Model Parameter-Scale Classification and Black-Box Provenance Auditing via Multi-Density Reverse Cloze Resonance Spectroscopy
Abstract
As generative foundation models become embedded across commercial chatbots, autonomous agents, and synthetic content streams, determining the underlying parameter capacity tier (8B, 24B, 70B, or 600B+ MoE) and architectural lineage of black-box text outputs has emerged as a fundamental challenge in AI forensics, regulatory compliance, and distillation defense. Prior watermarking methods fail without generation-side control, while white-box curvature estimators (DetectGPT, Binoculars) are rendered inoperable by logit-suppressing commercial APIs. In this paper, we introduce Parametric Cloze Resonance Spectroscopy, a privacy-preserving, zero-logit framework that classifies generator parameter scale directly from text streams. Rather than relying on naive absolute self-model preference, we establish the Capacity-Bounded Discourse Reconstruction (CBDR) principle: prober reconstruction congruence is a monotonically increasing function of prober capacity bounded by generator discourse entropy. Model scale is profiled via the Parametric Susceptibility Gradient and Discourse Transition Jitter (J), capturing the inter-paragraph root-mean-square friction of infill hypothesis matching under Kuhn-Munkres bipartite optimal assignment. Evaluated across 96 document-prober combinations spanning four parameter tiers (8B, 24B, 70B, 671B) and four model families (Meta LLaMA-3, Alibaba Qwen-2.5, Mistral, DeepSeek), we demonstrate that cloze resonance is predominantly model-family invariant (mean cross-family delta of only 4.38% on the 70B tier and 5.65% on the 8B tier). Furthermore, benchmarking against authenticated human academic controls demonstrates an overwhelming separation margin (Delta >= +40.7% to +55.1%, p < 1e-60, Cohen's d = 3.60), guaranteeing a 0.00% false positive rate at threshold tau = 35.0%.
