This dataset captures the operational telemetry of a data profiling and classification engine, logging 2.0 million events across 20 tables that track how datasets flow through analysis pipelines, how columns are catalogued and tagged, and how large language models are invoked to interpret schema semantics. The architecture is heavily skewed: **columns_catalog** alone holds 1.5 million rows—75% of all data—making this primarily a column-level metadata repository with auxiliary job execution and LLM usage logs. Referential integrity is strong at 94%, anchored by validated joins from **analysis_run_llm_usage** and **tables_catalog** back to their parent **analysis_runs** and **datasets** tables, enabling reliable reconstruction of which profiling jobs touched which datasets and what AI inference costs they incurred. The critical quality risk is **nia_messages.author_geo**, which is 100% null across all 5,279 records and renders geographic attribution of system messages impossible; if user location or compliance jurisdiction matters, this gap must be closed at the source. The dataset excels at retrospective analysis—tracing column discovery over time, auditing classification decisions through **classification_candidates** and **classification_overrides**, and correlating LLM token spend with job outcomes—but it is not transactional data suitable for real-time alerting or customer-facing analytics. Use it to optimize profiling workflows, benchmark model performance, and justify infrastructure investment; do not expect it to answer questions about the *business domains* being profiled, only about the profiling system itself.