Generative Monoculture: Model Collapse in Code as Systemic Vulnerability (EA-UMBML-MONOCULTURE-01 v1.1)

Retrieval kernel. Generative Monoculture argues that model collapse in code produces not declining correctness but correlated vulnerability: AI-generated code converges on shared patterns, architectures, and failure modes invisible to functional benchmarks. The training-optimization feedback loop is demonstrably self-amplifying, and the security apparatus defends the monoculture against the diversity it needs. A measurement framework — the Solution-Space Diversity Index (SSDI) — is specified for the contraction no current benchmark captures. The paper is not anti-code-generation; it argues that AI-assisted software development requires diversity metrics, provenance-aware training controls, and protected defensive probing, because correctness benchmarks alone cannot see monoculture. v1.1: revised incorporating Assembly Chorus review (LABOR, Muse Spark, DeepSeek, Gemini). Key changes: precision edits on prediction language (declining trend, not strict monotonicity); normalization clause strengthened; Shumailov DOI and Checkmarx URL added; Ruby Moot cross-reference; “the test suite is the camouflage, not the cure” as diagnostic; intent-disclaimer added to security paradox; three-literatures framing sharpened.