From 54534be0c17b5edba1044f1d440d55666d4357f2 Mon Sep 17 00:00:00 2001
From: YANG JIANKUAN <timyih@icloud.com>
Date: Wed, 4 Nov 2026 12:26:21 +0800
Subject: [PATCH] =?UTF-8?q?fix(evals):=20=E6=BC=8F=E5=93=8D=E5=BA=94?=
 =?UTF-8?q?=E5=9C=A8=E6=89=80=E6=9C=89=E8=A1=A8=E9=87=8C=E9=83=BD=E5=88=A0?=
 =?UTF-8?q?=EF=BC=9B=E8=AF=84=E5=88=86=E6=8C=89=E5=8F=A5=E5=AD=90=E9=87=8D?=
 =?UTF-8?q?=E5=8F=A0=E5=90=88=E5=B9=B6=E9=87=8D=E5=A4=8D=E7=9A=84=E6=A0=87?=
 =?UTF-8?q?=E6=B3=A8=E5=8D=95=E5=85=83?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
---
 evals/check/mutate.py |  7 +++++++
 evals/check/score.py  | 23 +++++++++++++++++++++--
 2 files changed, 28 insertions(+), 2 deletions(-)

diff --git a/evals/check/mutate.py b/evals/check/mutate.py
index 475174a..31ddf0e 100644
--- a/evals/check/mutate.py
+++ b/evals/check/mutate.py
@@ -82,6 +82,7 @@ class Mutator:
             for r in chunk:
                 self.mark(r.get("responds") or [], "D01", "删连续一段", "block", worded=False)
             del rows[start : start + n]
+            self.drop_everywhere([u for r in chunk for u in r.get("responds") or []])
             return
 
     def row_ops(self, n: int) -> None:
@@ -101,6 +102,11 @@ class Mutator:
                     done += 1
                     break
 
+    def drop_everywhere(self, units: list[str]) -> None:
+        """漏响应要在所有表里都漏：S04 的技术参数同时列在全量偏离表和 ★ 表里，只删一张表等于没漏。"""
+        for t in self.tables():
+            t["rows"] = [r for r in t["rows"] if not set(r.get("responds") or []) & set(units)]
+
     @staticmethod
     def col(t: dict[str, Any], field: str) -> int | None:
         cols = t.get("columns") or ["no", "requirement", "answer", "deviation", "note"]
@@ -114,6 +120,7 @@ class Mutator:
         units = r.get("responds") or []
         if op == "D01":
             t["rows"].remove(r)
+            self.drop_everywhere(units)
             self.mark(units, "D01", "删一行", worded=False)
         elif op == "D03":
             return self._unmet(r, tp, ans, dev)
diff --git a/evals/check/score.py b/evals/check/score.py
index fc0cecb..c7cae89 100644
--- a/evals/check/score.py
+++ b/evals/check/score.py
@@ -64,6 +64,20 @@ def main() -> None:
         for s in u.sids:
             sid_units[s].add(u.id)
 
+    # 标注集里同一条要求有时记了两个单元（两个类别各一条，或表格一行拆成几条）：句子有重叠的单元合成一组，按组计
+    parent = {u.id: u.id for u in units}
+
+    def find(x: str) -> str:
+        while parent[x] != x:
+            x = parent[x]
+        return x
+
+    for ids in sid_units.values():
+        first, *rest = sorted(ids)
+        for o in rest:
+            parent[find(o)] = find(first)
+    group = {u.id: find(u.id) for u in units}
+
     per_unit: dict[str, list[dict[str, Any]]] = defaultdict(list)
     unmapped: list[dict[str, Any]] = []
     for it in res["items"]:
@@ -82,8 +96,13 @@ def main() -> None:
     fp_by_trap: Counter[str] = Counter()
     ok_units = trap_units = 0
     other_findings = 0
-    for uid, e in manifest["units"].items():
-        items = per_unit.get(uid, [])
+    seen_groups: set[str] = set()
+    for uid, e in sorted(manifest["units"].items(), key=lambda kv: kv[1].get("expect") != "defect"):
+        g = group.get(uid, uid)
+        if g in seen_groups:
+            continue  # 同组的单元只算一次（挂了问题的那个优先）
+        seen_groups.add(g)
+        items = [it for x in manifest["units"] if group.get(x, x) == g for it in per_unit.get(x, [])]
         found = [it for it in items if it["finding"]]
         if e["expect"] == "defect":
             if not e.get("reviewed") and not a.include_unreviewed: