Infrastructure Induced Selection Convergence in Generative AI Enabled Information Production_

We define infrastructure-induced selection convergence as agreement in angles, claims, sources, frames, and headline focus beyond an event-specific independence null. A computational newsroom processed primary-source packets for 30 real events under shared-model, role-differentiated, heterogeneous-model, single-agent, and diversity-aware architectures.

The shared model converged more than heterogeneous models on angles, claims, sources, and headline focus; a blind classifier did not reproduce the trace-based frame contrast. Role prompts did not restore variety, and minimal early cues did not persist. A diversity-aware candidate-selection architecture increased the effective angle repertoire from 1.23 to 4.14 and reduced independently classified frame convergence without a meaningful decline in source fidelity.