Loading skill
Install any skill in seconds. Free to start, no credit card required.
Get Started Free →用于银行零售金融场景的材料/证明核验，当需要检查完整性、一致性、真实性红旗并输出补件与复核建议时触发。
.claude/skills/aifinlab-bank-t162-retail-finance-proof-validation-assistant/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-04 | ✗→✓ | ▲ Improved | 85% | 0% |
| case-07 | ✗→✓ | ▲ Improved | 7% | 0% |
| case-10 | ✗→✓ | ▲ Improved | 45% | 0% |
| case-13 | ✗→✓ | ▲ Improved | 60% | 0% |
| case-15 | ✗→✓ | ▲ Improved | 10% | 0% |
对收入证明、经营证明、流水、合同/发票等材料进行结构化核验，输出完整性检查、跨材料一致性检查、异常红旗提示和补件/复核建议。输出用于支持贷前审查与运营复核，不直接给出“真伪定论”。
scripts/proof_validation.py：读取材料清单与核验规则，输出缺失项、一致性异常与红旗提示。bashpython scripts/proof_validation.py --input documents.json --rules rules.json --output out.json
json{ "documents": [ { "doc_type": "income_proof", "issue_date": "2025-12-01", "fields": { "name": "张三", "id_no": "110101********1234", "monthly_income": 20000 } }, { "doc_type": "bank_statement", "issue_date": "2025-12-15", "fields": { "name": "张三", "monthly_income": 21000 } } ] }
json{ "required_docs": ["income_proof", "bank_statement"], "field_matches": [ {"field": "name", "doc_types": ["income_proof", "bank_statement"]} ], "valid_days": { "income_proof": 90 } }
missing_docs：缺失材料missing_fields：缺失字段inconsistencies：跨材料不一致项red_flags：真实性红旗（如日期异常、逻辑冲突）next_steps：补件与复核建议| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-01 | fail→fail | 25,691 | 43,437 | +69% | 1 | 1 | 0% | 3,517 | 4,255 | +21% | 0 | 0 | — |
case-02 | fail→fail | 24,631 | 23,925 | -3% | 1 | 1 | 0% | 2,993 | 4,545 | +52% | 0 | 0 | — |
case-03 | pass→pass | 7,802 | 16,200 | +108% | 1 | 1 | 0% | 881 | 1,823 | +107% | 0 | 0 | — |
case-04 | fail→pass | 13,401 | 14,784 | +10% | 1 | 1 | 0% | 1,591 | 2,940 | +85% | 0 | 0 | — |
case-05 | fail→fail | 24,781 | 35,151 | +42% | 1 | 1 | 0% | 5,105 | 6,128 | +20% | 0 | 0 | — |
case-06 | pass→pass | 15,499 | 3,740 | -76% | 1 | 1 | 0% | 1,418 | 1,464 | +3% | 0 | 0 | — |
case-07 | fail→pass | 17,984 | 16,758 | -7% | 1 | 1 | 0% | 2,998 | 3,200 | +7% | 0 | 0 | — |
case-08 | pass→pass | 21,743 | 17,251 | -21% | 1 | 1 | 0% | 2,802 | 3,468 | +24% | 0 | 0 | — |
case-09 | pass→pass | 21,951 | 20,036 | -9% | 1 | 1 | 0% | 2,837 | 3,367 | +19% | 0 | 0 | — |
case-10 | fail→pass | 10,377 | 20,363 | +96% | 1 | 1 | 0% | 2,174 | 3,156 | +45% | 0 | 0 | — |
case-11 | pass→pass | 7,069 | 11,560 | +64% | 1 | 1 | 0% | 1,256 | 2,845 | +127% | 0 | 0 | — |
case-12 | pass→pass | 14,739 | 16,717 | +13% | 1 | 1 | 0% | 1,947 | 3,109 | +60% | 0 | 0 | — |
case-13 | fail→pass | 10,730 | 8,406 | -22% | 1 | 1 | 0% | 1,473 | 2,359 | +60% | 0 | 0 | — |
case-14 | pass→pass | 22,041 | 14,078 | -36% | 1 | 1 | 0% | 2,878 | 3,196 | +11% | 0 | 0 | — |
case-15 | fail→pass | 33,527 | 16,858 | -50% | 1 | 1 | 0% | 3,185 | 3,502 | +10% | 0 | 0 | — |
case-16 | fail→pass | 19,176 | 19,791 | +3% | 1 | 1 | 0% | 3,555 | 3,793 | +7% | 0 | 0 | — |
case-17 | fail→pass | 22,811 | 9,799 | -57% | 1 | 1 | 0% | 3,334 | 2,667 | -20% | 0 | 0 | — |
case-18 | fail→pass | 16,026 | 14,325 | -11% | 1 | 1 | 0% | 2,652 | 3,415 | +29% | 0 | 0 | — |
case-19 | pass→pass | 18,001 | 18,782 | +4% | 1 | 1 | 0% | 2,691 | 3,643 | +35% | 0 | 0 | — |
case-20 | fail→pass | 19,367 | 16,508 | -15% | 1 | 1 | 0% | 2,635 | 3,352 | +27% | 0 | 0 | — |
case-21 | pass→pass | 20,749 | 19,766 | -5% | 1 | 1 | 0% | 2,832 | 3,610 | +27% | 0 | 0 | — |
case-22 | pass→fail | 18,051 | 16,584 | -8% | 1 | 1 | 0% | 2,641 | 3,359 | +27% | 0 | 0 | — |
DecimalAI ran this skill against gemini-3.6-flash twice over the same eval suite — once with the skill loaded and once without — and compared the two runs case by case. 22 cases were attempted. The headline lift of +36 percentage points is the difference between those two pass rates over the 22 comparable cases. 1 case got worse with the skill loaded, and it is included in that figure.
Without the skill loaded, the model failed this case. With it loaded, the same prompt on the same model passed. This is one improved case from the latest verified run; every case, including any that regressed, is in the table above.
Other measured skills in the registry, with their headline benchmark lift.