RedactionBench

In three linesRedactionBench is a manually annotated benchmark of 200 documents across 11 domains for evaluating PII redaction in context. Introduced with R-Score, a character-level metric, it shows 35 models (NER, SLM, frontier models) fail on contextual redactions: human consensus 89.4% for mandatory redactions, 47.7% for contextual ones.

Summary generated by Claude — human-verified