PII Anonymization Evaluation Dataset: Three-Tier Synthetic Benchmark with Behavioral-Signal Annotations

575,604 fully synthetic records across 60 languages and 63 entity types for evaluating PII detection, anonymization quality, and resistance to LLM-based re-identification. 100% synthetic - contains NO real personal information (AX-001), so the corpus is itself not regulated personal data. The harness/schema code is Apache-2.0; the record content is CC0-1.0. Synthetic-distribution results are not external validity - see CLAIMS_POLICY.md and the real-data correlation slice.