["target_id"] not in chunk["target_ids"]:
-            raise ModelError("模型证据未关联该片段的检索目标")
+            allowed = ",".join(str(key) for key in chunk["target_ids"] if key in by_target)
+            raise ModelError(f"模型证据未关联该片段的检索目标：chunk_id={chunk['chunk_id']}，"
+                             f"target_id={target['target_id']}；允许 target_id={allowed}")
         if not page_type_allowed(page_type, target["allowed_page_types"]):
             raise ModelError(f"模型证据来自不允许的页面类型（{page_type}），不能支持「{target['label']}」")
         if target["use"] == USE_CONTENT and not content_evidence_allowed(page_type):

backend/wisegrade/scoring.py

@@ -227,6 +227,8 @@ def _score_criterion(db, batch, project, criterion, client, *, retriever, multi_
     prompt = (batch.course_snapshot["prompts"]["score"] +
               " 证据quote必须从所引用chunk_id的text逐字复制连续片段，保留原文标点和换行；page必须等于该片段的page；"
               "target_id必须取自required_evidence_targets。不要改写、缩写或拼接引文。"
+              "每条引文的 target_id 必须属于该片段的 target_ids；同一片段可对应多个目标，"
+              "请按实际内容选择，不要按证据列表顺序分配编号。"
               "不得把目录、摘要或章节标题当作内容证据；无法据此判断内容时把score设为null。")
     error = ""
     result = None
@@ -239,7 +241,11 @@ def _score_criterion(db, batch, project, criterion, client, *, retriever, multi_
             error = str(exc)
             if attempt:
                 raise
-            prompt += f" 上次回答未通过校验（{error}）；请改用允许页面类型中的实质正文片段，并逐字复制quote。"
+            payload["validation_feedback"] = {
+                "error": error,
+                "instruction": "根据具体错误修正输出；引用目标只能从对应片段的 target_ids 选择，"
+                               "quote须为该片段原文。没有足够依据时保留空分，不要编造引用。"}
+            prompt += " 上次回答未通过校验；请按 validation_feedback 修正，不要重复错误引用。"
     score = db.scalar(select(CriterionScore).where(CriterionScore.id == score.id)
                       .with_for_update().execution_options(populate_existing=True))
     if score.status == STATUS_REVIEWED or score.teacher_score is not None: