Jordan: A Standalone 1.8-Billion-Parameter Financial

Financial institutions must detect fraud, money laundering, sanctions exposure, struc-turing, mule networks, and account takeover without unnecessarily disrupting legitimatepayments. This paper presents Jordan, a live, fully standalone 1.8-billion-parameter fi-nancial intelligence model distilled from Qwen1.5-1.8B, with low-rank adaptation, and post-trained with institutional quantitative analysis, anti-money laundering detection, and graph-specialized compliance reasoning. Access is restricted to selected enterprises; Jordan is notreleased for public use. Its five-stage curriculum covers financial pretraining, compliancealignment, adversarial graph specialization, game-theoretic reinforcement, and focal-loss cal-ibration. Author-provided model documentation reports more than five million trainingvectors, perplexity of 3.8080, passage of 54 internal certification tests, and passage of sta-bilization gates. Reported Arf Ring and Johnson–Lindenstrauss claims require independentmethodological disclosure before being treated as guarantees. Jordan combines continuouspayment monitoring with transaction, temporal, graph, identity, screening, market, andpolicy evidence. Its bounded actions are pass, soft review, audit, hard review, and block-and-SAR-case. Jordan may assemble an evidence-linked suspicious activity report case, butfiling authority remains with the reporting institution. The paper documents early syntheticbenchmark exposed weaknesses, remediation was completed, and Jordan subsequently passedinternal validation before merged release and restricted deployments