```json { "table_name": "llm_calls", "table_type_tags": [ {"tag": "event", "confidence": 0.90}, {"tag": "dimension", "confidence": 0.15} ], "primary_type": "event", "insufficient_evidence": false, "entity": { "columns": ["id"], "unique_per_row": true }, "time_axis": { "columns": ["created_at"], "kind": "event_time" }, "reasoning": "The entity id is unique per row (Uniqueness 1.000, RPUV 1.00), and created_at is a complete event timestamp (Uniqueness 1.000) with high monotonicity (1.000). Each row records an LLM API call occurrence — something that HAPPENED — with purpose, model, status, and performance metrics (tokens, cost, latency). The rows are bare occurrences to count and analyze; replaying them does not reconstruct a meaningful state of the id entity (the id is an artificial occurrence identifier, not a tracked object). This is event. The unique-per-row grain gives a weak secondary dimension reading, but the action-driven row existence dominates." } ```