first commit

backend/app/rules/assessment_rules.py

@@ -0,0 +1,57 @@
+"""Rule-based fraud amount assessment.
+
+Classifies each transaction into high / medium / low confidence fraud,
+and generates initial reason text.
+"""
+from app.models.transaction import TransactionRecord
+from app.models.assessment import ConfidenceLevel
+
+FRAUD_KEYWORDS = ["投资", "保证金", "手续费", "解冻", "税费", "充值", "提币", "提现"]
+
+
+def classify_transaction(tx: TransactionRecord) -> tuple[ConfidenceLevel, str, str]:
+    """Return (confidence_level, reason, exclude_reason)."""
+    if tx.is_transit:
+        return (
+            ConfidenceLevel.low,
+            f"该笔为本人账户间中转（{tx.source_app.value} -> {tx.counterparty_name}），不直接计入被骗损失。",
+            "本人账户间互转，仅作为资金路径展示。",
+        )
+
+    if tx.direction.value == "in":
+        return (
+            ConfidenceLevel.low,
+            f"该笔为收入方向交易（+¥{float(tx.amount):,.2f}），通常不属于被骗损失。",
+            "收入交易不计入损失。",
+        )
+
+    remark = tx.remark or ""
+    counterparty = tx.counterparty_name or ""
+    confidence = tx.confidence
+
+    has_fraud_keyword = any(kw in remark or kw in counterparty for kw in FRAUD_KEYWORDS)
+
+    if confidence >= 0.9 and has_fraud_keyword:
+        return (
+            ConfidenceLevel.high,
+            f"受害人经{tx.source_app.value}向「{counterparty}」转账¥{float(tx.amount):,.2f}，"
+            f"备注为「{remark}」，与诈骗常见话术吻合，OCR置信度{confidence:.0%}。",
+            "",
+        )
+
+    if confidence >= 0.85:
+        reason = (
+            f"受害人经{tx.source_app.value}向「{counterparty}」转账¥{float(tx.amount):,.2f}。"
+        )
+        if has_fraud_keyword:
+            reason += f"备注「{remark}」含涉诈关键词。"
+            return ConfidenceLevel.high, reason, ""
+        reason += "建议结合笔录确认是否受诱导操作。"
+        return ConfidenceLevel.medium, reason, "如经核实该笔为受害人主动日常消费，应排除。"
+
+    return (
+        ConfidenceLevel.medium,
+        f"受害人经{tx.source_app.value}向「{counterparty}」转账¥{float(tx.amount):,.2f}，"
+        f"OCR置信度较低({confidence:.0%})，需人工复核。",
+        "OCR置信度不足，可能存在识别误差。",
+    )