A token budget censors a measurement by the cost of the correct answer.
A reasoning model spends more tokens when the right answer is expensive to verify. Cap the tokens and you do not merely truncate the output — you silently drop exactly the answers that were costly to reach, biasing the result by ground truth. Statisticians already have a name for this: informative right-censoring (missing-not-at-random).
What Paper 0 pins down is that an experimenter's max_tokens cap, on a fixed model with unchanged weights, alone reproduces the exact signature that earlier work read as a model "getting worse." An inversion control — flipping which class is cheap to verify — makes the censoring follow, proving the mechanism is general, not a quirk of one task.
This is the methodology credential, not a drift finding. Before NL;NL publishes any change to a model, it must answer one question: is that the model moving, or your own measurement being censored? Paper 0 is that test. Report #1 — behavioural drift across model versions — follows once the Meter has accrued history.