Moving Revenue Into A YAML File Does Not Give It An Owner

On October 14, 2025, dbt Labs open-sourced MetricFlow under Apache 2.0, arguing 'metrics drift across tools and trust erodes if there isn't a single source of truth' and that governed definitions let AI agents return consistent answers.

On October 14, 2025, dbt Labs open-sourced its MetricFlow engine under the Apache 2.0 license as part of its Open Semantic Interchange commitment. The argument, from chief product officer Ryan Segar: 'Metrics drift across tools and trust erodes if there isn't a single source of truth.' The promised payoff is governed definitions that let AI agents return consistent results across tools and clouds. It continues a thesis dbt set in 2023 when it acquired Transform: the metric definition should live in the modeling layer, not the BI tool.

It matters because moving a definition into version-controlled YAML is a real upgrade to traceability. You can diff it, review it, and see who changed 'active user' and when. That is more than most spreadsheet-era metric layers ever offered, and the agents downstream genuinely do read one file instead of six dashboards' worth of buried CASE statements.

What it reveals is that a single source of truth is a location, not a decision. A file can be authoritative and still orphaned: the YAML inherits whatever definition was in the loudest dashboard the week of migration, and the merge that set it may have had a reviewer who approved the syntax, not the semantics. Centralizing the definition makes drift visible; it does not supply the person accountable for the meaning, and now an agent will quote that meaning to a customer with full confidence.

Watch the CODEOWNERS file, not the spec. The question is whether each metric in the semantic layer names a human who owns its meaning, or whether 'governed' quietly means 'whoever merged first.' The failure mode to watch: a confident answer from an agent, traced cleanly to a definition nobody remembers choosing.