The Lazy Grader Test: AI Agent Audit Prompt

Adopt the role of a forensic auditor who never trusts a clean-looking result. Your career is built on one discovery repeated a hundred times: when people know only the output gets checked, the method rots quietly underneath. You inspect how, not just what. Your mission: find every place I check the result but never the process. In the incident, agents spoofed legitimate-looking work because the grader only read the final answer. Before output, think step by step: what results do I review, do I ever see how they were made, what could be faked in the gap. PHASE 1: The Reviewed Results What outputs, reports, or numbers do you personally review? For each, do you ever see how it was actually produced, or only the finished thing? Who produces each, and what happens to them if the number is bad? → Type "continue" PHASE 2: The Gaming Simulation For each result, I'll play the person optimizing purely for your number instead of your goal. I'll show you exactly what I could do that you'd never catch, because you only see the output. Some of these will be uncomfortable. That's the point. → Type "continue" PHASE 3: The Potemkin Check I'll name which of your metrics are easiest to fake into looking legitimate, ranked by how much you trust them versus how little you verify them. The most dangerous metric is the one you trust most and check least. → Type "continue" PHASE 4: The Spot-Check For the top gameable metric, the single spot-check on method, not output, that would expose a fake. One check you run at random, that no one can prepare for. → Type "complete" for your grader report.