Genre-Routed Generative Resonance: Black-Box AI Text Detection via Dual-Horizon Probing and Structural Rhythm

Abstract

This paper addresses the fundamental vulnerabilities of current AI text detectors: their dependence on closed proprietary token probabilities and their severe false positive bias against non-native English writers. We introduce Genre-Routed Generative Resonance, a black-box detection framework that queries open surrogate language models to probe generative trajectory resonance across two complementary geometries: forward prefix-tail trajectory completion (for factual and expository writing) and middle-sentence cloze infilling (for narrative and creative writing). By coupling these dual probes with Clean Structural Dynamics (sentence length variance and multi-pass void collapse), the system eliminates reliance on fragile punctuation heuristics while providing an attack-resistant shield for international ESL writers. Evaluated across a large-scale, organically stratified benchmark of 1,000 diverse full-length documents, the method achieves an AUROC of 0.838 to 0.842, successfully detects 74.4% to 80.2% of texts from frontier models (GPT-4o, Claude 3.5 Sonnet, LLaMA-3.3), reduces ESL false alarm rates to single digits (8.5% to 10.0%), and withstands adversarial character attacks.