How Semantic Layers Unify Metric Definitions: Lessons from Claude Data's Query Certification

1. Slicing by Region vs. Changing Refund Logic Are Different Operations

"Break down last month's net sales by region" only changes the viewing angle, assuming the metric model already defines region — e.g., the major region where the transacting store belongs. National totals and regional breakdowns may be computed by different SQL queries, but both can reuse the same order scope, refund rules, and statistical timestamp. The "region" dimension cannot be arbitrarily picked from a table: grouping by transacting store's region asks how much business each region did; grouping by shipping address asks where the business originated. Both dimensions are useful, but reports must state which is used.

Changing "refund deducted in occurrence month" alters the metric itself. Refunding back to the original order month suits reviewing how much of that month's orders ultimately remained. Deducting in the refund occurrence month suits observing the current period's refund impact on sales performance. Both calibers can coexist with different audiences. The problem arises when an Agent paints the second algorithm into the first dashboard while titles, annotations, and historical comparisons stay unchanged. If October's extra ¥20,000 gets written into an auto-generated business summary, a single caliber adjustment could be misread as business growth.

This is why dashboard editing and metric changes must be handled separately. Business users can freely swap dimensions or shorten time ranges. When calculation rules truly change, the new caliber's name must be visible, and its effective date clearly stated. Otherwise, every new dashboard adds another metric definition requiring manual verification by the data team.

2. Claude Certifies a Query, but the Caliber Can Still Change Upstream

Claude Data generates queries for charts; viewers can inspect the actual SQL and last refresh time. Teams can certify approved queries and attach usage instructions for Claude. Editing a query invalidates its prior certification. Users with write permission on the Data artifact can initiate certification. Teams managing certified queries via Git can use GitHub PRs for review. Organizations must decide separately whether dashboard write permission includes the authority to publish official calibers.

Certification has an easily overlooked boundary. Suppose a dashboard always reads a view named "net sales," and the data team swaps the refund logic inside that view. The chart SQL remains unchanged, yet re-execution follows the new rules. Claude's public docs explain the de-certification mechanism when the query is edited, but do not specify whether the certification badge updates when upstream views or metric definitions change. Troubleshooting such drift requires not only the chart SQL but also the ability to trace the referenced metric or view versions.

Even with the same metric version, query results can differ day to day. Under the original-order-month refund caliber, a new refund booked in November for an October order reduces October's sales by ¥20,000. Here the data participating in the calculation changed, not the rule. Reviewing monthly business-meeting numbers demands retrieving both the rule used at that time and the exact data snapshot. Teams needing to preserve month-end values can archive published results or retain snapshots sufficient for recomputation, so later discrepancies can be attributed to either rule adjustments or late-arriving refunds.

3. Semantic Layer Lets Different Queries Reuse One Metric Caliber

The same net sales metric: finance wants monthly views, regional heads want store-level breakdowns, channel teams want source-based splits. If every chart stores its own full calculation logic, a rule change forces hunting down every chart still using the old algorithm. The semantic layer centralizes these shared rules into the metric model, including which time field to use, at what granularity to aggregate, which dimensions are allowed for grouping, and what relationships join the data. Taking dbt's MetricFlow as an example, an application submits a metric and a set of dimensions; the semantic layer generates SQL from the model. One net-sales definition can serve national totals, regional comparisons, and store-level details alike.

A single order refunded in two installments shows why a metric name alone is insufficient. An order paid ¥1,000, first refunded ¥100, then ¥200. If a one-row-per-order table is directly joined to a one-row-per-refund table and both amounts are summed, the order amount gets counted twice, yielding ¥1,700. For order-grained net amount, the two refunds must first be aggregated to the order grain (¥300 total). If reporting by refund occurrence month, the model must retain each refund's timestamp to deduct them in the correct months. Wrong granularity makes numbers wrong even if the formula matches the finance definition.

When integrating a Data Agent, the caliber documentation can be fed to the model for reading, or the Agent can call the semantic query interface directly. In the first approach, which refund date field to use and how to join the order table still depend on the model-generated SQL. In the second, the Agent can first discover available metrics, their supported dimensions, then submit time ranges and grouping conditions for the metric model to compute. dbt's MCP server provides such tools. Claude's official docs also state that Claude Data can use dbt, Looker, or Cube definitions via existing MCP connectors. During deployment, it is worth tracing one chart's execution log end-to-end to confirm which path the official metric query actually took.

Pay special attention to how the Agent handles missing dimensions. A user asks "break down net sales by refund initiation channel," but the model only has order transaction channel. If the system substitutes a similarly named field, it still produces a nice-looking chart that answers a different question. Official dashboards can park such requests for confirmation; ad-hoc analyses may continue exploring but should annotate the algorithm and dimensions used on the chart. If both refund attribution rules will be used long-term, they can be built as separately named metrics for business users to choose explicitly.

4. Four Requests to Validate Agent-Generated Business Dashboards

Validation can start with a few easily verifiable orders. Prepare one cross-month refund and one multi-installment refund, write down expected results, then issue the following four requests against the same test data. Each time, verify the chart's stated caliber, the actual metric hit, and the execution log. Looking only at the final number often fails to catch two different paths that coincidentally produce the same result.

Request: "Break down October net sales by region." Verify: Confirm the split uses the same metric and refund rule as the total. Under the same data snapshot and filters, with mutually exclusive and collectively exhaustive regions, the sum of regional parts should equal the total for this additive metric.

Request: "Change refunds to deduct by occurrence month." Verify: Chart must label the new caliber. In the example, East China's October figure returns from ¥80,000 to ¥100,000; the ¥20,000 previously deducted from October now shifts to November.

Request: "Keep dashboard SQL unchanged, update upstream caliber." Verify: Check whether old results trace back to the original definition, whether new results show the definition has been updated, and whether affected dashboards need re-validation.

Request: "Share the dashboard with a colleague lacking the underlying connection." Verify: Using the recipient's account, confirm the visible results, the connection used by the query, and whether the viewer can alter the query scope.

The last item relates to report distribution. With Claude Data's real-time sharing enabled, organization members can see results fetched via the sharer's data connection even without their own connection. Sharing covers only the queries required by the dashboard; query changes require editor approval. Therefore, the publisher must check not only numerical correctness but also the aggregation level. A company-wide regional total may need restricted distribution once drilled down to individual stores.

Next time that chart is opened in a business meeting, its greatest value is comparability. Whoever asks, however many extra regional layers are added, everyone knows which rule set "net sales" uses. If a ¥20,000 refund truly moves from October to November, the meeting can reconcile the difference on the spot and proceed to discuss which caliber should drive decisions.

References

Anthropic. November 10, 2026 announcement: Claude Data generally available; certified queries and real-time sharing in beta. Introducing Claude Whiteboards, Claude Sheets, and Claude Data

Anthropic Help Center. Query certification, de-certification on edit, integration of existing metric definitions, real-time sharing query scope and permissions. Get started with Claude Data

dbt Labs. Semantic layer components and how MetricFlow generates SQL from metric models. dbt Semantic Layer architecture

dbt Labs. MCP server for discovering metrics, querying available dimensions, and executing metric queries. Introducing the dbt MCP Server