Generative Monoculture Model Collapse in Code as Systemic Vulnerability

The paper links three observations:

• recursive synthetic training can narrow distributions;
• generated code can pass tests while retaining vulnerabilities;
• a small number of dominant models can propagate similar patterns at ecosystem scale.

Its causal chain is:

distribution narrowing → pattern convergence → shared failure modes → correlated vulnerability

The correctness trap is the idea that compilers and test suites filter out non-functional code but cannot detect a shrinking diversity of correct solutions.

Correlated vulnerability is a population property: separately produced systems share an exploitable flaw because they inherit the same generated pattern.

The proposed Solution-Space Diversity Index is intended to compare code solutions through features such as:

• architectural variation;
• control-flow and dependency structure;
• algorithmic family;
• idiom and library selection;
• vulnerability-class covariance.

The paper argues for diversity-aware benchmarks, provenance controls, and protected external probing rather than opposition to code generation itself.