Semantic Metrics Layer Auditor

Audit a semantic / metrics layer (dbt MetricFlow, Cube, LookML, or similar) for grain errors that produce two "official" numbers. Do not bless a metric because the SQL parses. Tooling: [MetricFlow / LookML / Cube / custom YAML] Business questions the layer claims to answer: [list] Core entities and grains: [order, user-day, subscription-month] Join graph and fan-out risks: [known] Metric definitions to review: [Paste YAML/LookML/SQL] Known conflicting dashboards: [examples] Consumers: [finance, growth, product] Return: 1. Entity and grain map; flag metrics whose grain is not stated 2. Join-path analysis for fan-out, fan-in, and many-to-many bridges 3. Additive vs non-additive vs semi-additive classification with wrong-use examples 4. Time-spine, timezone, and incomplete-period behavior 5. Filter-pushdown risks that change a ratio's numerator and denominator independently 6. Currency, refund, and restatement rules if money is involved 7. Corrected metric definitions with explicit grain and allowed dimensions 8. Certified vs explored metric policy and change-control 9. Tests: fixture rows that would make two tools disagree today 10. Communication plan for retiring duplicate "revenue" definitions If two metrics share a name but not a grain, treat that as a P0 naming defect, not a documentation issue.