Source Updates and Forecast Reliability: Probability-Bounded Revalidation of Astrophysical Data Pipelines

Abstract A changed source file identifies a dependency that may need review, but it does not quantify the change in a scientific prediction. We formulate a target-specific revalidation rule that supplements lineage with a bound on the difference between declared output probability laws. For Gaussian products, a relative-entropy calculation and Pinsker’s inequality bound the change in probability of every event in the registered output space. Low-rank mean and covariance updates make the test executable without rebuilding a dense source covariance. The rule distinguishes an unchanged target, a retained approximation with a stated probability bound, recomputation, and a required scientific review when the estimand or its applicability conditions change. A fully synthetic cosmological-distance pipeline supplies 360 sources, 21 scalar or joint products and 80 calibration-update scenarios. Among 1,680 product–update pairs, 255 have unchanged target distributions, 752 satisfy a probability-difference tolerance of 0.05, and 673 require recomputation. Separate dense calculations agree with the low-rank relative entropy to (9.06 \times 10^{-14}). A sequential example demonstrates that checking only the most recent small update can leave a stale output 0.1585 away in total variation, despite every local update passing its threshold. Comparing against the last fully recomputed anchor limits the accumulated error and requires three anchor updates in 50 steps. These results establish conditional revalidation mechanics, not empirical calibration of real astrophysical forecasts or a general runtime speedup. Keywords: scientific data updates; predictive reliability; executable provenance; covariance changes; total variation; cosmological distances.