How I stopped my research agent from inventing its sources

Описание и применение

Автоматизированный агент для исследований часто генерирует утверждения с ссылками, но детали в них могут быть неверными. Автор внедрил скрипт, который скачивает страницы, проверяет, что каждое слово и число в цитате действительно присутствуют на странице, и сравнивает цифры в утверждении с цифрами в цитате. Это позволяет отмечать неверные утверждения как «не проверено» и повышает надёжность отчётов.

Задача

Research agents often generate claims with linked sources that are not accurate, leading to misinformation.

Результат

The agent now reliably flags inaccurate claims, preventing fabricated quotes and mismatched numbers, and clearly indicates which claims are not machine‑checked.

Рабочий процесс

• 1. Researcher searches and fetches pages, outputs a citations file.
• 2. Writer drafts the brief from citations, no web access.
• 3. Fact‑Checker fetches cited pages but cannot search, reads each linked page.
• 4. Fact‑Checker verifies each claim by matching the quoted text against the page using a case‑insensitive, punctuation‑agnostic comparison.
• 5. If a quote is missing or altered, the report refuses to build the claim or marks it as not machine‑checked.
• 6. Added a numbers check to ensure figures in the claim appear in the quote, with tolerance for formatting and rounding.
• 7. Extended support to PDFs and allowed quote fragments in any order.
• 8. Claims that pass all checks are marked verified; otherwise they are marked partial or unverified.

Доказательства

The source often said something slightly different. A different number. A different year. "Experimental" where the brief said "default".

Источник

The Fact‑Checker having no search tool is deliberate. If it could search, it would "verify" a claim by finding some page that agrees.

Источник

If a quote isn’t on the page, the report refuses to build. If a page can’t be downloaded, the claim is marked not machine‑checked instead of quietly passing.

Источник

v1.2 adds a numbers check: every figure in a "verified" claim must also appear in its quote.

Источник