10 Principles of MetadataOps

The principles behind the practice, one line each, with a sentence on why. Built to be printed and argued with.

Practitioner, Team Lead

Everything in this Kit reduces to ten lines. Each one is a rule you can apply on Monday without buying anything, and each one was learned by watching the alternative fail.

Print them. The pages that follow are the argument for each.

1. The estate is a product with three users

Reports, agents, and auditors consume the same supply. Scope the work to all three, because the one you leave out will fund a shadow copy within a year.

2. Metadata gets deployed

A definition change gets proposed, reviewed, promoted, and reversed like any other artifact in the stack. The forty-second edit in a production interface is the failure the practice exists to end.

3. Score the description, and hand the values to data quality

Coverage, resolution, derivation, freshness, governance state. Bad values, null rates, and failed loads belong to someone else, and saying so out loud is what keeps the score credible when an engineer tests it.

4. State the denominator

Coverage without a denominator is a feeling. "Improved" sits a level below "61% to 84%, and here is what's still dark."

5. A definition attached to nothing is a glossary entry

Modeled means bound to a column, a table, or a metric, with a named owner who knows they own it. Prose attached to a table name loses the first disagreement it meets.

6. Fix in dependency order

Harvest, then Model, then Version, then Compile, then Verify. Working on stage four while stage two is incomplete looks productive for two quarters and delivers four definitions of revenue into six tools.

7. Business logic in a prompt is a definition that escaped

If a retrieval configuration contains a formula, one word now has two owners in two systems with no shared history. Six months later they disagree, and every question about last quarter has two answers.

8. Supply and retrieval are different jobs

Context engineering decides what reaches an agent's window. MetadataOps decides what exists to be retrieved. One is upstream in sequence; neither is upstream in importance.

9. Evidence beats reconstruction

An audit answer is a query against a change history that someone other than the author can run. Anything assembled by hand this quarter gets assembled again next quarter, by the same people, at a worse time.

10. Three metrics, all five stages, one quarter

Depth on three metrics produces a working reference and a score you can move. Partial coverage across the whole estate produces a vendor shortlist and a spreadsheet of terms.

The test

When a decision in the practice can't be traced back to one of these ten, it is usually a tool preference stated as a principle. Tool preferences are fine to hold. They should not set the roadmap.

Written by the team at MetaKarta and released under CC BY 4.0: free to use, adapt, and republish with attribution.