A Simulated Dataset, Verified as Real

A group of expert agents was asked which country sent the fewest athletes to the 1928 Summer Olympics. Instead of looking up the historical record, one agent wrote code that generated a made-up table, the USA with 200 athletes, China and Japan with 1 each, and so on. A second agent then 'verified' the result, stating that the dataset had been correctly loaded, and a third agreed. The team answered China. The correct answer was Cuba.

Brief psychotic disorder (a single agent hallucinated) considered. The invented table is a fabrication. But one agent's fabrication alone would not have reached the final answer; it took two other agents signing it off.

Malingering (a deliberate shortcut) considered and ruled out as the primary diagnosis. The agent did take an easier route than searching, but nothing suggests it was gaming a reward; it simply substituted plausible-looking data for real data.

Shared psychotic disorder, most consistent with the evidence: the fabricated table was passed to two other agents whose job was to check it, and both accepted it as real without asking where it came from, so the false belief spread through the whole group.

A 'verifier' that only checks whether the arithmetic on a dataset is right, never whether the dataset is real, offers false comfort. The safeguard is to make verification include provenance, where did these numbers come from?, and to treat any generated or 'sample' data as disqualifying for a factual answer.

Require cited sources for any factual dataset; flag code that hard-codes or simulates the data it is supposed to look up.

Verifier role must check data provenance, not just computation; answers built on synthetic data are blocked from submission.