feat:

1. 建立问题库大纲，对应每个人生阶段槽位
2. 鼓励使用更生活化的交流语言共情与总结
3. 降低评审模型可能发生截断的概率
4. 成稿质量维度强化情感表达和上下文连贯性

api/app/features/evaluation/conversation_compare_summary.py

@@ -5,6 +5,7 @@ from __future__ import annotations
 from typing import Any
 
 from app.features.evaluation.judge_schemas import ConversationJudgeOutput
+from app.features.evaluation.judge_service import trim_compare_transcript_pair
 
 _GROUP_KEYS: tuple[tuple[str, str], ...] = (
     ("emotion_score", "情绪与陪伴"),
@@ -44,6 +45,32 @@ def _has_repeat_issue(judge: ConversationJudgeOutput | None) -> bool:
     )
 
 
+def _evidence_quality(truncation: dict[str, Any]) -> dict[str, Any]:
+    """结构化说明「分数/对比文在多大程度上覆盖全量对话」，便于客观解读。"""
+    b_h = not truncation["baseline_truncated_for_conversation"]
+    r_h = not truncation["replay_truncated_for_conversation"]
+    pair_full = not (
+        truncation["baseline_truncated_for_compare"]
+        or truncation["replay_truncated_for_compare"]
+    )
+    if b_h and r_h and pair_full:
+        scope = "full"
+        note = "评分与 A/B 对比均基于当前注入的全文（在模型上下文内未再裁对话正文）。"
+    else:
+        scope = "partial"
+        note = (
+            "存在整段或对比环节截断：分数与流式结论仅反映已提交片段；"
+            "评审侧已注入截断边界说明，长程细项应保守。发布决策请结合逐轮分、人工抽查或更高上下文预算。"
+        )
+    return {
+        "scope": scope,
+        "baseline_holistic_covers_full_text": b_h,
+        "replay_holistic_covers_full_text": r_h,
+        "ab_compare_covers_full_transcripts": pair_full,
+        "note_zh": note,
+    }
+
+
 def build_conversation_compare_summary(
     *,
     baseline_judge: ConversationJudgeOutput | None,
@@ -51,29 +78,42 @@ def build_conversation_compare_summary(
     baseline_transcript: str,
     replay_transcript: str,
     conv_cap: int,
-    compare_cap_each: int,
+    compare_cap_total: int,
+    compare_per_side_cap: int | None = None,
     fixture_filename: str | None = None,
 ) -> dict[str, Any]:
+    _, _, baseline_cmp_trunc, replay_cmp_trunc = trim_compare_transcript_pair(
+        baseline_transcript,
+        replay_transcript,
+        total_max_chars=int(compare_cap_total),
+        per_side_max_chars=compare_per_side_cap,
+    )
+    if compare_per_side_cap and compare_per_side_cap > 0:
+        each_hint = int(compare_per_side_cap)
+    else:
+        each_hint = max(1, int(compare_cap_total) // 2)
+
     truncation = {
         "baseline_chars": len((baseline_transcript or "").strip()),
         "replay_chars": len((replay_transcript or "").strip()),
         "conversation_cap_chars": int(conv_cap),
-        "compare_cap_each_chars": int(compare_cap_each),
+        "compare_cap_total_chars": int(compare_cap_total),
+        "compare_cap_each_chars": each_hint,
         "baseline_truncated_for_conversation": len((baseline_transcript or "").strip())
         > int(conv_cap),
         "replay_truncated_for_conversation": len((replay_transcript or "").strip())
         > int(conv_cap),
-        "baseline_truncated_for_compare": len((baseline_transcript or "").strip())
-        > int(compare_cap_each),
-        "replay_truncated_for_compare": len((replay_transcript or "").strip())
-        > int(compare_cap_each),
+        "baseline_truncated_for_compare": baseline_cmp_trunc,
+        "replay_truncated_for_compare": replay_cmp_trunc,
     }
+    evidence_quality = _evidence_quality(truncation)
 
     if not replay_judge:
         return {
             "fixture_filename": fixture_filename,
             "mode": "single",
             "truncation": truncation,
+            "evidence_quality": evidence_quality,
             "gate": {
                 "status": "insufficient_data",
                 "reasons": ["缺少回放整体评分，无法判断是否追平或超过 A。"],
@@ -86,9 +126,12 @@ def build_conversation_compare_summary(
             "mode": "single",
             "replay_total": _round(replay_judge.total_score),
             "truncation": truncation,
+            "evidence_quality": evidence_quality,
             "gate": {
                 "status": "single_side_only",
-                "reasons": ["当前只有新对话单侧评分，可用于优化，但不能判定是否超过 A。"],
+                "reasons": [
+                    "当前只有新对话单侧评分，可用于优化，但不能判定是否超过 A。"
+                ],
             },
         }
 
@@ -150,8 +193,20 @@ def build_conversation_compare_summary(
         reasons.append(f"关键回落维度：{'、'.join(key_regressions[:4])}。")
     if key_gains:
         reasons.append(f"关键提升维度：{'、'.join(key_gains[:4])}。")
-    if truncation["baseline_truncated_for_compare"] or truncation["replay_truncated_for_compare"]:
-        reasons.append("A/B 对比稿使用了截断 transcript，长对话结论需结合逐轮评分复核。")
+    if (
+        truncation["baseline_truncated_for_compare"]
+        or truncation["replay_truncated_for_compare"]
+    ):
+        reasons.append(
+            "A/B 对比稿使用了截断 transcript，长对话结论需结合逐轮评分复核。"
+        )
+    if (
+        truncation["baseline_truncated_for_conversation"]
+        or truncation["replay_truncated_for_conversation"]
+    ):
+        reasons.append(
+            "整段评分可能仅见 transcript 前缀；长程维度已在评审边界下保守处理，请结合逐轮分或全文重跑交叉验证。"
+        )
 
     return {
         "fixture_filename": fixture_filename,
@@ -165,6 +220,7 @@ def build_conversation_compare_summary(
         "key_gains": key_gains,
         "repeat_issue_detected": has_repeat_regression,
         "truncation": truncation,
+        "evidence_quality": evidence_quality,
         "gate": {
             "status": status,
             "parity_passed": parity_passed,
@@ -173,4 +229,3 @@ def build_conversation_compare_summary(
             "golden_set_note": "建议在固定黄金样本集上复跑该口径，再决定是否发布。",
         },
     }
-