Weekly revenueFinance
Week of 2022-12-26$183,882Completed sales only
MetricGuard
Semantic conflict intelligenceVerified evidence replay
Case 01 · Weekly revenue
Two dashboards.$28.9M apart.
Both teams reported “weekly revenue.” MetricGuard inspected the SQL behind the dashboards, found the semantic conflict, and executed both definitions against Postgres to prove the business impact.
71 weeks executed13.07% mean gapHuman approval before write-back
Weekly revenueExecutive
Week of 2022-12-26$213,807All order statuses
$29,925 conflict
01 / FIND
The SQL looked almost identical. One missing predicate changed the business.
SELECT DATE_TRUNC('week', order_date), SUM(total_amount) FROM metric.orders GROUP BY 1
SELECT DATE_TRUNC('week', order_date), SUM(total_amount) FROM metric.orders WHERE order_status NOT IN ('canceled', 'returned', 'disputed') GROUP BY 1
High-severity semantic conflict · filtersExecutive revenue includes canceled, returned, and disputed orders; Finance excludes them.
02 / PROVE
Not an LLM opinion. Both definitions were executed.
$28.9M
Cumulative absolute disagreement across the recorded comparison window.
Executive$213,807
Finance$183,882
First divergent week$29,925 difference
03 / GOVERN
Evidence becomes organizational memory. A human still decides.
Three weekly-revenue definitions found across Finance, Executive BI, and Sales Ops.
Semantic differences and warehouse divergence recorded as deterministic evidence.
The agent proposes Finance as canonical. Nothing writes until a person approves.
Seven approved governance actions were live-verified in DataHub: tags, properties, descriptions, and a decision document.