Two AI readers agreed—and both were wrong

说明与用途

在 TransferDue 的工作流中，单次 AI 读取泰国房地产文件往往会误读数字。加入第二次读取和与 Gemini 的跨家族检查将未标记错误减少了 38.6%，但警告数量增加，使错误更易被发现，但也增加了人工审核工作。

问题

A single AI reading of Thai property documents can misinterpret digits, producing wrong but plausible values that go unchecked.

结果

The new rule set reduced unflagged wrong answers by 22 (from 57 to 35) and increased warnings by 30, making errors more visible but adding manual review work. Adding Mistral OCR increased correct answers by 19 but also added 3 more wrong answers without warnings.

工作流

• Run first Claude reader to extract values
• Run second Claude reader independently
• Run Gemini reader from a different model family
• If first two readers agree, skip third Claude read; otherwise request third read
• Apply business rules: if majority vote from same family, accept; if family conflict, flag as warning; if Gemini disagrees with majority, flag as warning; if Mistral OCR available, compare and possibly correct

证据

Wrong answers accepted without a warning fell from 57 to 35—a 38.6% reduction

来源

The new rules produced 22 fewer wrong answers without warnings

来源

The system can now flag the value for checking. In the code, that status is called family_conflict

来源