eric/bid-doctor
1
0
Fork 0
Code Issues Pull Requests Actions 1 Packages Projects Releases Wiki Activity

feat(check): 响应方式识别（C12）和「这条要不要投标人回答」；评测集纠错

Browse Source 
- checklist/response_mode.py：按类别（技术 / 商务）识别招标文件的响应方式（逐条响应 / 默认接受），规则优先；
  两种写法冲突按逐条，认不出按逐条并写进覆盖范围。
- check/applicability.py：无效情形（★ / 斜体下划线标注的条款除外）、日期与递交、没要求附材料的评分标准、
  无需提供证明或由采购人查询的资格，不逐条找响应，写进覆盖范围；要求附证明材料按原文判断，中标后才交的不算。
- 默认接受的要求没找到响应：一般要求记 implicit_accepted、不出问题；实质性要求记 implicit_review、提示级
  人工核对；要附证明材料的不免（迁移 0006）。
- match.py：修同一行单元格又是别的条目候选句时的 KeyError（每条清单项单独记候选句，句子从总表里查）。
- 评测集：注入器同一要求的所有表格行一起改；散文、附件、格式文件里也响应了的要求不挂漏响应类问题；
  错位响应必须是另一个参数；复核判不成立的单元不再挂；regen.py 按固定种子重新生成。

开发集 17 份：规矩投标文件报出的问题 639 → 229，误报 21.3 → 12.4 / 100，召回 68% → 65%，致命级 20 → 21 / 32
（docs/architecture/check-eval.md §4）。

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
This commit is contained in:
YANG JIANKUAN
2026-11-04 14:25:14 +08:00
parent b7432643dd
commit cb0c2fcafe
13 changed files with 492 additions and 21 deletions
Download Patch File Download Diff File Expand all files Collapse all files

75
evals/check/mutate.py
View File

@@ -41,10 +41,18 @@ UNIT_DOWN = {"TB": ("GB", 1024), "GB": ("MB", 1024), "Gbps": ("Mbps", 1000), "s"
class Mutator:
def __init__(self, spec: dict[str, Any], vocab: dict[str, Any], rng: random.Random) -> None:
self.spec, self.vocab, self.rng = spec, vocab, rng
self.vols = spec["volumes"]
self.unit_defects: dict[str, dict[str, Any]] = {}
self.rule_defects: list[dict[str, Any]] = []
self.umbrella: set[str] = set()
self.vols = spec["volumes"]
self.rejected: set[str] = set() # 以前复核判为不成立的单元（同一招标文件的任何投标文件里），不再挂问题
# 在表格以外（技术方案散文、附件、格式文件）也响应了的单元：删表格行、改响应值都不构成问题，不挂这几类
self.prose = [
{t[i : i + 2] for t in [re.sub(r"\W", "", b["text"])] for i in range(len(t) - 1)}
for b in self.blocks
if b.get("type") == "para" and b.get("form") == "prose"
]
self.elsewhere = {u for b in self.blocks if b.get("type") != "table" for u in b.get("responds") or []}
@property
def blocks(self) -> list[dict[str, Any]]:
@@ -80,7 +88,7 @@ class Mutator:
return out
def free(self, units: list[str]) -> bool:
return bool(units) and not any(u in self.unit_defects for u in units)
return bool(units) and not any(u in self.unit_defects or u in self.rejected for u in units)
def mark(
self, units: list[str], defect: str, variant: str, granularity: str = "single", worded: bool = True
@@ -103,10 +111,11 @@ class Mutator:
chunk = rows[start : start + n]
if not all(self.free(r.get("responds") or []) for r in chunk if r.get("responds")):
continue
if any(set(r.get("responds") or []) & self.elsewhere for r in chunk):
continue
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 []])
self.drop_everywhere([u for r in chunk for u in r.get("responds") or []], [(t, r) for r in chunk])
return
def row_ops(self, n: int) -> None:
@@ -126,10 +135,43 @@ 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。"""
def in_prose(self, text: str) -> bool:
"""技术方案散文里基本复述了这条要求（字 2-gram 有六成以上出现在同一段里）：表格里删掉或改错，散文里还在响应。"""
t = re.sub(r"\W", "", text)
grams = {t[i : i + 2] for i in range(len(t) - 1)}
return len(grams) >= 6 and any(len(grams & g) / len(grams) >= 0.6 for g in self.prose)
def _req_text(self, t: dict[str, Any], r: dict[str, Any]) -> str:
at = self.col(t, "requirement")
return re.sub(r"\W", "", r["cells"][at]) if at is not None else ""
def siblings(self, t: dict[str, Any], r: dict[str, Any]) -> list[tuple[dict[str, Any], dict[str, Any]]]:
"""同一条要求在投标文件里的所有表格行：挂同一个标注单元，或者要求原文相同（读标清单常把同一要求抽成几条，
S04 的参数同时列在两张表里）。注入的问题要在这些行上一起做，否则别处还留着正确的响应，就不是问题了。"""
units, text = set(r.get("responds") or []), self._req_text(t, r)
out = []
for t2 in self.tables():
for r2 in t2["rows"]:
if r2 is r:
continue
if units & set(r2.get("responds") or []) or (len(text) >= 6 and self._req_text(t2, r2) == text):
out.append((t2, r2))
return out
def propagate(self, t: dict[str, Any], r: dict[str, Any]) -> None:
"""把这一行改过的响应、偏离、说明照样写到同一要求的其他行上。"""
for t2, r2 in self.siblings(t, r):
for field in ("answer", "deviation", "note"):
a, b = self.col(t, field), self.col(t2, field)
if a is not None and b is not None:
r2["cells"][b] = r["cells"][a]
def drop_everywhere(self, units: list[str], rows: list[tuple[dict[str, Any], dict[str, Any]]] = ()) -> 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；
要求原文相同的重复行也一起删。"""
drop = {id(r2) for t, r in rows for _, r2 in self.siblings(t, r)} | {id(r) for _, r in rows}
for t in self.tables():
t["rows"] = [r for r in t["rows"] if not set(r.get("responds") or []) & set(units)]
t["rows"] = [r for r in t["rows"] if id(r) not in drop and not set(r.get("responds") or []) & set(units)]
@staticmethod
def col(t: dict[str, Any], field: str) -> int | None:
@@ -148,12 +190,14 @@ class Mutator:
numeric = tp.get("op") in (">=", "<=", "=") and tp.get("value")
if op == "D05" and (numeric or not FUNCTIONAL.search(req_text) or NON_FUNCTIONAL.search(req_text)):
return False # 「暂不支持」只套在功能类要求上（数值参数、证书、清单、服务义务不会这么写）
if op in ("D01", "D03", "D08", "D09") and (set(units) & self.elsewhere or self.in_prose(req_text)):
return False
if op == "D01":
t["rows"].remove(r)
self.drop_everywhere(units)
self.drop_everywhere(units, [(t, r)])
self.mark(units, "D01", "删一行", worded=False)
elif op == "D03":
return self._unmet(r, tp, ans, dev)
if not self._unmet(r, tp, ans, dev):
return False
elif op == "D05":
param = tp.get("param") or (r["cells"][req][:12] if req is not None else "该功能")
r["cells"][ans] = self.phrase("D05", "phrases").format(param=param)
@@ -171,11 +215,14 @@ class Mutator:
r["cells"][ans] = self.phrase("D08", "phrases")
self.mark(units, "D08", r["cells"][ans], worded=False)
elif op == "D09":
mine = str(tp.get("param") or "")
others = [
x
for _, x in cands
if x is not r
and (x.get("tp") or {}).get("param")
and x["tp"]["param"] != mine # 必须是另一个参数：同名参数（另一台设备的同一指标）的响应可能恰好满足
and (not mine or mine not in x["cells"][x.get("anchor", ans)])
and x["cells"][x.get("anchor", ans)] != r["cells"][ans]
]
if not others:
@@ -183,6 +230,8 @@ class Mutator:
o = self.rng.choice(others)
r["cells"][ans] = o["cells"][o.get("anchor", ans)]
self.mark(units, "D09", "另一条参数的响应", worded=False)
if op != "D01":
self.propagate(t, r)
return True
def _unmet(self, r: dict[str, Any], tp: dict[str, Any], ans: int, dev: int | None) -> bool:
@@ -205,7 +254,7 @@ class Mutator:
if dev is not None and self.rng.random() < 0.5:
r["cells"][dev] = "无偏离" # 一半在偏离栏仍写「无偏离」（另一半保持原样，也是无偏离或正偏离）
self.mark(r.get("responds") or [], "D03", variant, worded=False)
return True
return True # 同一要求的其他行由调用方（_apply）同步
def implicit_deviation(self, n: int) -> None:
"""只列偏离的表：挑默认接受的要求写成偏离（D06 负偏离未说明 / D05 写明不支持）。"""
@@ -421,6 +470,10 @@ def main() -> None:
optional = {u.id for u in units if u.optional}
modes = unit_modes(spec) # 注入前算：被删掉的行也要知道它本来该怎么响应
m = Mutator(spec, json.loads(VOCAB[a.split].read_text()), random.Random(a.seed)) # noqa: S311 评测数据，不用于安全
for f in OUT.glob(f"{tender}-*/manifest.json"):
for uid, e in json.loads(f.read_text())["units"].items():
if (e.get("review") or {}).get("ok") is False:
m.rejected.add(uid) # noqa: S311 评测数据，不用于安全
if not a.manifest_only:
if a.density == "high":
m.d01_block()