feat(v2): T11 改进治理与演示交付——对照门禁、泄漏作废、三模块演示脚本

覆盖 AC-15/16/34/37/41（BR-15/16/40、E-11/24）。后端： - contracts/improvement.py：ImprovementCandidate + ContrastReport/ContrastMetrics。零分母不得给比率或覆盖率（样本不足不是 0%）；非独立评测不得给改进/退化结论；已发布必须引用既有不可变版本；没有对照证据不得进入待审核或已发布。 - services/improvements.py： · evaluate_contrast 在服务端判定泄漏——预测锁定前读到标签即整次评测作废（E-11），保留集来源进开发集的样本作废并单列，指标按分子/可观察分母/覆盖率/未观察分开报； · transition 状态机：draft→evaluating→ready→in_review→approved→released，自批拒绝、退化与结论不明拒绝发布、发布幂等且只引用既有版本、撤回只影响新运行。 - api/prescreen.py：候选列表、提出与状态机三个端点；评测请求体提交样本，泄漏与指标一律由服务端计算，不接受客户端交来的"成品报告"。 - contracts/prescreen.py：补 E-11/E-24 错误码与 409 映射。前端一致性修正： - src/lib/prescreen-run.ts 的呈现判定改用后端既有词汇（empty/list/grouped/radar/faceted，与 services/review_plan.presentation 一致），避免前后端各一套判定导致界面画出后端判定为不该画雷达的图。演示交付： - scripts/xingbang_prescreen_demo.py：按 08 §7 八步演示，只驱动已实现入口，未实现的步骤（批量清单、历史提炼、卡片治理）如实标注原因；工程演示与模型效果分列，本地专项/Jev/真实概率一律标未验证或未接通。 - backend/test_demo_script.py：冒烟断言八步状态、未实现原因与分列口径。验证：backend 全量 unittest 519 项 OK；前端 npm run build 通过。 Co-Authored-By: Claude Code <noreply@anthropic.com>
2026-09-30 09:25:25 +08:00
parent f2bbac03e7
commit d6bc8485f5
10 changed files with 1018 additions and 14 deletions
@@ -43,10 +43,11 @@ assert.match(lib, /domain !== "demo"/, "lib 必须在非演示域返回 null");
 // 5. 雷达只对同刻度 3–8 维绘制，其余给完整列表；空维度不补占位轴（AC-44）
 assert.match(lib, /RADAR_MIN = 3/);
 assert.match(lib, /RADAR_MAX = 8/);
-assert.match(lib, /kind: "radar"/);
+assert.match(lib, /layout: "radar"/);
+assert.match(lib, /DimensionLayout = "empty" \| "list" \| "grouped" \| "radar" \| "faceted"/, "呈现词汇必须与后端 review_plan.presentation 一致");
 assert.match(lib, /EMPTY_DIMENSIONS = "当前方案尚未配置核查维度"/);
-assert.match(panel, /shown\.kind === "radar" && <RadarChart/, "只有判定为雷达时才画雷达");
-assert.match(panel, /shown\.kind === "empty" && <div className="ra-empty">/, "空维度用空状态，不补六轴");
+assert.match(panel, /shown\.layout === "radar" && <RadarChart/, "只有判定为雷达时才画雷达");
+assert.match(panel, /shown\.layout === "empty" && <div className="ra-empty">/, "空维度用空状态，不补六轴");
 assert.match(panel, /文字列表与图等价/, "图必须有等价文字列表");

 // 6. 乱序与过期响应必须丢弃（AC-31）
@@ -0,0 +1,218 @@
+"""兴邦预审 P0 模拟闭环演示脚本（T11）。
+
+依据：docs/v2/08-data-pilot-and-demo.md §7 的八步演示顺序。
+
+**诚实性要求**（08 §7 末、06 §6）：脚本只驱动本仓库已经实现并验证过的入口；
+没有实现的步骤打印"未实现"及其原因，不伪造屏幕状态；模型效果与工程演示
+分列，模拟标签不得包装成真实业务验证。
+
+用法：backend/venv/Scripts/python.exe scripts/xingbang_prescreen_demo.py
+全程在临时目录内使用临时数据库，不触碰真实 xingbang_data。
+"""
+
+from __future__ import annotations
+
+import sys
+import tempfile
+from pathlib import Path
+
+ROOT = Path(__file__).resolve().parent.parent
+sys.path.insert(0, str(ROOT / "backend"))
+
+from contracts.prescreen import ErrorCode  # noqa: E402
+from services import (assistant_store as store, improvements, judgment_adapters,  # noqa: E402
+                      judgment_provider as jp, prescreen, prescreen_policy,
+                      prescreen_sandbox, prescreen_store, review_plan, trajectory)
+
+DONE = "已执行"
+NOT_DONE = "未实现"
+MODEL_UNVERIFIED = "未验证（无权重与授权）"
+
+FINDINGS = [
+    {"code": "corporate.cashflow", "result": "unknown", "stage": 1, "evidence_ids": ["EV-1"]},
+    {"code": "corporate.title", "result": "unknown", "stage": 1, "evidence_ids": ["EV-2"]},
+    {"code": "corporate.deal", "result": "checked", "stage": 2, "evidence_ids": ["EV-3"]},
+]
+DIMENSIONS = [
+    {"id": "DC-CASH", "component_refs": [{"id": "corporate.cashflow"}]},
+    {"id": "DC-TITLE", "component_refs": [{"id": "corporate.title"}]},
+    {"id": "DC-DEAL", "component_refs": [{"id": "corporate.deal"}]},
+]
+PLAN = {
+    "plan": {"id": "corporate.default", "version": "1.0.0"},
+    "skills": [{"id": "corporate.cashflow", "version": "1.0.0"}],
+    "rules": [], "template": {"id": "template.ai-pre-review", "version": "1.0.0"},
+    "content_hash": "demo-hash", "dimensions": DIMENSIONS,
+}
+STEPS = [
+    (1, "3,000 条候选清单、生成版本与多结果分布", NOT_DONE,
+     "T04 的服务已在单测中验证，但本脚本不重复驱动批量生成；清单需真实数据授权"),
+    (2, "选开发集案例，展开问询/补件/会议事件，提炼 Skill 与 advisory Rule", NOT_DONE,
+     "历史提炼走 T05 的治理入口，需要真实历史材料；此处不伪造提炼结果"),
+    (3, "自然语言改适用范围 + 人工新增卡片，统一审核发布", NOT_DONE,
+     "卡片治理入口属 T05；演示脚本不代替审核动作"),
+    (4, "数据卡 + Skills + Rules 组方案，维度随方案变化，Tools 只在依据中展示", DONE, None),
+    (5, "缺证 → 主审核查 → 问题与补件 → 演示概率来源", DONE, None),
+    (6, "勾选动作比较情景、撤销恢复基线", DONE, None),
+    (7, "回放完整 trajectory，对齐模拟会议记录，查看材料版本差异", DONE, None),
+    (8, "从漏检提出改进候选，隔离样本对照后人工审核发布，旧版仍可回放", DONE, None),
+]
+
+
+def run_demo() -> dict:
+    facts: dict = {"steps": [], "sandbox": {}, "trajectory": {}, "improvement": {}}
+    with tempfile.TemporaryDirectory(ignore_cleanup_errors=True) as tmp:
+        db_path = Path(tmp) / "demo.sqlite3"
+        store.initialize(db_path)
+
+        # 步骤 4：方案维度（去重与刻度判定）
+        dimensions = review_plan.resolve_dimensions([
+            {"id": "DC-CASH", "label": "回款核验", "scale_id": "evidence-stage",
+             "scale_version": "1.0.0", "order": 1, "component_refs": [{"id": "corporate.cashflow", "version": "1.0.0"}],
+             "levels": [{"value": i, "label": name, "criterion": f"证据阶段：{name}"} for i, name in
+                        enumerate(("未提供", "仅有说明", "资料待核", "交叉核验", "人工复核"))]},
+            {"id": "DC-TITLE", "label": "权属核验", "scale_id": "evidence-stage",
+             "scale_version": "1.0.0", "order": 2, "component_refs": [{"id": "corporate.title", "version": "1.0.0"}],
+             "levels": [{"value": i, "label": name, "criterion": f"证据阶段：{name}"} for i, name in
+                        enumerate(("未提供", "仅有说明", "资料待核", "交叉核验", "人工复核"))]},
+            {"id": "DC-DEAL", "label": "交易要素", "scale_id": "evidence-stage",
+             "scale_version": "1.0.0", "order": 3, "component_refs": [{"id": "corporate.deal", "version": "1.0.0"}],
+             "levels": [{"value": i, "label": name, "criterion": f"证据阶段：{name}"} for i, name in
+                        enumerate(("未提供", "仅有说明", "资料待核", "交叉核验", "人工复核"))]},
+        ])
+        presentation = review_plan.presentation(dimensions)
+        facts["presentation"] = presentation["layout"]
+
+        # 步骤 5：预审执行（专项关闭，无校准概率）
+        config = jp.resolve_config(False)
+        decision = prescreen_policy.decide(FINDINGS)
+        decision["findings"] = FINDINGS
+        decision["missing_inputs"] = ["近六个月银行流水", "设备权属材料"]
+        screening = prescreen.assemble(
+            {"id": "run-demo", "project_id": "DEMO-026", "revision": 1,
+             "created_at": store.now(), "input_snapshot_id": "snap-demo"},
+            PLAN, decision, config)
+        with store.transaction(db_path) as db:
+            prescreen_store.save(db, screening)
+        facts["recommendation"] = screening["recommendation"]
+        facts["probability_is_null"] = screening["probability"] is None
+
+        # 步骤 6：沙盘（演示值只在演示域、组合查表）
+        with store.transaction(db_path) as db:
+            stored = prescreen_store.load(db, "DEMO-026", "run-demo")
+            baseline = prescreen_sandbox.simulate(stored, [])
+            both = prescreen_sandbox.simulate(stored, ["assume_bank_flow", "assume_title"])
+        facts["sandbox"] = {
+            "baseline": baseline["probability_view"]["demo_estimate"]["value"],
+            "combined": both["probability_view"]["demo_estimate"]["value"],
+            "label": baseline["probability_view"]["demo_estimate"]["label"],
+            "unchanged_dimensions": [item["id"] for item in both["dimensions"]
+                                     if not item["assumed"]],
+        }
+
+        # 步骤 7：轨迹回放与模拟会议结果
+        with store.transaction(db_path) as db:
+            store.put(db, "run", {"id": "run-demo", "project_id": "DEMO-026",
+                                  "status": "completed", "input_snapshot_id": "snap-demo",
+                                  "created_at": store.now(), "steps": [
+                                      {"run_id": "run-demo", "step_id": "corporate.cashflow",
+                                       "skill": {"id": "corporate.cashflow", "version": "1.0.0"},
+                                       "status": "completed", "input_snapshot_id": "snap-demo",
+                                       "started_at": store.now(), "outputs": {"note": "已核查"},
+                                       "findings": [{"code": "cashflow", "result": "unknown",
+                                                     "evidence_ids": ["EV-1"]}],
+                                       "errors": [], "missing_inputs": []},
+                                      {"run_id": "run-demo", "step_id": "corporate.title",
+                                       "skill": {"id": "corporate.title", "version": "1.0.0"},
+                                       "status": "failed", "input_snapshot_id": "snap-demo",
+                                       "started_at": store.now(), "outputs": {},
+                                       "findings": [], "missing_inputs": [],
+                                       "errors": [{"code": "SKILL_TIMEOUT",
+                                                   "message": "演示中的失败步骤"}]}]})
+            replay = trajectory.replay(db, "DEMO-026", "run-demo")
+            trajectory.record(db, "DEMO-026",
+                              {"raw_code": "附条件通过", "meeting_id": "M-DEMO",
+                               "occurred_at": store.now()}, "demo-reviewer",
+                              known_snapshots=["snap-demo"])
+            outcomes = trajectory.listing(db, "DEMO-026")
+        facts["trajectory"] = {
+            "steps": [item["status"] for item in replay["steps"]],
+            "failed_preserved": any(item["status"] == "failed" for item in replay["steps"]),
+            "unmatched": len(outcomes["unmatched"]),
+            "observed_denominator": outcomes["counts_toward_denominator"]["observed"],
+        }
+
+        # 步骤 8：改进候选（合成样本、独立评测、可回放旧版）
+        with store.transaction(db_path) as db:
+            candidate = improvements.propose(db, "DEMO-026", {
+                "target": "knowledge", "target_ref": {"id": "card-demo", "version": "1.0.0"},
+                "proposed_revision_ref": {"id": "card-demo", "version": "1.1.0"},
+                "change_summary": "把回款核验经验扩到设备直租",
+                "source_run_ids": ["run-demo"]}, "demo-manager", "demo")
+            evaluation = {"baseline_ref": {"id": "current", "version": "1.0.0"},
+                          "candidate_ref": {"id": "prescreen.v2", "version": "0.2.0"},
+                          "dataset_ref": "demo/synthetic-manifest.json",
+                          "split_ref": "demo-split.v1", "independent": True,
+                          "samples": [
+                              {"id": "S-1", "split": "dev", "observed": True,
+                               "baseline_pass": True, "candidate_pass": True,
+                               "predicted_at": "2026-09-29T01:00:00+00:00",
+                               "label_revealed_at": "2026-09-29T03:00:00+00:00"},
+                              {"id": "S-2", "split": "dev", "observed": True,
+                               "baseline_pass": False, "candidate_pass": True,
+                               "predicted_at": "2026-09-29T01:00:00+00:00",
+                               "label_revealed_at": "2026-09-29T03:00:00+00:00"}]}
+            evaluated = improvements.transition(db, candidate["id"], "evaluate", "demo-reviewer",
+                                                evaluation=evaluation)
+            improvements.transition(db, evaluated["id"], "submit", "demo-reviewer")
+            approved = improvements.transition(db, evaluated["id"], "approve", "demo-reviewer")
+            released = improvements.transition(
+                db, approved["id"], "release", "demo-reviewer",
+                publisher=lambda body: {"id": "card-demo", "version": "1.1.0"})
+        facts["improvement"] = {"verdict": evaluated["contrast"]["verdict"],
+                                "status": released["status"],
+                                "released_ref": released["released_ref"]}
+
+        # 专项：离线 fixture 可跑，live 未接通
+        facts["judgment_offline"] = judgment_adapters.load_fixture(
+            "question_bank_extraction.json")["rubric_version"]
+        facts["demo_table_version"] = prescreen_sandbox.DEMO_TABLE_VERSION
+        facts["budget_code"] = ErrorCode.JUDGMENT_CONTEXT_EXCEEDED.value
+    return facts
+
+
+def main() -> int:
+    facts = run_demo()
+    print("兴邦预审 P0 模拟闭环演示（合成数据 · 不连接真实数据与模型）")
+    print("=" * 78)
+    for index, title, status, reason in STEPS:
+        print(f"{index}. [{status}] {title}")
+        if reason:
+            print(f"   原因：{reason}")
+    print("-" * 78)
+    print("工程演示（本仓库已实现并验证的入口）")
+    print(f"  · 维度呈现：{facts['presentation']}")
+    print(f"  · 主审意见：{facts['recommendation']} · 真实概率为空：{facts['probability_is_null']}")
+    sandbox = facts["sandbox"]
+    print(f"  · 沙盘：基线 {sandbox['baseline']:.0%} → 组合 {sandbox['combined']:.0%}"
+          f"（{sandbox['label']}）")
+    print(f"  · 未变化的维度仍列出：{sandbox['unchanged_dimensions']}")
+    trace = facts["trajectory"]
+    print(f"  · 轨迹步骤状态：{trace['steps']} · 失败步骤保留：{trace['failed_preserved']}")
+    print(f"  · 会议结果：unmatched {trace['unmatched']} 条 · 已观察分母 {trace['observed_denominator']}")
+    imp = facts["improvement"]
+    print(f"  · 改进对照：{imp['verdict']} → {imp['status']}（{imp['released_ref']}）")
+    print("-" * 78)
+    print("模型效果（与上面的工程演示分开计）")
+    print(f"  · 本地专项：{MODEL_UNVERIFIED}；当前只有离线 fixture（{facts['judgment_offline']}）")
+    print(f"  · Jev：未接通（无端点与凭据）")
+    print(f"  · 真实审批概率：未校准，恒为空；界面上的百分数来自演示表 "
+          f"{facts['demo_table_version']}，不是模型输出")
+    print(f"  · 超预算时的拒绝码（{facts['budget_code']}）证明预算守卫生效，与概率无关")
+    print("=" * 78)
+    print("以上均为合成数据与演示值，不能作为真实业务效果或审批依据。")
+    return 0
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())