A metadata management framework is the set of decisions that makes metadata management repeatable. It says which metadata is in scope, who owns each definition, and what form a governed record takes. It says how a proposed meaning becomes an approved one, how the approved record reaches the systems that use it, and how you know the whole thing is working. The framework in this post has four stages. Each stage is defined by what an AI system can do with your metadata once you are there. A self-assessment table places one domain of your organization in about ten minutes.
Published frameworks for metadata tend to share one shape: a ladder of automation. The bottom rung is an inventory. The top rung is metadata that updates itself, and the word for the top is usually "active" or "autonomous". That ladder answers how fast your metadata moves. It never asks whether the meaning at the top was approved by anyone. The four stages below put that question in the middle, because an agent needs it answered before it acts.
The four stages are Collected, Described, Attested, and Served. Your stage is the highest one at which every statement in the self-assessment is still true. Your next job is the first false statement after it.
A framework is different from a strategy and from a tool. The strategy says why the organization is doing this. The tool is where the work happens. The framework sits between them and holds the decisions both depend on.
| Component | The decision it holds | The test that it exists |
|---|---|---|
| Scope | Which domains, systems, and terms are in the program, and in what order | A written list of the terms whose wrong meaning costs something |
| Roles | Who proposes, who approves, who consumes, and who inherits when the approver leaves | Every governed term resolves to a named steward with a successor on file |
| Standards | The form a record takes, so every tool reads the same thing | Records are held in RDF, SKOS, OWL, and SHACL, and export in full |
| The record | The fields a governed definition must carry before it counts | Definition, scope, steward, status, dates, version, mappings, validation rule, and provenance |
| Workflow | How a proposed meaning becomes an approved one, and when it is reviewed again | A queue where automation proposes and a person confirms, with a review interval on every approval |
| Serving | How consumers, including agents, get the approved record | One interface that returns the definition with its approver, date, and version attached |
| Measures | How you know the program is working | The share of consequential terms with a current approval, and the age of the oldest expired one |
In our post What Is Metadata Management? we state that the discipline has four jobs: collect, define, govern, and serve. The components above make those four jobs repeatable across domains. The stages below are what the organization looks like as each job gets done properly.
The stages are cumulative. Each adds one thing a machine could not do before. The names describe the state of the record, since the record is what an agent reads.
| Stage | What the organization can say | What the record holds | What an AI system can do with it |
|---|---|---|---|
| 1. Collected | We know what data exists and where it lives | Technical metadata harvested by scanners: schemas, tables, columns, types, refresh times | Find an asset. Infer everything else |
| 2. Described | We know what most of it is called and roughly what it means | Descriptions, glossary terms, tags, and inferred lineage, usually as prose, with nothing that separates a guess from a decision | Read a plausible meaning, with no way to tell whether it is approved or expired |
| 3. Attested | For the terms that matter, we know who approved the meaning and when | A structured definition with scope, a named steward, status, dates, a version, column mappings, and a validation rule | Resolve a term to ground truth, cite the approver, check a value against the rule, and refuse a record whose review date has passed |
| 4. Served | Every consumer reads the same approved record at the moment it asks | Everything in stage three, exposed through one interface with provenance attached, and the catalog syncing from it | Act on the record in production, show its basis to a person, and leave an audit trail a regulator can query |
Attested, or SME-approved, metadata is the stage that published ladders leave out. Our post Ground Truth: SME-Approved vs Inferred Metadata covers the difference between a description a crawler guessed and one a named expert confirmed. Stage two can be reached by buying a catalog and running its scanners. Stage three cannot. It requires a decision about who owns each meaning, and a record that carries the decision.
Stage four is where the framework proves it was real. Our post Active vs Passive Metadata: Why a Well-Organized Catalog Still Goes Unused argues that movement alone is the wrong measure of maturity. Stage four is movement of the right thing: the approved record, with its approver and its date attached.
The self-assessment below has twelve statements, three per stage. Mark each one true or false for one domain, since stages vary by domain and an average hides the gap. Your stage is the highest one at which all three of its statements are true.
| # | Statement | Stage |
|---|---|---|
| 1 | We can list every data asset in scope and the system that holds it | 1. Collected |
| 2 | Technical metadata for those assets is harvested automatically and refreshed on a schedule | 1. Collected |
| 3 | A new asset appears in the inventory without anyone filing a ticket | 1. Collected |
| 4 | Most assets in scope carry a description a person can read | 2. Described |
| 5 | A business glossary exists and its terms are linked to at least some assets | 2. Described |
| 6 | Lineage can be drawn for the assets that feed regulatory or financial reports | 2. Described |
| 7 | Every term whose wrong meaning carries a cost has one approved definition, a named steward, and a review date | 3. Attested |
| 8 | The approved definition is structured data, with mappings to physical columns and a rule that tests values against it | 3. Attested |
| 9 | A field on the record separates approved definitions from descriptions, so a consumer can decline the unapproved one | 3. Attested |
| 10 | An agent or application can request the approved record and receive it with its approver, date, and version attached | 4. Served |
| 11 | Our catalogs display definitions synced from the governed record and hold no competing version | 4. Served |
| 12 | "Who signed this off and when" is answered by a query, for any term, including past versions | 4. Served |
If seven is your first false statement, you are on the stage-two plateau. Our post Metadata Quality: How to Measure It and What to Do About It turns statements seven to twelve into metrics.
Stage two is where a catalog's own scores look best and an agent's answers look worst. Coverage is high. Descriptions exist. Lineage renders. Nothing in any of it is approved. The three things that stop a program there are all decisions, and automation cannot make them.
Each transition is one piece of work with a finish line. The order matters, because each stage produces the input the next one needs.
Our post Metadata Management Best Practices for AI-Ready Data sets out the practices that keep each stage from decaying, and we leave them there. The framework decides what the stages are and how you know which one you are at. The practices are how you stay.
A metadata framework is one of several a data organization runs. Each one has a job.
TopQuadrant is built for the move from stage two to stage four. Your catalog stays and keeps its discovery role. TQ Data Foundation sits above it as the governed source of meaning the catalog syncs from. Our metadata management use case page shows how TQ Data Foundation does this in practice.
In TQ Data Foundation, AI proposes definitions and mappings, and your experts confirm them. Nothing becomes authoritative without a person. Each approved record carries a named steward, a date it was last confirmed true, and a version. SHACL validation turns the definition into a test the data runs on load. Everything is held in RDF, SKOS, OWL, and SHACL and exports in full at any time. TopQuadrant helped author SHACL and RDF. We maintain 300 reusable standards-based models, so stage three starts from governed structure. The same governed record an analyst reads is what an agent resolves against, through the same interface. That is stage four.
Run the self-assessment on one domain this week. Find the first false statement. That is where TopQuadrant starts.