OEA: Structured Recursive Calibration for Generative Stability

This paper presents an empirical study of the OEA (Ontology, Epistemic, Agentic) framework as a three-layer protocol for improving reliability in recursive generative systems under synthetic-data exposure. The central thesis is: recursive generative stability depends more strongly on calibration direction and epistemic filtering than on unconstrained retrieval augmentation. The study reports four experiments: (1) a bigram-proxy ablation study across 12 variants, (2) a four-model real LLM validation on distilgpt2 (82M), gpt2 (124M), gpt-neo-125M (125M), and Qwen2.5-1.5B (1.5B) with corpus-grounded token-overlap retrieval, (3) a 30-step recursive memory drift benchmark, and (4) a baseline competition against temperature reduction, top-k, entropy filtering, and repetition penalty. Full OEA achieves TRR = 0.836, FRR = 0.081 (d = 4.56, p < 0.001). Results are fully reproducible with fixed seeds; all artifacts are on GitHub.