Recurrent Merger-Local Morphology in Gravitational-Wave Post-Model Data

Most post-model analyses begin with a named target: a waveform error, detector artifact, unmodeled transient, or parameterized departure from an established model. RIUD Paper I asks an earlier question: can recurrent merger-local structure be converted into a stable scientific object before its physical identity is known? The audited Paper-I development record comprises 612 preserved executions spanning successful, negative, unresolved, and technical branches. Development began with fixed physical phase-deformation models and reversed direction after those models failed to yield a stable population result. A residual-first reference in GW190521_074359 produced a two-stage merger-local morphology that survived synthetic-null, real off-source, waveform, posterior, and intrinsic-parameter-refit attacks, while a competing ringdown representation prevented the original basis from being treated as unique. The search was then generalized from fixed coordinates to an event-native family over local time, frequency, temporal width, inter-detector lag, and detector relation. In locked matched-background development tests, GW190521_074359, GW150914, and GW170104 each produced 0/1000 exceedances (p+1 = 0.000999); GW170814 entered as a non-blind catalog extension and also produced 0/1000 under its fixed-candidate validation. A later rigid temporal interpretation failed, separating the observational morphology from a stronger temporal story. The observation/scoring machinery itself was then audited. In a 227-row morphology/scorer audit, metadata-only leave-one-out classification reached 0.404, width-only classification reached 0.762, and frequency-width association gave Cramér’s V = 0.752, while direct mass dependence was negligible. Physical-sign posterior subtraction removed or weakened a stronger interpretation in 79/150 samples, whereas wrong-sign controls removed 0/12. An attractive 310.4-Hz branch failed an independent raw-FFT placebo in 3/3 target events and was discarded. The mature event-native score was deployed across a frozen 374-event native-strain universe, explicitly as a deployment census rather than 374 detections; the frozen score distribution comprised 31 FAIL, 162 PASS, 150 STRONG, and 31 VERY STRONG events. A later audit retained 374/374 finite scores, 343 H1/L1-ready events, 327 QC-valid events, and 300 prospective-eligible events after excluding 27 outcome-seen cases; the selection score was not retuned. Subsequent physical-identity tests rejected or bounded simple ppE-like, single-wavepacket, rigid common-center, and universal fixed-GR interpretations. A 29-event held-out cross-detector audit supported detector-related coordinate localization, but a conservative matched-SNR waveform-mismatch bank remained statistically compatible with the observed residual distribution. The surviving result is therefore deliberately bounded: RIUD Paper I establishes a recurrent event-native merger-local morphology as an auditable observational object, while leaving its unique physical cause unresolved.