Circular trading, in which firms pass invoices around a closed loop to manufacture input tax credit, is a leading form of evasion under India’s Goods and Services Tax (GST). Each participating firm files returns that reconcile, so the fraud is visible only as a pattern across firms, and research on it has relied on confidential departmental data that cannot be shared, so methods cannot be compared on common ground. We release a parameterised generator of GST invoice networks with labelled rings and realistic legitimate confounders, and propose a three-stage detector: a pass-through relation, which links two firms when one forwards the other’s consignment onward at a similar value and so makes cycle structure tractable; PT-SAGE (PassThrough GraphSAGE), a graph neural network that learns from both invoice and pass-through links; and a ring-recovery step that returns suspect rings rather than isolated firms. Across 10 independently generated networks, PT-SAGE reaches an average precision of 0.948, against 0.854 for gradient boosting with engineered cycle features (paired gain +0.095, 10/10 networks) and 0.089 for rule-based red flags. It recovers 88.0% of held-out rings as complete chains, and its advantage persists when most ring members are genuine firms.

GST,input tax credit,circular trading,fraud detection,graph neural networks,synthetic benchmark