An agent is asked how many active customers the retail division had last quarter. It finds a column called acct_status, reads the description a crawler wrote for it, and answers. That description was inferred from the column name, a sample of values, and the queries that touched it. Nobody in Finance has seen it. The answer is plausible. It is also high, because the wealth division uses a different window and the description never said so.
Inferred metadata is a description a system guessed from evidence. SME-approved metadata is a record a named subject-matter expert confirmed, with a date, a version, and a scope. Both describe the same column. Only one of them is something the business agreed. Ground truth is the name for the second kind. It is what an AI system needs when being wrong carries a cost.
Ground truth is an organization's expert-approved, authoritative account of what its data means. It is held in a form a machine can read, and it names who approved it, when, and until when.
In a catalog the two look identical: a sentence in a description field. The difference is everything around the sentence: how it got there, who stands behind it, and what a machine can do with it.
| Dimension | Inferred metadata | SME-approved metadata |
|---|---|---|
| How it is produced | A model reads column names, data profiles, query history, and nearby descriptions, then writes a draft | A person accountable for the domain confirms, edits, or rejects a proposal |
| What it asserts | A probable meaning | The meaning the business agreed, with its scope |
| Who is accountable | Nobody. The activity log says a model wrote it | A named approver, with a successor on file |
| When it was last right | Unknown. It matched the evidence when the evidence was read | A review date and a version number on the record |
| When it is wrong | The error is silent and looks like every correct answer | Traceable: someone approved it, and the approval can be revisited |
| What a consumer can do | Read it | Cite it, check its date, refuse it when expired, and escalate when none exists |
Inferred metadata is the fastest way to put a first description on fifty thousand undocumented columns. In our post Context Management: The Complete Guide we state that inference is the supply line into ratification, and that the mistake is treating one as a substitute for the other. This piece is about what happens when the substitution is made.
Suppose the descriptions a model writes are right 85% of the time. For a data discovery search that is a good result. A person reads the description, compares it with the column, and corrects the wrong ones as they go. That correction was free. The catalog's accuracy was never the only safeguard.
An agent removes the correction. It reads the description, treats it as true, and acts. The 15% is now spread across every answer with no mark on it. Nobody can tell in advance which answers fall into it. Whether that is acceptable depends on the cost of a wrong answer, which varies by use.
| Use | What a wrong description costs | Verdict |
|---|---|---|
| Discovery search by an analyst | A minute, then a correction | Inferred is fine |
| First draft of a dashboard | A review comment | Inferred is fine |
| Agent reconciling a regulatory report | A restatement, a refiling, a fine | Approved only |
| Agent setting a credit limit | A loss, or a declined customer with no reason on file | Approved only |
| Clinical or safety query | Harm, then a question about who signed off | Approved only, with the approver on the record |
| Any answer an auditor will ask about | The question "who decided this meant that", with no answer | Approved only |
85% accurate isn't good enough for the bottom four rows. Ground truth is what those rows require, and it is a small fraction of the estate. In our post Context Management: The Complete Guide we state that the authoritative slice is consistently under 5% of all data in TopQuadrant customer engagements. The whole catalog does not need an expert's signature. The slice that carries consequences does.
Regulation is starting to say the same thing. Article 14 of the EU AI Act requires that the people overseeing a high-risk system can "disregard, override or reverse the output". Overriding an answer requires seeing what it was resolved against. An inferred description with no approver gives the overseer nothing to check.
The mechanism is the same across the catalog tools that offer it, so it is worth describing plainly.
Steps one to three are the "AI proposes" half of the workflow we recommend, and a proposal beats an empty field. The problem is what steps four and five leave out.
So enrichment raises coverage and leaves the question of truth where it was. It is the right first step and the wrong last one.
Four fields turn a description into ground truth, and together they change what the consumer of the record can do.
| Field | What it holds | What the agent can now do |
|---|---|---|
| Named approver | A person accountable for the domain, with a successor on file | Cite who stands behind the meaning when a human asks |
| Review date | When it was confirmed and when it is due again | Tell a current definition from a stale one |
| Version | A number, with the previous versions and the reason for each change | Explain why last year's report and this year's disagree |
| Expiry rule | What happens when the review date passes | Refuse to resolve against the record and escalate to the steward |
Refusal is the one that matters most. An agent that resolves "active customer" against a version-three definition approved in March can say so. An agent that finds the review date has passed can decline to answer and route the question to the person on the record. A description with no date supports none of this.
Two more fields make the record testable. Provenance written in W3C PROV-O records who produced the definition, from what, and when, as data the agent can traverse. A SHACL shape turns the definition into a check that data passes or fails on every load. We walk through a full governed record, field by field, in our post What Is Metadata Management?. The four fields here are the part of that record an agent acts on.
The objection to approval is cost: fifty thousand columns and ten stewards will never meet. The answer is scope and workflow.
Our post Metadata Quality: How to Measure It and What to Do About It covers the metrics, starting with the share of consequential terms that hold a current approval. Our post Active vs Passive Metadata: Why a Well-Organized Catalog Still Goes Unused covers why an approved record that never reaches the tools is a policy document. Our post Metadata Management for AI: Why Machine-Readability Beats Catalog Structure covers the form a record needs for an agent to read the four fields at all. Our post What Is Data Governance for AI? covers who owns a definition in the first place.
TQ Data Foundation is built for the workflow above. AI proposes, your experts confirm: a model drafts definitions and mappings, and a named steward approves them onto a record that carries the approver, the date, and the version. SHACL validation turns each approved definition into a test the data runs on load. The record is held on open standards that TopQuadrant helped author, SHACL and RDF, so you can export all of it with its history. We maintain 300 reusable standards-based models, so the first approval starts from governed structure. Our metadata management use case page shows how TQ Data Foundation does this in practice.
Your catalog stays. It syncs from the approved record. What changes is that the meaning an agent resolves against has a name on it. Inferred metadata gets you to a description. Ground truth gets you to an answer someone will stand behind. Decide which questions need the second, and build it there first.