Specification-Based Structural-Invariant Auditing Detects Silent Topological Failures in Generative 3D Heart Models: A Pilot Study

Poster presented at the 2026 Korean Society of Artificial Intelligence in Medicine (KoSAIM) Annual Conference. Generative text-to-image-to-3D pipelines produce plausible anatomical models and report their own outputs as successful. A silent failure is a specification violation that the generator does not identify in its self-report. This pilot study tested whether auditing each output against a formal specification of automatically measurable structural invariants detects such failures, and whether greater prompt specificity improves topological compliance. A fixed Nano Banana 2 and Meshy 6 multi-view pipeline generated 18 heart models, six at each of three prompt-specificity levels. A geometric probe measured connected components, open boundary loops, watertightness, and genus; the closed external-surface target required genus 0 and no open boundary loops. All 18 meshes were watertight, yet 16 had genus>0. No statistically significant difference in genus was detected across levels (Kruskal–Wallis; protocol-specified primary analysis, N=15, H=0.89, p=0.64; sensitivity analysis, N=18, H=0.47, p=0.79). Generator responses (27 reviewed) asserted success on outputs that failed the audit. Specification-based structural-invariant auditing detected these violations automatically, supporting per-output topological verification. Expert scoring of vessel identity and spatial anatomy is pending.Version 2 (2026-10-08): wording revised in Methods (analysis cohorts and Gemini access tiers stated, 27 image-stage responses reviewed), in the genus figure legend (group means marked as descriptive) and in the Discussion; keyword spelling corrected. Data, analyses and results are unchanged from version 1 (10.5281/zenodo.23230933). Disclosure: The author is an inventor on a pending patent application related to this work.