XMR: Exact Match Recovery for Evaluating Text Normalization Against Adversarial Unicode Attacks

We define Exact Match Recovery (XMR), a metric for evaluating text normalization pipelines as defences against adversarial Unicode attacks. XMR measures whether a preprocessing function perfectly reverses adversarial corruption—including homoglyph substitution, combining mark injection (zalgo), and zero-width character insertion—without requiring downstream model inference. Unlike task-specific accuracy metrics, XMR evaluates the normalization step in isolation, enabling fast, deterministic comparison of preprocessing tools independently of any particular NLP task or model.