I run a travel site whose core dataset is legal-status data: whether cannabis is recreational, medical-only, decriminalized, or illegal in 213 countries and all 50 US states. Getting one of those entries wrong isn't a typo. Someone plans a trip around it.
The problem is that this kind of data rots quietly. Laws moved constantly through 2025 and 2026: one US state opened its first dispensaries three weeks before I ran this audit, a Caribbean nation decriminalized this year, and one country famous for legalizing reversed course almost entirely. A dataset that was accurate when written is a dataset that is wrong now — you just don't know where.
So I gave an AI agent (Claude Code, in my case) a blunt instruction: go through every state and every country, compare each stored status against current law, and tell me what's wrong. No edits until I approve.
I expected a list of stale entries. I got that — plus two other categories of wrong I wasn't expecting, and they were both worse. I thought I was auditing facts. It took three kinds of wrong to see I was auditing a system.
Way #1 to be wrong: the facts moved
