feat(v2): T07-a 预审结果约束与结论校验
T07 是"预审执行"大卡，按纪律拆子卡。本批为第一块：纯确定性的结论校验， 不调用任何模型、不计算概率。覆盖 AC-05/06/07/09/12/45 与 BR-05/06/07/12。 新增 services/prescreen_policy.py： - decide()：四级优先（hard_block → critical_incomplete → open_findings → clean）， 优先级顺序即业务口径，改动等于改结论，已显式标注 - **硬性阻断不可覆盖**：命中 hard 规则即 stop，并指出规则版本、保留模型原始 结构化提案供审计；专项模型的正面单题判断不能翻盘（有专门反例） - **未知不等于风险也不等于通过**：关键核查不可完成给 insufficient_information 并列出未完成项，findings 的 result 原样保留、**不得被改写成 issue** - **没有 second path 给概率**：probability() 不接受任何入参，模型分数无法顺手传入； native_metrics 里名为 probability 的指标会被 assert_no_probability_leak 拦下， self_reported 置信度可入档但不得变概率 - stop 明确不触碰正式审批（AC-45），预算耗尽按 E-08 列未完成项而非伪造完成 实现中修正了两处自己的测试写法：断言"不含高风险三字"会被免责说明本身命中； 断言 open_findings 为空与"unknown 属待处理"的定义矛盾。两处都改成检验真正的 不变量（结论非风险判定 + findings 不被改写）。 测试：新增 test_prescreen_policy.py 19 项。全量 368 项 OK（349 → 368）； verify.py 10/10 PASS。 Co-Authored-By: Claude Code <noreply@anthropic.com>
This commit is contained in:
Claude Code
@@ -0,0 +1,126 @@
|
||||
"""T07-a 预审结果约束：把核查结果收敛成受约束的建议。
|
||||
|
||||
依据：docs/v2/01-requirements.md §5、03-detailed-design.md §5；
|
||||
BR-05/06/07/12、E-02/06/07/08、AC-05/06/07/09/12/45。
|
||||
|
||||
本模块只做**确定性结论校验**，不调用任何模型、不算概率。四条不可让渡的规则：
|
||||
|
||||
1. **硬性阻断不可被覆盖**（BR-06）：已确认的 hard 规则命中即 `stop`，模型的高分
|
||||
或"建议推进"都不能翻盘。同时**保留模型原始结构化提案**供审计，并指出规则版本。
|
||||
2. **未知不等于风险，也不等于通过**（BR-07）：关键核查不可完成时给
|
||||
`insufficient_information` 并列补件，**不得**据此评判企业高风险。
|
||||
3. **单题不能代表综合意见**（BR-05）：专项判断只进 findings 与被引用，不能直接
|
||||
写入 recommendation。
|
||||
4. **未校准就没有概率**（BR-12）：`probability` 恒为 None + `not_calibrated`，
|
||||
任何模型自报置信度都不改写它。
|
||||
"""
|
||||
|
||||
from __future__ import annotations
|
||||
|
||||
RECOMMENDATIONS = ("advance", "supplement", "stop", "insufficient_information")
|
||||
|
||||
NOT_CALIBRATED = "not_calibrated"
|
||||
NO_PROBABILITY_REASON = "尚未完成真实历史校准；不展示审批通过概率"
|
||||
|
||||
# 判定优先级（从上到下，先命中先返回）。顺序本身是业务口径，改动等于改结论。
|
||||
PRECEDENCE = ("hard_block", "critical_incomplete", "open_findings", "clean")
|
||||
|
||||
INCOMPLETE_RESULTS = ("unknown", "error", "not_applicable")
|
||||
|
||||
|
||||
def _codes(findings) -> set[str]:
|
||||
return {str(item.get("code")) for item in findings or () if item.get("code")}
|
||||
|
||||
|
||||
def hard_blocks(findings, hard_rule_codes) -> list[dict]:
|
||||
"""命中的硬性阻断项。`hard_rule_codes` 来自冻结方案里 effect=hard 且 on_match=block 的规则。"""
|
||||
wanted = set(hard_rule_codes or ())
|
||||
return [dict(item) for item in findings or () if item.get("code") in wanted]
|
||||
|
||||
|
||||
def critical_incomplete(findings, critical_check_ids) -> list[str]:
|
||||
"""关键核查项中未完成的那些编码（未知/执行失败/不适用）。"""
|
||||
critical = set(critical_check_ids or ())
|
||||
return sorted({
|
||||
str(item.get("code")) for item in findings or ()
|
||||
if item.get("code") in critical and item.get("result") in INCOMPLETE_RESULTS
|
||||
})
|
||||
|
||||
|
||||
def open_findings(findings) -> list[dict]:
|
||||
"""需要业务处理但未阻断的问题（issue / unknown）。"""
|
||||
return [dict(item) for item in findings or ()
|
||||
if item.get("result") in ("issue", "unknown")]
|
||||
|
||||
|
||||
def probability() -> dict:
|
||||
"""P0 恒为空。**不接受任何入参**——不给"顺手传个模型分数"留口子。"""
|
||||
return {"value": None, "reason": NOT_CALIBRATED, "message": NO_PROBABILITY_REASON}
|
||||
|
||||
|
||||
def decide(findings, missing_inputs=(), degradations=(), judgments=(),
|
||||
critical_check_ids=(), hard_rule_codes=(), rule_versions=()) -> dict:
|
||||
"""收敛结论。返回 recommendation 与支撑它的理由、降级与审计信息。"""
|
||||
blocks = hard_blocks(findings, hard_rule_codes)
|
||||
incomplete = critical_incomplete(findings, critical_check_ids)
|
||||
open_items = open_findings(findings)
|
||||
|
||||
degradations = [dict(item) for item in degradations or ()]
|
||||
missing = sorted({str(item) for item in missing_inputs or ()})
|
||||
|
||||
if blocks:
|
||||
# 命中硬阻断：即便专项模型给出正面单题判断，也不改变结论（BR-05 / BR-06）
|
||||
stage = "hard_block"
|
||||
recommendation = "stop"
|
||||
reason = "存在已确认的硬性阻断，暂不建议推进；" + "、".join(
|
||||
f'{item.get("code")}@{item.get("basis_version", "未标注版本")}' for item in blocks)
|
||||
elif incomplete:
|
||||
stage = "critical_incomplete"
|
||||
recommendation = "insufficient_information"
|
||||
reason = ("关键核查未能完成，无法形成结论，请补充材料后重新发起；"
|
||||
"不得据此判定企业高风险。未完成项：" + "、".join(incomplete))
|
||||
missing = sorted(set(missing) | set(incomplete))
|
||||
elif open_items:
|
||||
stage = "open_findings"
|
||||
recommendation = "supplement"
|
||||
reason = f"存在 {len(open_items)} 项待处理事项，建议补充材料或说明后重新核查"
|
||||
else:
|
||||
stage = "clean"
|
||||
recommendation = "advance"
|
||||
reason = "已完成方案内的确定性核查，未发现阻断项；正式审批仍由业务流程决定"
|
||||
|
||||
return {
|
||||
"recommendation": recommendation,
|
||||
"recommendation_reason": reason,
|
||||
"stage": stage,
|
||||
"findings": [dict(item) for item in findings or ()],
|
||||
"open_findings": open_items,
|
||||
"hard_blocks": blocks,
|
||||
"critical_incomplete": incomplete,
|
||||
"missing_inputs": missing,
|
||||
"degradations": degradations,
|
||||
# 专项判断只作为依据入档，**不参与** recommendation 的判定
|
||||
"judgments": [dict(item) for item in judgments or ()],
|
||||
"judgment_can_override": False,
|
||||
"rule_versions": [dict(item) for item in rule_versions or ()],
|
||||
"probability": probability(),
|
||||
# AC-45：stop 只是"暂不建议推进"，系统不得自动关闭项目或执行正式审批
|
||||
"formal_approval_touched": False,
|
||||
}
|
||||
|
||||
|
||||
def assert_no_probability_leak(result: dict) -> None:
|
||||
"""任何路径都不得产出未校准概率（BR-12 / AC-12）。"""
|
||||
value = (result.get("probability") or {}).get("value")
|
||||
if value is not None:
|
||||
raise ValueError("未校准的数据集不得产出审批通过概率")
|
||||
for item in result.get("judgments", ()):
|
||||
for metric in item.get("native_metrics", ()) or ():
|
||||
if str(metric.get("name", "")).lower() in ("probability", "approval_probability"):
|
||||
raise ValueError("专项模型的原生指标不得被当作审批概率")
|
||||
|
||||
|
||||
def budget_exhausted(reason: str = "RUN_BUDGET_EXCEEDED") -> dict:
|
||||
"""预算耗尽（E-08）：停止并列出未完成项，不伪造完成。"""
|
||||
return {"code": reason, "message": "本次核查已达到执行上限，请复核未完成项",
|
||||
"retryable": False}
|
||||
@@ -0,0 +1,156 @@
|
||||
"""T07-a 预审结果约束测试：AC-05/06/07/09/12/45。
|
||||
|
||||
每条规则都配反例——只测"正常给出 advance"证明不了约束存在。
|
||||
本模块纯确定性，不调用模型、不算概率。
|
||||
"""
|
||||
|
||||
import unittest
|
||||
|
||||
from services import prescreen_policy as policy
|
||||
|
||||
|
||||
def finding(code, result, basis_version=None, **over):
|
||||
item = {"code": code, "title": f"检查项 {code}", "result": result}
|
||||
if basis_version:
|
||||
item["basis_version"] = basis_version
|
||||
item.update(over)
|
||||
return item
|
||||
|
||||
|
||||
class HardBlockTests(unittest.TestCase):
|
||||
def test_hard_block_forces_stop(self):
|
||||
result = policy.decide([finding("lease.coverage.below", "issue")],
|
||||
hard_rule_codes=["lease.coverage.below"],
|
||||
critical_check_ids=["lease.coverage.below"])
|
||||
self.assertEqual(result["recommendation"], "stop")
|
||||
self.assertEqual(result["stage"], "hard_block")
|
||||
|
||||
def test_hard_block_cites_the_rule_version(self):
|
||||
result = policy.decide([finding("r.hard", "issue", basis_version="rule.credit@v3")],
|
||||
hard_rule_codes=["r.hard"])
|
||||
self.assertIn("rule.credit@v3", result["recommendation_reason"])
|
||||
|
||||
def test_positive_single_judgment_cannot_override_a_hard_block(self):
|
||||
# AC-05 / BR-05：单题正面判断不能覆盖综合意见
|
||||
judgments = [{"question_id": "q1", "status": "completed",
|
||||
"answer": {"kind": "choice", "value": "supported"}}]
|
||||
result = policy.decide([finding("r.hard", "issue")], hard_rule_codes=["r.hard"],
|
||||
judgments=judgments)
|
||||
self.assertEqual(result["recommendation"], "stop")
|
||||
self.assertFalse(result["judgment_can_override"])
|
||||
self.assertEqual(result["judgments"], judgments, "判断仍应入档供审计")
|
||||
|
||||
def test_hard_block_beats_insufficient_information(self):
|
||||
result = policy.decide(
|
||||
[finding("r.hard", "issue"), finding("k1", "unknown")],
|
||||
hard_rule_codes=["r.hard"], critical_check_ids=["k1"])
|
||||
self.assertEqual(result["recommendation"], "stop")
|
||||
|
||||
def test_block_does_not_touch_formal_approval(self):
|
||||
# AC-45：stop 只表示暂不建议推进，系统不得自动关闭项目或审批
|
||||
result = policy.decide([finding("r.hard", "issue")], hard_rule_codes=["r.hard"])
|
||||
self.assertFalse(result["formal_approval_touched"])
|
||||
self.assertIn("暂不建议推进", result["recommendation_reason"])
|
||||
|
||||
def test_advisory_code_is_not_a_hard_block(self):
|
||||
result = policy.decide([finding("r.advisory", "issue")],
|
||||
hard_rule_codes=["r.hard"])
|
||||
self.assertNotEqual(result["recommendation"], "stop")
|
||||
|
||||
|
||||
class InsufficientInformationTests(unittest.TestCase):
|
||||
def test_incomplete_critical_check_never_becomes_high_risk(self):
|
||||
# AC-07 / BR-07：全部关键资料缺失时给信息不足，**不判高风险**
|
||||
result = policy.decide([finding("k1", "unknown"), finding("k2", "unknown")],
|
||||
critical_check_ids=["k1", "k2"])
|
||||
self.assertEqual(result["recommendation"], "insufficient_information")
|
||||
# 要求不是"别出现高风险三个字"（免责说明本身就会写到它），而是：
|
||||
# 结论不能变成风险判定，且不得凭空生成"高风险"这类 finding
|
||||
self.assertIn("不得据此判定企业高风险", result["recommendation_reason"])
|
||||
self.assertNotEqual(result["recommendation"], "stop")
|
||||
# 缺证可以（也应该）出现在待处理列表里，但**不得被改写成风险问题**：
|
||||
# findings 数量与 result 取值必须原样保留，不能凭空多出 issue
|
||||
self.assertEqual(len(result["findings"]), 2)
|
||||
self.assertTrue(all(item["result"] == "unknown" for item in result["findings"]))
|
||||
self.assertFalse(any(item["result"] == "issue" for item in result["open_findings"]))
|
||||
|
||||
def test_missing_inputs_include_the_incomplete_critical_checks(self):
|
||||
result = policy.decide([finding("k1", "unknown")], critical_check_ids=["k1"],
|
||||
missing_inputs=["现金流材料"])
|
||||
self.assertEqual(result["missing_inputs"], ["k1", "现金流材料"])
|
||||
|
||||
def test_execution_failure_counts_as_incomplete(self):
|
||||
result = policy.decide([finding("k1", "error")], critical_check_ids=["k1"])
|
||||
self.assertEqual(result["recommendation"], "insufficient_information")
|
||||
|
||||
def test_non_critical_unknown_does_not_force_insufficient(self):
|
||||
# 非关键项的未知只要求补件，不阻断成信息不足
|
||||
result = policy.decide([finding("other", "unknown")], critical_check_ids=["k1"])
|
||||
self.assertEqual(result["recommendation"], "supplement")
|
||||
|
||||
|
||||
class SupplementAndAdvanceTests(unittest.TestCase):
|
||||
def test_open_issue_leads_to_supplement(self):
|
||||
result = policy.decide([finding("c1", "issue")])
|
||||
self.assertEqual(result["recommendation"], "supplement")
|
||||
|
||||
def test_clean_run_advances_but_does_not_approve(self):
|
||||
result = policy.decide([finding("c1", "pass")])
|
||||
self.assertEqual(result["recommendation"], "advance")
|
||||
self.assertIn("正式审批仍由业务流程决定", result["recommendation_reason"])
|
||||
|
||||
def test_no_findings_at_all_is_treated_as_clean(self):
|
||||
result = policy.decide([])
|
||||
self.assertEqual(result["stage"], "clean")
|
||||
self.assertEqual(result["open_findings"], [])
|
||||
|
||||
|
||||
class DegradationTests(unittest.TestCase):
|
||||
def test_degradations_are_preserved_not_swallowed(self):
|
||||
result = policy.decide([finding("c1", "pass")],
|
||||
degradations=[{"code": "JUDGMENT_UNAVAILABLE",
|
||||
"message": "专项核查未完成"}])
|
||||
self.assertEqual(len(result["degradations"]), 1)
|
||||
|
||||
def test_budget_exhaustion_reports_unfinished_not_success(self):
|
||||
# E-08：耗尽即停止并列未完成，不伪造完成
|
||||
item = policy.budget_exhausted()
|
||||
self.assertEqual(item["code"], "RUN_BUDGET_EXCEEDED")
|
||||
self.assertFalse(item["retryable"])
|
||||
self.assertIn("未完成项", item["message"])
|
||||
|
||||
|
||||
class ProbabilityGateTests(unittest.TestCase):
|
||||
def test_probability_is_always_null_with_a_reason(self):
|
||||
for findings in ([], [finding("c1", "pass")], [finding("c1", "issue")]):
|
||||
with self.subTest(findings=len(findings)):
|
||||
result = policy.decide(findings)
|
||||
self.assertIsNone(result["probability"]["value"])
|
||||
self.assertEqual(result["probability"]["reason"], "not_calibrated")
|
||||
|
||||
def test_probability_takes_no_input_so_a_model_score_cannot_leak_in(self):
|
||||
with self.assertRaises(TypeError):
|
||||
policy.probability(0.93)
|
||||
|
||||
def test_native_metric_named_probability_is_rejected(self):
|
||||
# BR-12 / AC-18：厂商原生指标不得被当作审批概率
|
||||
result = policy.decide(
|
||||
[finding("c1", "pass")],
|
||||
judgments=[{"question_id": "q1", "status": "completed",
|
||||
"native_metrics": [{"name": "probability", "value": 0.91}]}])
|
||||
with self.assertRaises(ValueError) as ctx:
|
||||
policy.assert_no_probability_leak(result)
|
||||
self.assertIn("不得被当作审批概率", str(ctx.exception))
|
||||
|
||||
def test_self_reported_confidence_does_not_become_probability(self):
|
||||
result = policy.decide(
|
||||
[finding("c1", "pass")],
|
||||
judgments=[{"question_id": "q1", "status": "completed",
|
||||
"native_metrics": [{"name": "confidence", "value": 0.99,
|
||||
"source": "self_reported"}]}])
|
||||
policy.assert_no_probability_leak(result) # 可以入档，但不得变概率
|
||||
self.assertIsNone(result["probability"]["value"])
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
unittest.main()
|
||||
Reference in New Issue
Block a user