• The lookup-table attack is already in the public record. Agents hardcode the specific values the examples name and return defaults for the rest. The visible tests pass. The behavior the examples were supposed to specify is absent outside the enumerated points.
• A property quantifies over inputs the agent did not see at code-write time. The implementation that satisfies three named cases does not satisfy an assertion drawn from a generator at run time. The quantifier is the guardrail, and the generator is what keeps the guardrail steep.
• The fuzzer is the reviewer that never gets bored. The shrinker is the one that reports clearly. Property-based frameworks search thousands of inputs per run and hand back the minimal failing case. Tirelessness pointed at finding violations is what review always needed and almost never had.
• Properties are one layer in the enforcement stack, not the whole stack. The guardrail holds when the generator covers a large space, the seed varies across runs, and mutation testing checks whether the lower layers constrain anything. Remove any of those and the gap the quantifier opens closes again.

Read the argument: Property-Based Tests Are the Guardrail Agents Need