The gold layer holds business-ready tables: metrics, aggregates and dimensional models at a declared grain, built for dashboards, semantic models and reports. Gold tables apply business definitions once, so every report uses the same numbers. They are usually smaller than silver and shaped around questions rather than sources.
For example, a gold table gold.monthly_sales might hold one row per month and region with revenue, order count and average basket size, computed from the silver orders and stores tables. The definition of revenue, for instance whether returns and tax are excluded, is written once in the SQL that builds this table. A finance dashboard, a sales report and an analyst's notebook that all read gold.monthly_sales show the same revenue, which is the main reason the layer exists.
Gold tables come in two broad kinds: dimensional models, with fact tables and dimensions that tools can slice in many ways, and aggregated tables built for one use, such as a dashboard that must load fast. Both need a stated grain, and both are usually rebuilt or merged on a schedule after their inputs are refreshed.
Because many people read gold tables, they are where changes hurt most. Renaming a column or changing a definition can break reports that the team building the table never sees, so gold tables are good candidates for a data contract and for tests that check totals against known values. See also semantic model and the sample dashboard, which reads gold tables of synthetic data.
On Databricks and Microsoft Fabric, gold tables are usually Delta tables in a schema that BI tools can read; on Snowflake, they are often tables or dynamic tables refreshed after their inputs. Wherever they live, they are the tables a dashboard should read, so a dashboard never repeats the business logic itself. See also grain and fact table.
Emolysis's Aggregation / gold shape builds gold tables, and a gold table is what the agent builds when you ask a business question. See Dashboards from a question. Because a gold table answers a question at a stated grain, its totals can be checked against known values.