A quality label becomes useful only when a reader can see what was measured, how it was measured, whose judgement it records and what the result can support
The W3C Data Quality Vocabulary extends dataset descriptions with structured quality evidence. It does not declare one universal score; it helps publishers and users describe measurements, annotations, policies, standards and provenance precisely enough to inspect.
Data quality is always quality for a purpose. A dataset can be current but incomplete, complete but inaccessible, or accurate for one task and unsuitable for another. DQV keeps these claims separate by distinguishing dimensions, metrics and individual measurements.
The vocabulary also makes room for human annotations, certificates, policies and standards. That breadth matters: a number without a method is weak evidence, while a certificate or opinion should not be mistaken for a fresh measurement.
Chapter contents
- 1. Separate a quality dimension from a metric and a measurement
- 2. Define a metric before publishing its value
- 3. Attach measurements to the correct dataset or distribution
- 4. Record parameters, units and observation windows
- 5. Publish provenance for quality metadata and individual results
- 6. Worked example: trace one quality claim from question to evidence
- 7. Use annotations for feedback, reviews and certificates
- 8. Represent quality policies and standards without inventing conformance
- 9. Keep quality classifications separate from raw measurements
- 10. Validate DQV graphs and their evidence links
- 11. Publish quality evidence so people can discover and compare it
- 12. Monitor, correct and retire quality statements over time
Choose a reading route
- Route 1: Model quality — dimensions, metrics and measurements
- Route 2: Record context — parameters, provenance and derivation
- Route 3: Work an assessment — dataset-to-evidence trace
- Route 4: Express judgement — annotations, policies and conformance
- Route 5: Govern change — publishing, monitoring and repair
How to use this guide
Start with the chapters that match your immediate decision, then return to the worked example and audit chapters. Each chapter separates a model from the evidence needed to use it responsibly. A claim is not strengthened merely by being technical; it is strengthened when assumptions, observations, alternatives and limits can be inspected.
| Layer | Question | Minimum record |
|---|---|---|
| Concept | What exactly is being represented? | Definitions and boundary |
| Method | How will the reasoning be carried out? | Steps, inputs and assumptions |
| Transfer | Would it work in a changed setting? | A deliberately altered case |
| Audit | What could prove it wrong? | Evidence, limits and revision trigger |
Chapter 1. Separate a quality dimension from a metric and a measurement
The idea
A dimension names a quality characteristic relevant to a user, such as availability or completeness. A metric defines how that characteristic is assessed. A measurement applies the metric to a dataset or distribution and records a value.
A disciplined method
Write the dimension in user language, define an operational metric, then create a measurement that identifies the assessed resource, metric, value, unit or datatype and observation time.
A useful working record makes the sequence visible: first state the question and unit of analysis; next record the observations and their provenance; then apply the stated rule; finally compare the result with plausible alternatives. For this chapter, the decisive choice is whether the statement describes a concern, a procedure or one observed result. Keeping that choice explicit prevents a later reader from mistaking a convenient convention for a fact about the world.
Worked example
Availability is the dimension; a daily HTTP retrieval success rate is the metric; 98.7 per cent during September is the measurement.
Work the example twice. On the first pass, follow the stated assumptions. On the second, reverse one assumption, remove one piece of evidence and ask whether the conclusion, ranking or boundary changes. This small stress test matters because polished outputs can hide sensitivity. The objective is not to force certainty; it is to locate the evidence on which the conclusion actually turns.
Changed-context transfer check
Replace file retrieval with an authenticated API and redesign the metric rather than reusing a procedure that no longer matches the access route.
A transfer succeeds only when the structure of the problem survives the change. Match roles, dependencies and measurement conditions rather than surface vocabulary. Record what was preserved, what changed, and why the method still applies—or why it does not. If the new context changes the data-generating process or the people bearing the burden, re-specify the model instead of importing the old answer.
Limits, evidence and revision
Limit: Dimensions and metrics are not universal facts. Different communities can organise dimensions differently and choose different procedures for the same concern.
Evidence to retain: dimension identifier, metric definition, assessed resource, value, datatype, unit, observation window and assessor
Common misconception: A quality dimension already tells a machine how to calculate a score.
A gold-standard conclusion is therefore conditional. It should name the observation period, excluded cases, uncertainty and the event that would trigger review. Negative findings and unresolved contradictions belong in the record. They protect future readers from repeating a hidden assumption and make correction possible without reconstructing the entire analysis from memory.
Practical evidence check
Can another reader recover the input, reproduce the intermediate decision and identify at least one observation that would change the result? If not, the chapter has produced an explanation but not yet an auditable finding.
Chapter 2. Define a metric before publishing its value
The idea
A metric is an executable or at least reproducible account of how an indicator becomes a result. The label completeness score is too vague unless the denominator, eligible records, missing-value rules and aggregation are defined.
A disciplined method
Publish the procedure, inputs, parameters, software or query, inclusion rules and interpretation. Give the metric a stable identifier so measurements across versions can refer to the same method.
A useful working record makes the sequence visible: first state the question and unit of analysis; next record the observations and their provenance; then apply the stated rule; finally compare the result with plausible alternatives. For this chapter, the decisive choice is whether another assessor could obtain a comparable result from the same inputs. Keeping that choice explicit prevents a later reader from mistaking a convenient convention for a fact about the world.
Worked example
A metric counts records with valid geographic coordinates divided by records required to have coordinates, excluding records explicitly outside the geographic scope.
Work the example twice. On the first pass, follow the stated assumptions. On the second, reverse one assumption, remove one piece of evidence and ask whether the conclusion, ranking or boundary changes. This small stress test matters because polished outputs can hide sensitivity. The objective is not to force certainty; it is to locate the evidence on which the conclusion actually turns.
Changed-context transfer check
Introduce multiple coordinate systems and decide whether syntactic validity, transformation success and positional accuracy require separate metrics.
A transfer succeeds only when the structure of the problem survives the change. Match roles, dependencies and measurement conditions rather than surface vocabulary. Record what was preserved, what changed, and why the method still applies—or why it does not. If the new context changes the data-generating process or the people bearing the burden, re-specify the model instead of importing the old answer.
Limits, evidence and revision
Limit: Perfect reproducibility may be limited by proprietary tools, changing services or stochastic methods; document those limits and preserve the strongest available evidence.
Evidence to retain: metric specification, code or query, parameters, test fixtures, software version and expected examples
Common misconception: A percentage is self-explanatory because its range is zero to one hundred.
A gold-standard conclusion is therefore conditional. It should name the observation period, excluded cases, uncertainty and the event that would trigger review. Negative findings and unresolved contradictions belong in the record. They protect future readers from repeating a hidden assumption and make correction possible without reconstructing the entire analysis from memory.
Practical evidence check
Can another reader recover the input, reproduce the intermediate decision and identify at least one observation that would change the result? If not, the chapter has produced an explanation but not yet an auditable finding.
Chapter 3. Attach measurements to the correct dataset or distribution
The idea
Quality can differ across the conceptual dataset, one file distribution, an API service or the metadata record. Attaching every result to the dataset can hide that only one representation was tested.
A disciplined method
Identify the assessed resource explicitly, use stable URIs and state whether the measurement concerns content, delivery, metadata or a service response.
A useful working record makes the sequence visible: first state the question and unit of analysis; next record the observations and their provenance; then apply the stated rule; finally compare the result with plausible alternatives. For this chapter, the decisive choice is which object was actually observed and what broader claim is justified. Keeping that choice explicit prevents a later reader from mistaking a convenient convention for a fact about the world.
Worked example
A checksum and parse-success measurement applies to the CSV distribution, while coverage completeness applies to the underlying dataset after all records are considered.
Work the example twice. On the first pass, follow the stated assumptions. On the second, reverse one assumption, remove one piece of evidence and ask whether the conclusion, ranking or boundary changes. This small stress test matters because polished outputs can hide sensitivity. The objective is not to force certainty; it is to locate the evidence on which the conclusion actually turns.
Changed-context transfer check
Publish the same data as CSV and Parquet and compare serialisation-specific measurements without duplicating dataset-level judgements.
A transfer succeeds only when the structure of the problem survives the change. Match roles, dependencies and measurement conditions rather than surface vocabulary. Record what was preserved, what changed, and why the method still applies—or why it does not. If the new context changes the data-generating process or the people bearing the burden, re-specify the model instead of importing the old answer.
Limits, evidence and revision
Limit: Evidence about one sample, distribution or access path does not automatically generalise to every version and service.
Evidence to retain: resource URI, version, distribution, access path, sample, assessment scope and justified inference
Common misconception: If two files represent one dataset, they must share every quality result.
A gold-standard conclusion is therefore conditional. It should name the observation period, excluded cases, uncertainty and the event that would trigger review. Negative findings and unresolved contradictions belong in the record. They protect future readers from repeating a hidden assumption and make correction possible without reconstructing the entire analysis from memory.
Practical evidence check
Can another reader recover the input, reproduce the intermediate decision and identify at least one observation that would change the result? If not, the chapter has produced an explanation but not yet an auditable finding.
Chapter 4. Record parameters, units and observation windows
The idea
Many quality metrics depend on thresholds, reference sets, geographic bounds, languages, sampling intervals or tolerances. Values without these parameters invite false comparison.
A disciplined method
Model parameters alongside the metric or measurement, use recognised units where available, state time windows and preserve the reference dataset or rule version.
A useful working record makes the sequence visible: first state the question and unit of analysis; next record the observations and their provenance; then apply the stated rule; finally compare the result with plausible alternatives. For this chapter, the decisive choice is which hidden setting could change the result while the metric name stays constant. Keeping that choice explicit prevents a later reader from mistaking a convenient convention for a fact about the world.
Worked example
A positional-accuracy measurement records a five-metre tolerance, the reference map edition and the sample selection date.
Work the example twice. On the first pass, follow the stated assumptions. On the second, reverse one assumption, remove one piece of evidence and ask whether the conclusion, ranking or boundary changes. This small stress test matters because polished outputs can hide sensitivity. The objective is not to force certainty; it is to locate the evidence on which the conclusion actually turns.
Changed-context transfer check
Tighten the tolerance to one metre and publish a new measurement rather than overwriting the earlier result.
A transfer succeeds only when the structure of the problem survives the change. Match roles, dependencies and measurement conditions rather than surface vocabulary. Record what was preserved, what changed, and why the method still applies—or why it does not. If the new context changes the data-generating process or the people bearing the burden, re-specify the model instead of importing the old answer.
Limits, evidence and revision
Limit: DQV offers a vocabulary framework, not a complete parameter-validation language; SHACL or application rules may be needed.
Evidence to retain: parameters, units, thresholds, reference version, temporal coverage, sample design and validation report
Common misconception: Measurements are comparable whenever they use the same metric label.
A gold-standard conclusion is therefore conditional. It should name the observation period, excluded cases, uncertainty and the event that would trigger review. Negative findings and unresolved contradictions belong in the record. They protect future readers from repeating a hidden assumption and make correction possible without reconstructing the entire analysis from memory.
Practical evidence check
Can another reader recover the input, reproduce the intermediate decision and identify at least one observation that would change the result? If not, the chapter has produced an explanation but not yet an auditable finding.
Chapter 5. Publish provenance for quality metadata and individual results
The idea
Quality evidence has its own lifecycle. Readers need to know who assessed, when, using which activity and inputs, and whether a result was derived from another result. DQV deliberately works with PROV-O for this purpose.
A disciplined method
Describe the assessing agent, activity, inputs, software, generated measurement, timestamps and derivation links. Preserve enough evidence to distinguish publisher claims from independent assessments.
A useful working record makes the sequence visible: first state the question and unit of analysis; next record the observations and their provenance; then apply the stated rule; finally compare the result with plausible alternatives. For this chapter, the decisive choice is who produced the statement and which process supports it. Keeping that choice explicit prevents a later reader from mistaking a convenient convention for a fact about the world.
Worked example
A portal publishes a completeness measurement generated by a named validation run, while a consumer group publishes a separate annotation derived from its field test.
Work the example twice. On the first pass, follow the stated assumptions. On the second, reverse one assumption, remove one piece of evidence and ask whether the conclusion, ranking or boundary changes. This small stress test matters because polished outputs can hide sensitivity. The objective is not to force certainty; it is to locate the evidence on which the conclusion actually turns.
Changed-context transfer check
Re-run the assessment after a source-data correction and connect the new result to the prior measurement without treating them as identical.
A transfer succeeds only when the structure of the problem survives the change. Match roles, dependencies and measurement conditions rather than surface vocabulary. Record what was preserved, what changed, and why the method still applies—or why it does not. If the new context changes the data-generating process or the people bearing the burden, re-specify the model instead of importing the old answer.
Limits, evidence and revision
Limit: Provenance can show how a claim arose but does not prove the method was sound, impartial or lawful.
Evidence to retain: agent, activity, inputs, plan, software, timestamps, outputs, derivation and responsibility
Common misconception: A provenance trail makes the quality result objectively correct.
A gold-standard conclusion is therefore conditional. It should name the observation period, excluded cases, uncertainty and the event that would trigger review. Negative findings and unresolved contradictions belong in the record. They protect future readers from repeating a hidden assumption and make correction possible without reconstructing the entire analysis from memory.
Practical evidence check
Can another reader recover the input, reproduce the intermediate decision and identify at least one observation that would change the result? If not, the chapter has produced an explanation but not yet an auditable finding.
Chapter 6. Worked example: trace one quality claim from question to evidence
The idea
A complete assessment begins with a use question, selects a dimension, defines a metric, executes it against a named resource and publishes the result with provenance and limits.
A disciplined method
Start with the intended decision, define eligibility and failure, freeze inputs, run the metric, inspect exceptions, publish the DQV measurement and write a plain-language interpretation.
A useful working record makes the sequence visible: first state the question and unit of analysis; next record the observations and their provenance; then apply the stated rule; finally compare the result with plausible alternatives. For this chapter, the decisive choice is where a user concern becomes a reproducible and bounded claim. Keeping that choice explicit prevents a later reader from mistaking a convenient convention for a fact about the world.
Worked example
A fictional bus-stop dataset is assessed for coordinate completeness. Of 1,200 eligible stops, 1,188 contain valid coordinates under a published validation rule, yielding 0.99 for that release. Twelve exceptions remain listed for repair.
Work the example twice. On the first pass, follow the stated assumptions. On the second, reverse one assumption, remove one piece of evidence and ask whether the conclusion, ranking or boundary changes. This small stress test matters because polished outputs can hide sensitivity. The objective is not to force certainty; it is to locate the evidence on which the conclusion actually turns.
Changed-context transfer check
Change the use case from route planning to wheelchair-accessible navigation and add dimensions for accessibility attributes and path connectivity.
A transfer succeeds only when the structure of the problem survives the change. Match roles, dependencies and measurement conditions rather than surface vocabulary. Record what was preserved, what changed, and why the method still applies—or why it does not. If the new context changes the data-generating process or the people bearing the burden, re-specify the model instead of importing the old answer.
Limits, evidence and revision
Limit: A high coordinate-completeness score says nothing about positional accuracy, current service status or accessibility.
Evidence to retain: use question, resource version, metric, code, counts, value, exceptions, provenance and interpretation
Common misconception: One high score establishes that the dataset is high quality overall.
A gold-standard conclusion is therefore conditional. It should name the observation period, excluded cases, uncertainty and the event that would trigger review. Negative findings and unresolved contradictions belong in the record. They protect future readers from repeating a hidden assumption and make correction possible without reconstructing the entire analysis from memory.
| Layer | Example record | Check |
|---|---|---|
| Dimension | coordinate completeness | relevant to the use |
| Metric | valid eligible coordinates / eligible stops | rules and denominator explicit |
| Measurement | 0.99 for release 2026-09 | resource and time fixed |
| Interpretation | twelve repair cases remain | claim stays bounded |
Practical evidence check
Can another reader recover the input, reproduce the intermediate decision and identify at least one observation that would change the result? If not, the chapter has produced an explanation but not yet an auditable finding.
Chapter 7. Use annotations for feedback, reviews and certificates
The idea
DQV QualityAnnotation can represent feedback or certificates about a dataset or distribution, commonly using the Web Annotation model. An annotation records a motivated assertion; it is not automatically a quantitative measurement.
A disciplined method
Identify the target, body, motivation, creator and date; distinguish a certificate from a consumer comment; link supporting measurements or standards when they exist.
A useful working record makes the sequence visible: first state the question and unit of analysis; next record the observations and their provenance; then apply the stated rule; finally compare the result with plausible alternatives. For this chapter, the decisive choice is whether the statement is observed evidence, professional judgement, user experience or formal certification. Keeping that choice explicit prevents a later reader from mistaking a convenient convention for a fact about the world.
Worked example
A mapping association annotates a distribution with a certificate and links the assessment policy and supporting measurements. A user separately reports that the download instructions are inaccessible.
Work the example twice. On the first pass, follow the stated assumptions. On the second, reverse one assumption, remove one piece of evidence and ask whether the conclusion, ranking or boundary changes. This small stress test matters because polished outputs can hide sensitivity. The objective is not to force certainty; it is to locate the evidence on which the conclusion actually turns.
Changed-context transfer check
Withdraw the certificate after expiry while preserving the historical annotation and its validity period.
A transfer succeeds only when the structure of the problem survives the change. Match roles, dependencies and measurement conditions rather than surface vocabulary. Record what was preserved, what changed, and why the method still applies—or why it does not. If the new context changes the data-generating process or the people bearing the burden, re-specify the model instead of importing the old answer.
Limits, evidence and revision
Limit: Annotations can be subjective, malicious, outdated or issued under unknown criteria; consumers need provenance and status.
Evidence to retain: annotation target, motivation, body, creator, criteria, dates, status and supporting evidence
Common misconception: Every annotation is a measurement because it says something about quality.
A gold-standard conclusion is therefore conditional. It should name the observation period, excluded cases, uncertainty and the event that would trigger review. Negative findings and unresolved contradictions belong in the record. They protect future readers from repeating a hidden assumption and make correction possible without reconstructing the entire analysis from memory.
Practical evidence check
Can another reader recover the input, reproduce the intermediate decision and identify at least one observation that would change the result? If not, the chapter has produced an explanation but not yet an auditable finding.
Chapter 8. Represent quality policies and standards without inventing conformance
The idea
A quality policy describes commitments, requirements or agreements. A standard can be linked as a conformance target. Neither link alone proves that the assessed resource conforms.
A disciplined method
Publish the policy or standard identifier, scope, version and responsible authority; state how conformance was evaluated; distinguish claimed, tested, partial and failed conformance in visible text and evidence.
A useful working record makes the sequence visible: first state the question and unit of analysis; next record the observations and their provenance; then apply the stated rule; finally compare the result with plausible alternatives. For this chapter, the decisive choice is what test supports the conformance statement and which version was applied. Keeping that choice explicit prevents a later reader from mistaking a convenient convention for a fact about the world.
Worked example
A catalogue links its metadata profile, validation shapes and report. Passing required shapes supports a bounded metadata-conformance statement.
Work the example twice. On the first pass, follow the stated assumptions. On the second, reverse one assumption, remove one piece of evidence and ask whether the conclusion, ranking or boundary changes. This small stress test matters because polished outputs can hide sensitivity. The objective is not to force certainty; it is to locate the evidence on which the conclusion actually turns.
Changed-context transfer check
Add recommended shapes and decide whether warnings change the conformance category or remain advisory.
A transfer succeeds only when the structure of the problem survives the change. Match roles, dependencies and measurement conditions rather than surface vocabulary. Record what was preserved, what changed, and why the method still applies—or why it does not. If the new context changes the data-generating process or the people bearing the burden, re-specify the model instead of importing the old answer.
Limits, evidence and revision
Limit: Standards can be ambiguous, profiles can narrow them, and automated validators rarely test every semantic obligation.
Evidence to retain: standard URI, profile, version, test suite, shapes, report, exceptions, assessor and decision rule
Common misconception: Linking dcterms:conformsTo is enough evidence that every requirement was tested.
A gold-standard conclusion is therefore conditional. It should name the observation period, excluded cases, uncertainty and the event that would trigger review. Negative findings and unresolved contradictions belong in the record. They protect future readers from repeating a hidden assumption and make correction possible without reconstructing the entire analysis from memory.
Practical evidence check
Can another reader recover the input, reproduce the intermediate decision and identify at least one observation that would change the result? If not, the chapter has produced an explanation but not yet an auditable finding.
Chapter 9. Keep quality classifications separate from raw measurements
The idea
A classification translates evidence into a category such as bronze, silver or fit for a named purpose. Its decision rule, evidence set and validity period must remain visible.
A disciplined method
Publish the classification scheme, thresholds, required dimensions, weighting, missing-evidence treatment and assessment date; retain the underlying measurements rather than only the badge.
A useful working record makes the sequence visible: first state the question and unit of analysis; next record the observations and their provenance; then apply the stated rule; finally compare the result with plausible alternatives. For this chapter, the decisive choice is which rule converted several pieces of evidence into the label. Keeping that choice explicit prevents a later reader from mistaking a convenient convention for a fact about the world.
Worked example
A trusted-use classification requires current provenance, licence clarity, ninety-nine per cent schema validity and a manual disclosure review.
Work the example twice. On the first pass, follow the stated assumptions. On the second, reverse one assumption, remove one piece of evidence and ask whether the conclusion, ranking or boundary changes. This small stress test matters because polished outputs can hide sensitivity. The objective is not to force certainty; it is to locate the evidence on which the conclusion actually turns.
Changed-context transfer check
Change one threshold and reclassify old results without pretending the scheme version stayed the same.
A transfer succeeds only when the structure of the problem survives the change. Match roles, dependencies and measurement conditions rather than surface vocabulary. Record what was preserved, what changed, and why the method still applies—or why it does not. If the new context changes the data-generating process or the people bearing the burden, re-specify the model instead of importing the old answer.
Limits, evidence and revision
Limit: Composite classifications can hide trade-offs and encourage gaming; a single label should never erase the evidence profile.
Evidence to retain: scheme version, criteria, weights, missing rules, measurements, result and expiry
Common misconception: A gold badge means the dataset is best for every purpose.
A gold-standard conclusion is therefore conditional. It should name the observation period, excluded cases, uncertainty and the event that would trigger review. Negative findings and unresolved contradictions belong in the record. They protect future readers from repeating a hidden assumption and make correction possible without reconstructing the entire analysis from memory.
Practical evidence check
Can another reader recover the input, reproduce the intermediate decision and identify at least one observation that would change the result? If not, the chapter has produced an explanation but not yet an auditable finding.
Chapter 10. Validate DQV graphs and their evidence links
The idea
A syntactically valid RDF graph can still attach a value to no metric, omit the assessed resource or use incompatible datatypes. Structural validation and semantic review are complementary.
A disciplined method
Use SHACL or equivalent constraints for required relationships and datatypes, test identifiers and links, then manually review metric meaning, derivations and interpretation.
A useful working record makes the sequence visible: first state the question and unit of analysis; next record the observations and their provenance; then apply the stated rule; finally compare the result with plausible alternatives. For this chapter, the decisive choice is whether the graph is complete enough to consume and truthful enough to trust. Keeping that choice explicit prevents a later reader from mistaking a convenient convention for a fact about the world.
Worked example
A validation shape requires each QualityMeasurement to name its metric, assessed resource and value, while a reviewer checks that the metric actually matches the published procedure.
Work the example twice. On the first pass, follow the stated assumptions. On the second, reverse one assumption, remove one piece of evidence and ask whether the conclusion, ranking or boundary changes. This small stress test matters because polished outputs can hide sensitivity. The objective is not to force certainty; it is to locate the evidence on which the conclusion actually turns.
Changed-context transfer check
Allow a textual qualitative value and revise shapes without rejecting legitimate non-numeric assessments.
A transfer succeeds only when the structure of the problem survives the change. Match roles, dependencies and measurement conditions rather than surface vocabulary. Record what was preserved, what changed, and why the method still applies—or why it does not. If the new context changes the data-generating process or the people bearing the burden, re-specify the model instead of importing the old answer.
Limits, evidence and revision
Limit: Shapes enforce declared rules, not scientific validity or fitness for use; over-strict constraints can exclude useful evidence.
Evidence to retain: graph, shapes, validation report, review notes, exceptions and repaired version
Common misconception: Passing RDF validation proves that the quality assessment is sound.
A gold-standard conclusion is therefore conditional. It should name the observation period, excluded cases, uncertainty and the event that would trigger review. Negative findings and unresolved contradictions belong in the record. They protect future readers from repeating a hidden assumption and make correction possible without reconstructing the entire analysis from memory.
Practical evidence check
Can another reader recover the input, reproduce the intermediate decision and identify at least one observation that would change the result? If not, the chapter has produced an explanation but not yet an auditable finding.
Chapter 11. Publish quality evidence so people can discover and compare it
The idea
Quality metadata should be reachable from the dataset record, serialised in documented formats and accompanied by human-readable explanations. Comparisons require aligned scope and methods, not merely aligned property names.
A disciplined method
Link quality metadata from DCAT records, negotiate useful representations, document profiles, keep identifiers stable and present side-by-side evidence only after checking metric and parameter compatibility.
A useful working record makes the sequence visible: first state the question and unit of analysis; next record the observations and their provenance; then apply the stated rule; finally compare the result with plausible alternatives. For this chapter, the decisive choice is whether a consumer can find the evidence and understand what can be compared. Keeping that choice explicit prevents a later reader from mistaking a convenient convention for a fact about the world.
Worked example
A catalogue exposes DQV in RDF and a readable panel that lists dimensions, methods, dates, results and caveats for each distribution.
Work the example twice. On the first pass, follow the stated assumptions. On the second, reverse one assumption, remove one piece of evidence and ask whether the conclusion, ranking or boundary changes. This small stress test matters because polished outputs can hide sensitivity. The objective is not to force certainty; it is to locate the evidence on which the conclusion actually turns.
Changed-context transfer check
Harvest the record into another catalogue and verify that identifiers, profiles, provenance and links survive.
A transfer succeeds only when the structure of the problem survives the change. Match roles, dependencies and measurement conditions rather than surface vocabulary. Record what was preserved, what changed, and why the method still applies—or why it does not. If the new context changes the data-generating process or the people bearing the burden, re-specify the model instead of importing the old answer.
Limits, evidence and revision
Limit: Not every consumer supports DQV, and central portals may flatten local nuance; keep source evidence resolvable.
Evidence to retain: catalogue links, representations, profiles, identifiers, harvest trace, readable view and comparison rules
Common misconception: Two DQV values can be ranked as soon as their numerical datatypes match.
A gold-standard conclusion is therefore conditional. It should name the observation period, excluded cases, uncertainty and the event that would trigger review. Negative findings and unresolved contradictions belong in the record. They protect future readers from repeating a hidden assumption and make correction possible without reconstructing the entire analysis from memory.
Practical evidence check
Can another reader recover the input, reproduce the intermediate decision and identify at least one observation that would change the result? If not, the chapter has produced an explanation but not yet an auditable finding.
Chapter 12. Monitor, correct and retire quality statements over time
The idea
Datasets, metrics, policies and access routes change. A once-valid measurement can become stale even when its URL still resolves.
A disciplined method
Set review or expiry expectations, trigger reassessment on material change, version metric definitions, preserve superseded evidence, publish corrections and assign repair owners.
A useful working record makes the sequence visible: first state the question and unit of analysis; next record the observations and their provenance; then apply the stated rule; finally compare the result with plausible alternatives. For this chapter, the decisive choice is when the evidence no longer supports a current claim. Keeping that choice explicit prevents a later reader from mistaking a convenient convention for a fact about the world.
Worked example
A monthly availability measurement expires after an endpoint migration; the publisher reruns the test and marks the older result as historical.
Work the example twice. On the first pass, follow the stated assumptions. On the second, reverse one assumption, remove one piece of evidence and ask whether the conclusion, ranking or boundary changes. This small stress test matters because polished outputs can hide sensitivity. The objective is not to force certainty; it is to locate the evidence on which the conclusion actually turns.
Changed-context transfer check
Discover that a denominator bug inflated years of scores and publish corrected values, derivation notes and affected decisions.
A transfer succeeds only when the structure of the problem survives the change. Match roles, dependencies and measurement conditions rather than surface vocabulary. Record what was preserved, what changed, and why the method still applies—or why it does not. If the new context changes the data-generating process or the people bearing the burden, re-specify the model instead of importing the old answer.
Limits, evidence and revision
Limit: Historical preservation must respect privacy and legal duties, and not every error can be fully reconstructed.
Evidence to retain: change event, validity period, reassessment, correction notice, versions, affected uses and owner sign-off
Common misconception: Quality metadata remains valid until someone files a complaint.
A gold-standard conclusion is therefore conditional. It should name the observation period, excluded cases, uncertainty and the event that would trigger review. Negative findings and unresolved contradictions belong in the record. They protect future readers from repeating a hidden assumption and make correction possible without reconstructing the entire analysis from memory.
Practical evidence check
Can another reader recover the input, reproduce the intermediate decision and identify at least one observation that would change the result? If not, the chapter has produced an explanation but not yet an auditable finding.
Field workbook
Choose one real or fictional dataset and one distribution. Define two quality dimensions, a metric for each, parameters, a measurement, one annotation and one conformance claim. Add PROV-O evidence, a validation rule, a plain-language interpretation and a changed-context reassessment.
- Write the decision question in one sentence and name the decision owner.
- List observations separately from interpretations, with dates and provenance.
- Record at least two plausible alternatives before choosing a preferred account.
- Run the changed-context test and note which assumptions fail.
- State the limit, confidence and next review trigger in plain language.
| Layer | Evidence | Failure signal |
|---|---|---|
| Meaning | dimension and user purpose | generic score |
| Method | metric, parameters and code | undefined calculation |
| Result | resource, value and window | orphan number |
| Accountability | provenance, annotation and repair | stale unsupported claim |
Continue through the verified ecosystem
- Dataset distribution discovery — DCAT resources and truthful access routes
- Content negotiation by profile — profiles and representation selection
- Web Annotation — targets, bodies and motivation semantics
- PREMIS preservation evidence — events, agents, objects and evidence
- ResourceSync — change, fixity and synchronisation
Frequently asked questions
Does DQV define one universal data-quality score?
No. It provides terms for dimensions, metrics, measurements, annotations, policies, standards and provenance; communities define the relevant frameworks and methods.
What is the difference between a metric and a measurement?
A metric defines the procedure; a measurement is one result from applying that procedure to a named resource.
Can DQV describe qualitative evidence?
Yes. Measurements may be qualitative, and annotations can express feedback, reviews or certificates with provenance.
Does a conformance link prove compliance?
No. The applied standard or profile, test method, report, exceptions and decision rule should support the claim.
Why record provenance for quality metadata?
It shows who assessed what, when, through which activity and inputs, helping readers distinguish and reproduce claims.
Can two DQV scores always be compared?
Only when scope, resource versions, metrics, parameters, units and observation windows are sufficiently aligned.
Authoritative sources and further reading
- W3C: Data Quality Vocabulary — quality dimensions, metrics, measurements, annotations, policies and examples
- W3C: Data Catalog Vocabulary Version 3 — datasets, distributions, catalogues and services
- W3C: PROV-O — provenance entities, activities, agents and derivation
- W3C: Web Annotation Data Model — annotation targets, bodies, motivations and provenance
- W3C: Shapes Constraint Language — structural validation of RDF graphs
This guide explains how to publish and interpret DQV evidence. It does not certify a dataset, prescribe a universal quality framework or replace domain review, statistical validation, legal assessment or user testing.