Data governance
7 min

Ground Truth: SME-Approved vs Inferred Metadata

Key takeaways

  • Inferred metadata is a description a model guessed from column names, data profiles, and query history. SME-approved metadata is a record a named expert confirmed, with a review date, a version, and a scope. Only the second is ground truth.
  • For discovery, an inferred description is enough, because a person corrects the wrong ones as they read. An agent removes that correction, so for answers that carry a cost the unmarked 15% is the whole problem.
  • AI enrichment raises coverage and leaves the question of truth where it was. The accept click produces a string on an asset, with no scope, no review date, no version, and no expiry.
  • Four fields change what an agent can do: a named approver to cite, a review date to check, a version to explain a change, and an expiry rule that lets it refuse a stale record and escalate.
  • Scope approval to the terms behind consequential answers, approve at the concept, let AI propose and your experts confirm, then sync the catalog from the approved record.
In this article
Share

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.

What is the difference between inferred and SME-approved metadata?

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.

Six ways the two records differ

DimensionInferred metadataSME-approved metadata
How it is producedA model reads column names, data profiles, query history, and nearby descriptions, then writes a draftA person accountable for the domain confirms, edits, or rejects a proposal
What it assertsA probable meaningThe meaning the business agreed, with its scope
Who is accountableNobody. The activity log says a model wrote itA named approver, with a successor on file
When it was last rightUnknown. It matched the evidence when the evidence was readA review date and a version number on the record
When it is wrongThe error is silent and looks like every correct answerTraceable: someone approved it, and the approval can be revisited
What a consumer can doRead itCite 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.

Why is 85% accurate not good enough?

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.

The cost of a wrong description, by use

UseWhat a wrong description costsVerdict
Discovery search by an analystA minute, then a correctionInferred is fine
First draft of a dashboardA review commentInferred is fine
Agent reconciling a regulatory reportA restatement, a refiling, a fineApproved only
Agent setting a credit limitA loss, or a declined customer with no reason on fileApproved only
Clinical or safety queryHarm, then a question about who signed offApproved only, with the approver on the record
Any answer an auditor will ask aboutThe question "who decided this meant that", with no answerApproved 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.

How does AI enrichment produce inferred metadata, and where does it stop?

The mechanism is the same across the catalog tools that offer it, so it is worth describing plainly.

What an enrichment agent does, step by step

  1. Reads the evidence. Column schema, a data profile, query history, linked glossary terms, and the descriptions of sibling columns.
  2. Writes a draft. One description per asset: this table, this column.
  3. Offers it to a reviewer. The reviewer accepts, edits, or rejects. Some tools attach a confidence score and a preview.
  4. Records the acceptance. The activity log notes that a model wrote it and a person accepted it.
  5. Counts coverage. The asset now has a description, so the coverage number goes up.

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.

Three things the accept click does not produce

  • A record of meaning. The output is a string on an asset. The concept behind it may resolve to forty columns in twelve systems. They get forty separate strings and forty clicks, or one click and thirty-nine inferences.
  • A date and a scope. The log says who clicked and when. It does not say that the definition holds for retail and excludes wealth, or that it should be confirmed again in twelve months.
  • A distinction a machine can read. Coverage counts a description from any source as a description. A model-written sentence and a steward-written one look the same to the next agent that reads them. A rating of how well a sentence is written says nothing about whether it is the meaning the business agreed.

So enrichment raises coverage and leaves the question of truth where it was. It is the right first step and the wrong last one.

What does an expert-approved record change for an AI system?

Four fields turn a description into ground truth, and together they change what the consumer of the record can do.

The four fields, and what each changes

FieldWhat it holdsWhat the agent can now do
Named approverA person accountable for the domain, with a successor on fileCite who stands behind the meaning when a human asks
Review dateWhen it was confirmed and when it is due againTell a current definition from a stale one
VersionA number, with the previous versions and the reason for each changeExplain why last year's report and this year's disagree
Expiry ruleWhat happens when the review date passesRefuse 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.

How do you build ground truth without approving everything?

The objection to approval is cost: fifty thousand columns and ten stewards will never meet. The answer is scope and workflow.

AI proposes, your experts confirm: the workflow in order

  1. Scope to consequence. List the questions an agent will answer where a wrong answer costs something. The terms those questions depend on are the authoritative slice. Start there.
  2. Approve at the concept. One definition of "active customer", owned by Finance, with its scope. The forty columns map to it. Each mapping is proposed by a model with a score and confirmed by the steward.
  3. Let automation do the reading. Drafts of definitions, candidate mappings, suggested stewards. Every proposal arrives in a queue with the evidence it was drawn from.
  4. Confirm, with a cadence. The steward approves, edits, or rejects. The approval writes the approver, the date, the version, and the review interval onto the record.
  5. Serve the record with the fields attached. The agent reads the definition, the approver, the date, and the version through one interface, and refuses the record on expiry.
  6. Sync the catalog from it. The approved definition flows down into the catalog descriptions it governs. Your catalog stays, and it shows approved meaning instead of inventing it.

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.

Where does TopQuadrant fit?

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.

Give your AI the context it's been missing

See how the TQ Data Foundation turns your enterprise knowledge into trusted, Al-ready context.