ReturnRing: fair return decisions within EU law

The same quick return can mean a wrong size or a ring of linked accounts. ReturnRing helps EU fashion merchants tell them apart: restrictions apply only to future orders, and the 14-day right of withdrawal is never denied. In the US, $86B of the goods returned in 2025 was return abuse, 6× return fraud (Appriss Retail, industry estimate). Statistical models make every prediction, never the language model: XGBoost estimates return probability and abuse risk, with SHAP reasons per case; an account graph proposes links from shared identifiers and shows the evidence for each; a product-anomaly check separates catalogue problems from customer behaviour. Customer value never feeds the risk score. A rules engine picks the admissible actions, a legal guard checks them, and a simulator estimates net benefit under explicit assumptions, never counting the refund owed as a saving. Gemma 4 31B on AMD Token Factory explains each decision and drafts the customer message in Italian and English; a validator checks every draft before it is shown, and a person approves every restriction, with a review date. XGBoost and a GraphSAGE encoder were trained on an AMD Instinct MI300X with ROCm, on a 20,000-customer synthetic store calibrated on real ASOS aggregates. Measured: F1 0.739 against 0.721 for a threshold rule on 1.4M real ASOS purchases; AUPRC 0.981 against 0.633 for rules alone on 6,009 held-out synthetic customers, with 2 of 1,722 high-value customers flagged by mistake. The public demo is live and read-only, with 1,000 fictional customers and real Gemma output. ReturnRing is a working prototype for a merchant pilot: abuse detection on real data and commercial impact are still to be validated.