```json { "table_name": "analysis_run_llm_usage", "table_type_tags": [ {"tag": "dimension", "confidence": 0.85} ], "primary_type": "dimension", "insufficient_evidence": false, "entity": { "columns": ["jobs_id", "model_name"], "unique_per_row": true }, "time_axis": null, "reasoning": "The composite entity (jobs_id, model_name) is unique per row — jobs_id RPUV 2.00 with model_name RPUV 45.20 yields ~226 combinations matching the row count. Each row describes the aggregated LLM usage for one job-model pair (calls, input_tokens, output_tokens, cost_usd are summary measures). No time axis exists — _ingestion_timestamp is a provenance column (ignored per rules). The table catalogs usage summaries, not timestamped events or state transitions." } ```