Structural Fuzzing: Adapting Input Mutation Testing for Parameterized Model Validation

Software fuzzing---the automated generation of mutated inputs to discover program crashes---is a cornerstone of modern software testing. Tools such as AFL and LibFuzzer mutate program (byte sequences) and monitor for assertion failures, guided by code-coverage metrics. We observe that parameterized predictive models face an analogous validation challenge: standard techniques (cross-validation, information criteria, Bayesian model comparison) test whether a model data but not whether its structure is , , or under parameter perturbation. We introduce , a six-component validation framework that adapts the adversarial mindset of software fuzzing to model validation by mutating structural rather than program inputs. The six components are: (1)~ ---exhaustive search over parameter subsets to find minimal sufficient structure; (2)~ ---identifying models optimal at each complexity level; (3)~ ---ablation to rank parameters by predictive contribution; (4)~the (MRI)---a weighted-percentile robustness metric adapted from the Bond Index for LLM evaluator robustness; (5)~ ---systematic log-spaced search for the minimal perturbation that flips a prediction; and (6)~ ---greedy parameter addition to reveal optimal ordering and diminishing returns. We demonstrate the framework on a nine-dimensional geometric economics model predicting 16 behavioral targets across game theory, prospect theory, and published meta-analyses. The campaign reveals that a three-dimensional subset achieves = 2.70\ All structural parameters are discovered by automated search, not hand-tuned. We further validate on a software defect prediction model, where the framework discovers that complexity and process metrics suffice while traditional size metrics (LOC) contribute minimally---consistent with decades of empirical SE findings. The methodology is domain-general and applicable to any parameterized predictive model, including machine learning models, agent-based simulations, and scientific computing pipelines. Target venue: ACM TOSEM. Author preprint deposited for archival and citation. Draft — pending author review.