Testing AI Agents with a Rigged Ledger

Speedydooo · reddit · 2026-07-15

The author spent a month testing their AI agents using a "fake ledger": creating realistic-looking invoices, payments, and churn curves, then deliberately planting 6 errors, including duplicate invoices, outdated rates, missing approval thresholds, seemingly normal but unprofitable customer segments, revenue leaks only visible in churn months, and an unapproved payment.

As a result, the agents caught 5 out of 6 errors unprompted. The missed error wasn't a math issue but a missing sign-off rule in the author's scoring logic. Furthermore, two agents actually proactively pointed out this scoring gap.

More importantly, they demonstrated excellent "refusal" behaviors:

The author also found two mistakes in their own answer key: a math typo and an incorrect business rule, both discovered during reviews prompted by discrepancies between the agents' outputs and the expected answers.

The entire experiment cost about $5.5 and took less than 30 minutes. The author concludes that knowing when to say "insufficient information" is far more valuable than simply getting the right answer.

Original post →

More from coding & agent

coding & agent channel →