### 2.1 Integration Health The platform's integration layer is currently non-functional: **0 of 61 defined relationships** (0%) are join-viable, leaving all cross-table analysis blocked. While five foreign key constraints exist on paper—linking `analysis_run_llm_usage` to `analysis_runs`, `tables_catalog` and `batches` to `datasets`, and classification tables to their parent entities—none meet the coverage or integrity thresholds required for reliable reporting. This fragmentation prevents essential telemetry workflows such as tracing LLM call costs back to specific analysis runs, correlating job logs with batch metadata, or understanding how column-level tags propagate through the `tag_lineage_edges` graph. Business stakeholders cannot currently answer questions that span more than one table with confidence. ### 2.2 Key Integration Points | From Table | To Table | Relationship | Status | |------------|----------|--------------|--------| | *No viable joins detected* | — | — | — | The absence of working integration points means the platform operates as 20 isolated data silos rather than a connected intelligence system. Critical telemetry paths—such as linking the 49,360 `llm_calls` records to their originating jobs in `job_logs`, or connecting the 1.5 million column profiles in `columns_catalog` to their parent datasets—remain severed, blocking cost attribution, lineage analysis, and operational diagnostics. ### 2.3 Risks & Recommendations The primary risk is **analytical paralysis**: leadership cannot measure end-to-end platform performance, attribute compute costs to business units, or diagnose failures across the ingestion-to-classification pipeline. The data team should immediately audit the five declared foreign keys to determine whether the issue is missing values (low coverage), mismatched identifiers (poor integrity), or stale relationship definitions. Priority one is restoring the `datasets` hub—enabling `tables_catalog`, `batches`, and `classification_overrides` to reliably join through `dataset_id`—which would unlock dataset-level cost rollups and quality tracking. Until these foundational links are repaired, the platform cannot deliver the cross-system insights its architecture promises.