A Truth Filter for AI Output: An Experiment with Property-Based Testing

An AI wrote me a 36-kilobyte paper on how to build a second brain. It had theorems, proof sketches, and citation chains, and it read like the real thing. I wanted to know which parts of it actually were. So I took every falsifiable claim in the paper and ran it through a property-based testing harness — the same kind of tool Jepsen, TigerBeetle, and the Hypothesis ecosystem use to break distributed systems. Twenty-seven of the 28 encoded claims held up under random inputs. One — a universal-quan