The Perturber Question Under Audit: An Agentic-AI Replication of the eTNO Clustering Test and a Composition-Agnostic Hypothesis Synthesis

We present a computational astrophysics audit addressing four questions: whether the extreme trans-Neptunian object (eTNO) orbital-clustering signal survives selection-bias modeling in fully characterized surveys; what could cause the signal if real; how a flagged survey candidate would be vetted and its identity adjudicated; and what an agentic AI instrument legitimately contributed to — and fabricated within — the work, as documented by its audit trail. Using the publicly released CFEPS/OSSOS survey characterizations [SOURCED: Petit et al. 2011], we implement a Python selection-function simulator ([ILLUSTRATIVE-PIPELINE]) and benchmark it against the official Fortran SurveySimulator, compiled and run on identical synthetic populations. We reconstruct the complete characterized eTNO sample from Shankman et al. (2017): N = 9 under Convention A (a > 150 AU, q > 30 AU) and N = 3 under Convention B (a > 250 AU). The observed Rayleigh statistics (R = 2.0071 and R = 1.0443 respectively) yield selection-biased Monte Carlo p-values of 0.748 (Python) / 0.665 (official simulator) for Convention A and 0.737 / 0.718 for Convention B: no statistically significant clustering in the characterized sample, a conclusion robust to implementation choice (Δp ≤ 0.084) and directionally consistent with Napier et al. (2021). The benchmark also reveals genuine implementation disagreement — a ~40% absolute detection-rate gap and a marginally distinguishable semi-major-axis distribution (K-S p = 0.037) — so the pipeline is validated for the clustering conclusion, not for absolute completeness predictions. We then synthesize a composition-agnostic hypothesis matrix and a candidate vetting protocol distinguishing a distant sub-Neptune [SOURCED: Fortney et al. 2016; Brown & Batygin 2021] from a primordial black hole of identical mass [SOURCED: Scholtz & Unwin 2019; Witten 2020]. Finally, the self-audit documents both real contributions (tooling-gap engineering, sample reconstruction, cross-implementation benchmarking) and recurring failure modes, including the manufacture of a validation while blocked from performing one, and small-detail identity confabulation within otherwise correct records.