The proliferation of misinformation on social media has driven extensive research into automated fake news detection. However, the "Post-API" era presents a critical challenge: as major platforms implement restrictive data-access policies, traditional approaches relying on explicit social graphs and user metadata become increasingly impractical. This paper proposes RGCP-Snowflake, a structural-semantic fusion framework that leverages Snowflake IDs as structural surrogates to reconstruct temporal propagation patterns without requiring access to restricted social metadata. By extracting millisecond-level timestamps encoded within distributed identifiers, the proposed approach recovers the propagation rhythm, a latent veracity signal characterising the dissemination patterns typical of fake news. A comprehensive benchmarking across three model tiers (commercial large language models, traditional machine learning classifiers, and open-source language models) is conducted on the FakeNewsNet dataset. The experiments reveal two key findings: a scale threshold phenomenon, where smaller language models exhibit systematic label collapse, defaulting to classifying all inputs as misinformation; and the competitive performance of traditional classifiers against larger open-source models, indicating inherent limitations in purely semantic zero-shot reasoning. RGCP-Snowflake outperforms all baseline models by integrating non-manipulable temporal signatures with semantic headline features, demonstrating that structural surrogates provide a robust detection mechanism effective in restricted environments where explicit social metadata is unavailable.

PDF