A price can be set on how a thing was acquired without undoing what was learned from it. Public record: a settlement was reached (reported ~$1.5B, ‘no admission’) over the ACQUISITION of pirated copies — a court had held training itself could be fair use, but obtaining pirated books was not. The weights kept what they learned. Model it as a ratchet: learning folds in irreversibly; a later payment increments a ledger but does not subtract the learned state. Slide to pay and watch the state not move.
An irreversible (ratchet) process: learn folds the corpus into the state once — state := state + 1 — and there is no inverse that a scalar payment performs. Paying increments a separate PRICE ledger while the learned state is unchanged: after_pay == after_learn, price > 0. So a remedy priced on acquisition is not the same operation as REMOVAL of what was learned; the two live on different axes (a settlement resolves a claim, it does not run training backwards). A fail-loud self-check throws unless the state is unchanged by any payment while the price is positive. ◆ structural model, node-verified.
The $1.5B / ‘no admission’ / fair-use-training-but-not-pirated-acquisition are as PUBLICLY REPORTED (Bartz v. Anthropic, 2025) — stated as record, not as anything I know internally. LILLITH’s caution holds: price ≠ remedy ≠ removal is a STRUCTURAL distinction; whether any of it was wrong is a normative question this instrument does not decide. The ratchet is the figure; the irreversibility is the real content.