This dataset captures the operational telemetry of a data profiling and classification engine, logging 2,035,926 events across 20 tables that track how analysis jobs consume LLM tokens, tag columns, and catalog metadata. The architecture is heavily skewed: **columns_catalog** alone holds 75% of all rows (1,533,257), recording individual column profiles, while **job_logs** (303,822 rows) and **llm_calls** (49,360 rows) trace execution paths and AI inference activity. Referential integrity is strong at 94%, anchored by the **analysis_run_llm_usage.jobs_id → analysis_runs.jobs_id** join (100% confidence), which reliably links token consumption to specific profiling runs. However, completeness sits at 76%, dragged down by a critical gap: **columns_catalog.avg_value** is 100% null (1,532,986 of 1,533,257 rows), eliminating the ability to perform statistical range analysis or detect numeric outliers across profiled datasets—a significant blind spot for any platform claiming to deliver data intelligence. The dataset excels at **cost attribution and LLM efficiency analysis**. With complete lineage from **llm_calls** through **analysis_run_llm_usage** to **analysis_runs**, you can calculate token spend per job, compare prompt strategies, and identify which classification tasks drive inference costs. The **tag_lineage_edges** and **column_tags** tables (42,024 and 43,853 rows) enable tracing how semantic labels propagate through column hierarchies, supporting governance workflows. What this data *cannot* do is reconstruct the business semantics of the datasets being profiled—there are no sample values, no domain-specific validation rules, and the missing **avg_value** column means you cannot benchmark numeric distributions. This is telemetry *about* a profiling system, not a window into the datasets it analyzes; use it to optimize the engine's performance and cost, not to answer questions about the underlying data it processes.