A Power BI report under change control

A .pbix file is one binary, so Git can only say "the file changed." Saved as a Power BI Project (.pbip), the same report becomes text: PBIR files for pages and visuals, TMDL files for tables, measures and relationships. Once it is text, a report can carry the controls any regulated change carries. Before I built software, I spent years in broker-dealer supervision, where the job is making sure a change that matters was reviewed, evidenced and traceable. This is that checklist, pointed at a report: twenty control points in five stages.

Synthetic data. Seeded generator, no real company.

Review (4)

• Save as .pbip from day one; treat .pbix as a build output.
• Every model change goes through a pull request, and the TMDL diff shows the exact measure or relationship that changed.
• Branch protection on main requires one approval for any change to a published model.
• git blame names who last touched a measure, and a bad edit reverts in one command.

CI (6)

• The workflow triggers on every pull request.
• It runs on a Windows runner, because the Power BI tooling is Windows-first.
• It installs the free Tabular Editor 2 command-line runner.
• A Best Practice Analyzer rule file flags missing format strings, unused columns, implicit measures, naming drift and wrong cardinality.
• Any non-zero exit fails the build and blocks the merge.
• A value assertion checks that Gross Margin % still reads 42.6 percent for January 2026 on the seeded data.

Publish (3)

• A merge to main publishes through a service principal, not a person clicking Publish.
• The same parameterized project promotes through dev, test and prod workspaces, with the data source swapped by a parameter.
• Confirm the current publish API against Microsoft's docs before wiring it. On Fabric, prefer its native Git integration and deployment pipelines.

Audit trail (3)

• Who changed a number, when and why: the commit message, the PR discussion and the linked issue.
• The required approval, kept as evidence rather than a memory.
• Traceability from a requirement to the change to the release that shipped it.

Data safety (4)

• Commit a seeded generator instead of data, so anyone can regenerate identical CSVs.
• A validation script exits non-zero on broken referential integrity, gaps in the date table, duplicate keys or a failed tie-out.
• Ignore the analysis cache (cache.abf), .pbi/localSettings.json, credentials and .env files.
• Read git status and the staged diff before every commit. No ignore rule catches a customer name in a screenshot.

The 42.6 percent in the CI gate is the January margin reconciled across Databricks, SQL and DAX in one margin bridge, three tools.