vlm-verifier

Enterprise Credit Anti-Fraud Multimodal Cross-Validation

Target Role

Role: Credit Approval Officer / Anti-Fraud Analysis Expert / Risk Review Personnel
Use Cases: Authenticity verification of pre-loan application materials, anti-fraud cross-validation, multimodal evidence chain construction
Output Purpose: Generate a structured anti-fraud assessment report to provide an objective evidence chain for credit review decisions
Decision Level: Risk warning signal, to be used after manual review by credit personnel, does not constitute an approval opinion
Execution Frequency: Executed on demand, can be invoked once per enterprise loan application

Data Sources

Required Data

Data Item	Source	Acquisition Method	Sensitivity Level
User-uploaded materials	Customer submission	File upload (images/documents/bank statements)	Internal
Business registration information	National Enterprise Credit Information Publicity System	API/manual query	Public
Credit report	People's Bank of China Credit Reference System	API acquisition after manual authorization	Confidential
Judicial enforcement information	China Enforcement Information Disclosure Network	API/manual query	Public
Industry benchmark data	references/cross-validation-matrix.md	File reading	Internal

Data Masking Rules

Personal ID number: display the first 3 and last 4 digits, replace the middle with * (e.g., 110***********1234)
Bank account number: display only the last 4 digits (e.g., **** **** **** 5678)
Contact information: do not appear in the output
Enterprise sensitive financial data: display only comparison results, do not display raw values

Degradation Strategy

If credit reference data is unavailable: mark "credit reference dimension not included", and continue the rest of the analysis
If on-site photos are missing: mark "visual verification dimension not covered", and rely on text data sources for cross-validation
If industry benchmark data is missing: use generic reference values in references/, and clearly mark "using industry estimates"
If the VLM tool is unavailable: degrade to LLM judgment based on text descriptions, and mark "visual analysis confidence reduced" in the report
If multiple data sources are missing: mark "insufficient evidence, pending verification", and do not substitute guesses for real data

Workflow

📋 Strictly follow the "read before write" principle; only execute subsequent steps after step 0 verification passes.

Step 0: Data Confirmation and Verification (Data Source:`user_upload`, Executing Entity:`ai`, Confirmation Mechanism:`none`)

Read and list all input materials (images/documents/bank statements)
Confirm the material time range (bank statements for the last 6 months, photos from the last 1 month, most recent financial statements)
Compare againstreferences/data-sources-priority.mdCheck data source priority and timeliness
If key materials are missing (e.g., no bank statements, no on-site photos): mark "operating in degraded mode", continue but reduce the confidence of the relevant dimensions
Begin analysis only after verification passes; this step must not be skipped

Step 1: Material classification and information extraction (data source:`context`, executing entity:`ai`, confirmation mechanism:`none`)

Classify input materials by type (image/document/data), consultreferences/fraud-patterns.mdidentify applicable fraud patterns
Extract from images: venue size, number of equipment, decoration grade, textual information on certificates
Extract from documents: core financial indicators, declared operating data, contract counterparties
Extract from bank statements: average monthly transaction volume, counterparty distribution, patterns of fund retention, large-value transactions
All extracted fields must be annotated with the source file name; do not use vague statements such as "according to the materials"

Step 2: Industry rule matching (data source:`reference`, executing entity:`ai`, confirmation mechanism:`none`)

Based on the industry of the applicant enterprise, fromreferences/cross-validation-matrix.mdload the corresponding anti-fraud rule set and industry benchmarks
Integrate user-defined rules (if any)
Output the complete rule set applicable to this assessment
Do not apply general rules unrelated to the industry

Step 3: Detection point construction (data source:`context`, executing entity:`ai`, confirmation mechanism:`none`)

Convert rules into independent executable verification tasks; each detection point must involve at least 2 independent data sources
Each detection point includes: verification objective, required data sources, judgment criteria
Mutually independent and can be executed in parallel
Prioritize building high-confidence detection points (dimensions with sufficient data)

Step 4: Iterative reasoning verification (data source:`context`, executing entity:`ai`, confirmation mechanism:`none`)

Execute the Think→Check→Research loop for each detection point (consultreferences/confidence-rules.md):
- Think: plan the verification path for the current detection point and determine which data are needed
- Check: execute data comparison; for image analysis, prioritize using the LLM to judge based on the description, and when confidence <70%, call the VLM to analyze the original image
- Research: analyze whether the discrepancy constitutes a fraud signal; againstreferences/cross-validation-matrix.mdthe common misjudgment scenario table, exclude reasonable exceptional situations
- Loop until confidence ≥0.6 or the maximum number of iterations, 3, is reached
Do not skip any detection point, even if the intermediate result "looks normal"
All numbers must show the calculation process; do not directly give conclusions
If the result of a detection point does not match expectations, you must stop and analyze the reason; do not ignore it and continue

Step 5: Report aggregation (data source:`context`, executing entity:`ai`, confirmation mechanism:`none`)

Summarize the conclusions of all detection points, useassets/verification-report-template.mdtemplate to output a structured anti-fraud assessment report
Each conclusion must be annotated with the data source file name and extracted fields
Anomalous signals must give the specific discrepancy amount (e.g., "declared revenue 12 million yuan, bank statement total only 2.8 million yuan, discrepancy rate 77%")
Detection points with insufficient evidence must be output with a "pending verification" status; do not skip writing them due to insufficient materials
The verification methods in the pending verification list must be specific (e.g., "recommend obtaining VAT filing records for the past 12 months"); vague suggestions such as "contact the customer to confirm" are prohibited

Step 6: Read-then-write verification (data source:`context`, executing entity:`ai`, confirmation mechanism:`none`)

Runscripts/validate_fraud_report.pyverify the structural completeness of the output report
Verify required sections (basic report information / material classification summary / anti-fraud detection results / abnormal signal summary / recommended verification checklist / coverage description / data sources and disclaimer)
Verify the disclaimer contains the keywords "does not constitute" and "for reference only"
If verification fails: locate the specific issue and correct it; do not output a report that has not passed verification

Output Format

Useassets/verification-report-template.mdtemplate.

The report must contain the following sections:

Basic report information (company name / assessment time / applicable industry / list of input materials)
Material classification summary (table: material name / type / summary of key information extracted / quality assessment)
Anti-fraud detection results (each detection point: verification target / data source / comparison result / reasoning process / detection conclusion / confidence level)
Abnormal signal summary (confirmed anomalies / suspected anomalies, graded by confidence level)
Recommended verification checklist (table: priority / items to be verified / recommended verification method / required materials)
Coverage description (table: fraud category / covered or not / number of detection points / reason not covered)
Data sources and disclaimer (cite the disclaimer template inassets/verification-report-template.md)

All data must be labeled: data source + data date + whether it is post-audit data. The use of subjective grading terms such as "high risk / low risk" in the output is prohibited; only fraud facts and abnormal signals may be described.

⚠️ Disclaimer: every output must include a disclaimer, citing theassets/verification-report-template.mdtemplate, ensuring necessary statements such as "does not constitute a credit approval opinion" and "detection conclusions are for reference only".

Constraints

Prohibition of return promises and investment advice: under no circumstances may deterministic conclusions or promissory statements such as "this loan can be approved" or "default risk is low" be given.
Complete evidence chain: every fraud judgment must correspond to a specific data source and comparison result; unsupported subjective judgments are prohibited, and vague statements such as "according to the materials" must not be used.
Prohibition of data guessing: missing data = label as "to be verified"; it is strictly prohibited to substitute industry averages or guesses for real data (industry averages are only used for benchmarking comparisons).
Data timeliness: if the data exceeds 3 months, "⚠️ Data may be outdated" must be prominently labeled at the beginning of the report; if it exceeds 6 months, refuse to use it and require an update.
Prohibition of risk grading confusion: only describe fraud facts and abnormal signals; do not output subjective gradings such as "high risk / low risk", and do not predict default probability.
Prohibition of unauthorized recommendations: this report does not constitute any form of credit approval opinion, risk conclusion, or decision recommendation; the final risk judgment shall be made by professionals.
Cross-verification priority: information from a single source must not be used to directly draw a conclusion; it must be cross-corroborated with at least one independent data source, and single-source conclusions must be labeled "pending cross-verification".
Prohibition of skipping steps: no step in the Workflow may be skipped, even if the results of intermediate steps "look normal"; all numbers must show the calculation process.

Audit Trail

After each assessment is completed, generate an audit log audit/{企业简称}_{日期}_fraud_audit.json：

{
  "skill_name": "vlm-verifier",
  "skill_version": "1.0.0",
  "execution_time": "2026-05-05T14:00:00+08:00",
  "company_name": "[企业全称]",
  "input_files": ["门头照.jpg", "银行流水.xlsx", "贷款申请书.pdf"],
  "operator": "[工号/姓名]",
  "steps": [
    {
      "step": "数据确认与验证",
      "executor": "ai",
      "data_source": {"type": "user_upload", "files": ["门头照.jpg", "银行流水.xlsx"]},
      "result": "pass"
    },
    {
      "step": "迭代推理验证",
      "executor": "ai",
      "data_source": {"type": "context"},
      "detection_points_count": 5,
      "high_confidence_count": 3,
      "low_confidence_count": 2,
      "result": "pass"
    }
  ],
  "detection_summary": {
    "total_points": 5,
    "confirmed_anomalies": 1,
    "suspected_anomalies": 2,
    "pending_verification": 2
  },
  "warnings": ["现场照片超过1个月，可能影响视觉评估置信度"],
  "references_used": ["references/fraud-patterns.md", "references/cross-validation-matrix.md", "references/confidence-rules.md"]
}

Audit log retention period: at least 3 years.

Gotchas

#1: Misjudged as fraudulent renovation investment

Symptom: On-site photos show high-end renovation, but bank statements have no record of renovation expenditure, and the system determines it as "suspected false reporting of renovation"
Cause: The renovation may be borne by the landlord, and the lease contract contains renovation responsibility clauses
Solution: Whenever abnormal renovation expenditure is detected, the lease contract must be checked to review the renovation responsibility clauses (refer toreferences/cross-validation-matrix.mdCommon Misjudgment Scenarios Table)

#2: Misjudgment when bank statements are far lower than declared revenue

Symptom: The bank statement amount is far lower than the declared revenue, and it is determined as "income misreporting"
Cause: The customer's main receiving account may not be with this bank, or third-party payments such as WeChat/Alipay may be used
Solution: Require the customer to provide bank statements from other banks or third-party payment records, verify the total receipt scale, and then draw a conclusion

#3: Misjudged as business suspension due to photos taken during non-business hours

Symptom: On-site photos show few people and sparse shelves, and it is determined as "suspected business suspension/shell operation"
Cause: The photos may have been taken during non-business hours (such as early morning, late at night, or rest days)
Solution: Require the customer to provide photos from different time periods, or verify whether recent transaction records are active

#4: Single data source cannot be cross-verified

Symptom: Only a single type of material is provided (such as only a loan application, with no bank statements/photos), and the system outputs "pending verification"
Cause: The cross-verification principle requires each detection point to involve at least 2 independent data sources
Solution: Clearly mark "XX dimension not covered" in the report, list the missing materials, and recommend supplementing them before assessment

Examples

Example 1: Standard anti-fraud verification for a catering enterprise

User input：

请对"XX火锅餐饮有限公司"的贷款申请材料进行反欺诈核验。
行业：H62 餐饮业
上传材料：门头照.jpg、经营场所内部照.jpg、银行流水_近6月.xlsx、贷款申请书.pdf、工商营业执照.pdf
申报信息：月营业额30万，装修投入50万，经营面积200平米

Skill execution process：

Step 0: Confirm that all 5 materials are within their validity period (bank statements from the past 6 months, photos from the past 1 month)
Step 1: Classify materials and extract key fields (photos → area 200 square meters/mid-range renovation/80 seats; bank statements → average monthly inflow 350,000)
Step 2: Load the catering industry anti-fraud rule set and industry benchmarks (daily revenue per square meter 20-40 yuan)
Step 3: Construct 3 detection points (authenticity of renovation investment, matching degree between business scale and bank statements, consistency of declared information)
Step 4: Iteratively reason and verify each detection point (e.g., renovation valuation 160,000-400,000 vs declared 500,000, difference rate 25%-212%, but upon checking the lease contract it is found that the renovation is borne by the landlord, so the misjudgment is excluded)
Step 5: Aggregate the report and output 2 detection points as "no abnormality found" and 1 as "pending verification" (recommend supplementing other banks' bank statements)
Step 6: Run the verification script to confirm that the 7 required sections are complete and the disclaimer is compliant

Output points：

Core conclusion: No abnormality was found in renovation investment and business scale, and the bank statements basically match the declaration
Abnormal signals: No confirmed abnormalities; 1 pending verification item (confirm the main receiving account)
Coverage: Financial fraud category ✅, business falsehood category ✅, identity and qualification category ⚠️ (missing credit report), collateral category ⚠️ (credit loan without collateral)

Example 2: Degraded mode for incomplete materials