Label-Conditional Degradation: Regression-Discontinuity Evidence from 11,431 Stability Lots, and the STAMP-1 Randomised Trial of Printed Expiry Dating

Potency in 11,431 pharmaceutical stability lots is flat for the whole of a product’s shelf life and then falls $11.8$ percentage points of label claim (95% CI $11.0$–$12.6$) within thirty days of the date printed on the carton. Before that date the fitted decay slope is $-0.021\%\mathrm{LC}$/month; after it, $-0.019$. The field reads the printed date as a forecast of this drop. We test the reverse. We first show that the question is not identified in observational data: the printed date is a fixed administrative offset from the manufacturing date, so elapsed chemical time and label arrival are perfectly collinear in every stability programme ever run, and no such programme can distinguish them. We break the collinearity by randomising the label. In STAMP-1, 1,080 cartons drawn from one homogeneous manufacturing lot of each of three index products were assigned printed shelf lives of 24, 36 or 60 months and stored together on one shelf in one chamber. Potency fell at whichever date was printed — a between-arm difference of $11.0$ points at month 24 (95% CI $10.2$–$11.8$, $p<0.001$) among cartons of identical chemistry and identical storage history. Storage temperature had no detectable effect (main effect $0.3$ points, $p=0.71$; fitted $Q_{10}=1.02$). The effect scales with the legibility of the print, and cartons whose date was covered by an opaque label at $t=0$ showed no detectable loss ($-1.2$ points, 95% CI $-2.9$ to $0.5$). We propose no mechanism. Overprinting a revised date costs $0.0031 per carton, or $0.42 per DALY averted.