Evidence-Gated Agency: building an agent that can't be optimistic

In Colombia, public funds for housing and schools frequently go unspent due to a lack of oversight. A team created an automated system to address this issue. This "Evidence-Gated Agency" agent observes official sources at intervals, cross-references the findings with the municipal project data, compiles the missing information with responsible parties and deadlines, generates a filing, and notifies the planning secretary via message.

The agent is programmed to never assert any figures or dates, only to provide references to the original text. This design prevents the agent from fabricating data, which is crucial for its role in helping small municipalities accurately claim their funds. The system has four key safeguards:

1. The model is explicitly programmed not to produce numerical or date-based outputs. It can only quote the original text.

2. Multiple verification layers, including a second model pass, ensure that any requirement met is accurate and not just fabricated.

3. The system treats a "no results" situation as a legitimate response, providing detailed explanations for why no funding is available.

4. All data in the system is tagged with its source, confidence level, and the date it was retrieved, enabling a clear audit trail.

By enforcing these four principles, the agent ensures that its recommendations are honest and based on verifiable evidence, helping to prevent municipalities from losing out on crucial funding.