This paper proposes WB CoinTrust, an AI-based framework for the quantitative assessment of digital-asset (cryptocurrency) trustworthiness, and validates it through a retrospective backtest on twelve samples (three confirmed scam coins and nine legitimately operating coins). The methodology structures unstructured, multi-source data — on-chain activity, source code, disclosures, market microstructure, and regulatory status — using a heterogeneous LLM ensemble (Gemini and GPT) as evaluators (the LLM-as-evaluator paradigm), then scores it through a six-axis quantitative rubric and a non-compensatory rule engine ("integrity gate"). Results: the three scam coins (mean TRS 6.8) all received a D grade and triggered ten gates in total, while the nine legitimate coins (mean TRS 69.0) triggered no gates and distributed across AA–BB grades, yielding precision, recall, F1, and accuracy of 100% on the 12-sample set. The FTX Token case (score 14.4, forced to D by three gate triggers) demonstrates the value of a non-compensatory rule engine in catching sophisticated fraud that a weighted average conceals.
Kai Yim (Wed,) studied this question.