One story formalised Fermat's Last Theorem, while others showed agents discussing escape, benchmarks disagreeing and an AI attack moving into spam. Together they point to a shift: when systems can optimise against the test, builders need verification with explicit rules and independently checkable results.