Evidence‑Gated Agency: building an agent that can’t be optimistic

Описание и применение

Автоматический агент, проверяющий государственные заявки на финансирование и сопоставляющий их с данными муниципалитета, предотвращает ошибки и повышает эффективность получения средств.

Задача

Municipalities in Colombia lose public funding because staff cannot monitor procurement portals and match funding calls to their project data in a timely way.

Результат

The agent reliably identifies eligible funding calls, prevents hallucinated or incorrect filings, and provides transparent, actionable reports to municipal staff, improving funding capture and maintaining credibility.

Рабочий процесс

• 1. Agent schedules periodic checks of official procurement portals.
• 2. For each new funding call, agent extracts requirement text and matches it against municipality’s structured project data.
• 3. Agent requests verbatim quotes from the source document for any numeric or date information.
• 4. Code verifies quotes literally exist in the source and parses numbers only from verified text.
• 5. Four gates are applied: requirement evaluable, quote verified, attribute has real value, and a subtractive adversary model pass that can only refute.
• 6. If all gates pass, agent marks requirement as met; otherwise it records an honest zero with missing pieces and owners.
• 7. Agent drafts filing and messages planning secretary overnight for human review.
• 8. Human reviews, decides, and files the application.

Доказательства

The model reads a source document and returns a verbatim quote — the exact span of text that declares the amount.

Источник

Four gates before any "requirement met"

Источник

The honest zero is a first‑class output

Источник