Verification Without a Referee: Authorship-Dependent Null Testing of Gravitational-Wave Population Features

As data-driven and generative tools scale, scientific progress is increasingly constrained by hypothesis verification rather than hypothesis generation. In observational astrophysics, flexible non-parametric population models (such as B-splines, Gaussian processes, and autoregressive mixtures) serve as the simplest, low-dimensional baseline of automated hypothesis generators: they propose structural features (peaks, breaks, and subpopulations) without imposing rigid functional forms. However, because flexible models readily fit Poisson noise and measurement uncertainties in finite catalogs ($N \sim 100\text{--}300$), synthetic injection null tests are required to calibrate false-alarm structure. We audit a live debate in gravitational-wave (GW) astrophysics regarding a claimed off-axis peak in the black hole spin-tilt distribution, and expand this to a cohort of $N=4$ recent population-feature families, quantifying co-authorship overlap across discovery, extension, and verification studies using the Jaccard index $J$. We show that while the field has developed rigorous mock-catalog verification tooling, its use is uneven: in our cohort the one exhaustive injection null contributed by a discovery-adjacent group shares authors with the discovery paper ($J=0.25$), the one fully independent rebuttal is author-disjoint ($J=0$), and the highest-overlap follow-on extension ($J=0.27$) skips a matched null for its newly claimed features---even though essentially the same core team ran such a null a month later in a narrower follow-up. This dynamic illustrates a structural risk for AI-driven discovery: when verification tools are accessible but not systematically mandated, validation tracks social proximity rather than standardized protocols. We propose the \emph{Null-Audit \& Independence Statement} (NAIS) as a lightweight disclosure standard to make verification provenance transparent.