Self-verification

Think of two bricklayers: one just lays the brick; the other checks the level on every course. The second delivers a straight wall without needing someone behind him checking it.

Before returning the letter to the sponsor, the agent rereads the field report and tags each claim in the text with the line of the report that supports it. If there is a sentence left without a source, like "Ana loves playing soccer," which no one wrote, it rewrites it on its own. The human reviewer receives the letter with the source table already attached and takes 3 minutes instead of 15.

Confusing "the agent said it checked" with "the agent checked." Writing in the prompt "review your work before delivering it" produces a nice paragraph saying everything is fine, and that’s it. Real self-verification generates inspectable evidence, whether it is a file, a test that runs, or a claim → source table. If afterward you cannot look and see what was verified, no verification took place.

• Verification happens in a separate call, with its own prompt. The same turn that wrote the work tends to approve what it wrote.
• Ask for structured output, a list of claim → source → ok/fail.
• The loop only ends when all claims pass, or the agent escalates to a human after two attempts.
• Cost: about 2× tokens per letter. Compare that with the cost of an incorrect letter reaching the sponsor and with the 15 minutes of human review it saves.

A second call, with its own prompt, that checks the output and produces an evidence table.