Classification Binding

Binding as a Reified Assertion

A classification binding is a first-class n-ary assertion (subject, scheme, version, term, role, agent, time, basis), not a bare property value. Reification is required because the same subject-term pair may be asserted repeatedly by different agents, at different times, with different authority and different validity, and each such assertion must remain independently addressable, contestable and retractable.

1. What minimum tuple constitutes one binding assertion, and which components are mandatory versus optional? definition
2. When must a coded value be recorded as a full binding record rather than as an inline coded attribute of the subject? decision
3. May two distinct binding records hold the same subject, scheme and term at the same time, and if so what distinguishes them? identity
4. How is an individual binding addressed and cited by an external system without dereferencing the subject? interoperability

Classification Versus Ontological Typing

Binding a subject to a scheme term is subject indexing or categorisation and carries no entailment that the subject is an instance of a formal class. The SKOS Primer directs subject indexing through dcterms:subject to a skos:Concept and notes that OWL-DL prevents treating SKOS concepts as OWL classes; conflating the two produces unsound inference and unsafe schema migration.

1. Does this binding license any inference about the subject's type, and if not how is that non-entailment recorded? classification
2. Where a scheme term is also published as a formal class, which reading governs the binding? constraint
3. Is a binding to a narrower term to be treated as also binding the subject to broader terms? relationship

Subject and Scope of Application

What a binding attaches to: a whole entity, a specific part or aspect (Web Annotation SpecificResource), a registered metadata item (ISO/IEC 11179 Classifiable_Item), a dataset or an identifiable artefact in a statistical model (SDMX Hierarchy Association). The scope must be explicit because the same term applied to a whole versus a part means different things.

1. At what granularity does this binding apply: whole subject, a named part, a time slice, or an aspect? composition
2. Which subject kinds are eligible for binding under this slot, and which are explicitly excluded? constraint
3. Does the subject correspond to the statistical or classificatory unit the scheme was designed to classify? validation

Binding Identifier and Natural Key

The binding needs both a surrogate identifier and a declared natural key so that repeated ingestion is idempotent. Identity priority is: identifier from the authoritative master system that owns the binding; else a governed global IRI; else a UUID or ULID minted by the adopting Dimension. Dates are never identifiers.

1. Which system is the master of record for this binding's identifier, and what identifier form does it issue? ownership
2. What natural key makes repeated ingestion of the same binding idempotent? identity
3. Which external identifiers of the same binding must be carried alongside the internal one? interoperability

Scheme, Version and Term Reference Identity

A binding must reference the scheme by canonical URI rather than by name, pin the scheme version used at assignment time, and carry the term by its identifier or notation. FHIR requires specified code system URIs to be used in preference to any other identifying mechanism, records Coding.version as the code system version used when choosing the code, and treats Coding.display as a rendering, not identity.

1. What is the canonical URI of the scheme, and which registry governs it? identity
2. Which scheme version was in force when the term was selected, and is the binding version-pinned or version-floating? provenance
3. Is the term carried by URI, by notation, or by both, and how is a notation resolved to a term? interoperability
4. How is a stored display label reconciled when the scheme later changes that label? quality

Display, user text and language

FHIR Coding.display is the representation defined by the system; CodeableConcept.text is the human language representation selected or uttered by the user. Computation on display alone is unsafe. SKOS separates prefLabel, altLabel and hiddenLabel with language tags. DCMI prefers a URI for subject but allows a literal if needed.

1. What display string does the referenced scheme define for this item, in which language, and is it the scheme’s recommended display? definition
2. What text did the user or agent enter as the intended meaning, and does it stand alone without a code? evidence
3. Are preferred labels unique per language within the scheme, and has any process treated display text as a computable identifier? quality
4. If no code is present, is a literal subject or keyword recorded, and is it preferred to be aligned later to a controlled vocabulary URI? interoperability

Binding Slot Declaration

A binding slot names a classification role in a context (subject type, purpose, jurisdiction, process step) and declares the scheme or value space that fills it, mirroring FHIR's ElementDefinition.binding with name, strength, value set reference and description, and DCMI's association of a property with a Vocabulary Encoding Scheme.

1. What is the slot's stable name, the subject type it applies to, and the decision it feeds? definition
2. Which scheme fills this slot, and on what documented basis was it chosen over alternatives? authority
3. Does the permitted scheme vary by jurisdiction, tenant or process step, and how is the applicable variant resolved at runtime? constraint

Permitted Value Space

The exact set of terms admissible in a slot: a whole scheme, an enumerated subset, a computed selection, a restriction to a specific classification level, or a restriction to selectable leaf terms. XKOS levels with notationPattern, SDMX hierarchies over one or more codelists, and FHIR value set expansions all express this restriction.

1. How is the permitted term set defined: by enumeration, by scheme subtree, by level, or by a computed rule? constraint
2. Are non-leaf or non-selectable terms permitted, and at which classification level must the term sit? constraint
3. Is a point-in-time expansion of the permitted set captured, and when is it refreshed? temporal
4. May the slot draw terms from more than one scheme simultaneously? composition

Value set applied to a slot

A terminology binding names the value set whose expansion is the allowed or suggested set of codes for an element. The value-set canonical URL is not the code-system URL. Additional bindings may tighten the main binding in a context.

1. Which value-set canonical URL is bound to this element or slot, and is it version-pinned? identity
2. Is the bound object a value set rather than a code system, and which code systems does that value set draw from? composition
3. Are additional bindings defined for particular usage contexts, and what is each additional binding’s purpose? constraint
4. What human-readable description states the intent of this element binding for implementers? definition

Binding Strength

The conformance obligation attached to a slot. FHIR defines four strengths: required (data SHALL contain a value from the expansion), extensible (SHALL use a value set code if any code applies, otherwise an alternate code or text), preferred (encouraged but not required for conformance) and example (illustrative only). Strength determines validation severity and whether text fallback is legitimate.

1. What binding strength applies to this slot, and what validation severity does each strength map to? requirement
2. Under an extensible binding, what test decides that no code in the value set can apply? decision
3. May a local profile tighten or loosen the inherited strength, and what governance approves that? authority

Multiplicity, Exhaustiveness and Mutual Exclusivity

How many terms may be bound in a slot and whether the term set must be exhaustive and mutually exclusive. The UNSD guidelines state that mutual exclusivity is mandatory for statistical classifications and that a classification should be exhaustive for the values the variable can take for its primary units; multi-label domains deliberately relax exclusivity and must say so explicitly.

1. What minimum and maximum number of terms may be bound in this slot? constraint
2. Is the term set mutually exclusive at the applicable level, and is multi-labelling permitted? constraint
3. When several terms are bound, how is the primary or governing term determined? decision

Applied Term and Binding Role

The applied term plus the role it plays. Web Annotation distinguishes purposes such as tagging, classifying, identifying and assessing; the same term applied for identification, categorisation or evaluation carries different obligations. Post-coordinated expressions (permitted where a code system declares itself compositional) must be flagged because they are not simple term references.

1. What role does this applied term play: classifying, identifying, assessing or tagging? classification
2. Is the applied value a simple term reference or a post-coordinated expression, and does the scheme permit expressions? constraint
3. What qualifiers modify the applied term (negation, suspected, historical, laterality, degree)? constraint

Origin of the Assertion

Whether the term was chosen directly by a person, produced by a rule or model, inherited from a related subject, or derived through a mapping. FHIR's Coding.userSelected marks codings chosen by a user directly off a pick list, and Web Annotation separates creator/created from generator/generated; both distinctions must survive into the binding record because they change how much trust and how much rectification obligation the binding carries.

1. Was this term selected directly by a person, produced automatically, inherited, or derived from a mapping? provenance
2. If produced automatically about a natural person, does the binding feed a decision based solely on automated processing? privacy
3. Which software serialised or generated this binding, distinct from the agent responsible for its content? provenance

Classification role, facet and multiplicity

A binding has a role or facet: topical theme, type/genre, ISO 19115 keyword type, statistical variable coding, or local primary/secondary rank. DCAT allows multiple themes. GSIM statistical classifications used as variable domains are typically mutually exclusive and exhaustive. ISO 19115 topic categories explicitly overlap and the user is asked to pick the most appropriate. These rules conflict and must be recorded as policy on the slot, not assumed globally.

1. What facet or role does this binding play (subject/theme, type/genre, place, discipline, taxon, statistical variable, or other keyword type)? classification
2. Is this the primary classification for the facet, a secondary class, or an unordered additional theme? relationship
3. Does the slot allow multiple simultaneous classes, require mutual exclusivity, or require a single most-appropriate class despite known overlap? constraint
4. If both a type/genre and a subject/theme are recorded, are they separate bindings rather than one overloaded code? decision

Residual and Unclassifiable Handling

Exhaustive schemes achieve exhaustiveness partly through residual 'not elsewhere classified' or 'other' categories. A binding must distinguish a deliberate residual classification from an unknown, a not-applicable, a refused and a not-yet-attempted state, because these have different downstream consequences for aggregation and for follow-up.

1. How are residual, unknown, not-applicable, refused and not-yet-classified states distinguished in the record? state
2. Does a residual or unknown binding create a follow-up obligation, and with what deadline? process
3. Does the scheme claim exhaustive coverage of the subject population, and what happens when it demonstrably does not? quality

Uncoded Text Fallback and Candidate Terms

Under extensible, preferred or example strengths, an implementation may legitimately record an alternate code or free text. FHIR's CodeableConcept.text is defined as the human-language representation as seen, selected or uttered by the user, sitting alongside zero or more codings. Uncoded text is also the primary evidence stream for proposing new scheme terms.

1. Under which strengths is uncoded text a conformant outcome rather than a validation failure? requirement
2. Is the recorded text the user's original wording or a system-generated rendering of a code? provenance
3. How does recurring uncoded text become a proposal for a new scheme term? process

Assignment Actor, Activity and Method

Every binding is generated by an activity attributed to an agent that may have followed a plan. PROV-O supplies wasAttributedTo, wasGeneratedBy, generatedAtTime, used and hadPlan; for classification the plan is typically a coding index, a decision rule set, a general interpretative rule, or a model version. Method identity must be versioned so that a past assignment can be reproduced.

1. Which agent is responsible for this binding's content, and in what capacity did they act? ownership
2. Which coding method, rule set, coding index or model version produced the assignment? process
3. Which inputs did the assignment activity consume, and are they retained for re-derivation? provenance
4. At what instant did the assignment activity complete, and in which time zone was it recorded? temporal

Evidence and Justification

The material and reasoning that support the chosen term: source documents, measurements, applied interpretative rules, and a rationale note. SKOS documentation notes (scopeNote, historyNote, changeNote) show the same need at scheme level; at binding level the rationale is what makes a contested classification defensible.

1. What concrete evidence supports the chosen term, and where is each item held? evidence
2. Which interpretative rules or scheme notes were applied in reaching the term? authority
3. What is the human-readable rationale, and is it sufficient for an independent reviewer to reach the same term? quality

Assignment Authority and Legal Effect

Bindings differ sharply in force. An EU Binding Tariff Information decision is binding on all EU customs administrations and on the holder, is generally valid for three years, may be annulled where it was issued on inaccurate or incomplete information, and may be revoked or cease to be valid following nomenclature change, CJEU rulings or WCO decisions. Most bindings carry no such force; the model must record which regime applies rather than assuming one.

1. Who is entitled to assert a binding in this slot, and what qualification or delegation is required? authority
2. What force does the binding carry: advisory, internally binding, contractually binding, or legally binding on third parties? authority
3. By what route may the binding be challenged, and with what effect on its validity while under challenge? exception
4. What distinguishes annulment, revocation and cessation of validity for this binding? lifecycle

Authority, ownership and jurisdiction

FHIR routes external code-system URIs through HL7 Terminology Authority where listed. Statistical classifications such as ISIC have UN custodianship; regional schemes (NACE, NAICS) apply in particular jurisdictions. ISO/IEC 11179 registration records stewardship of classified items. Ownership of the binding record is the adopting Dimension’s data owner, which is not necessarily the scheme maintainer.

1. Who is the custodial authority of the referenced scheme, and is the URI the one authorised by that authority or by a naming system such as HL7 THO/HTA? authority
2. Which party owns this binding record in the adopting Dimension, and who may update or withdraw it? ownership
3. In which jurisdiction or statistical territory is this classification valid (for example NAICS versus NACE versus ISIC)? spatial
4. Is this assignment an official legal or statistical coding, a community convention, or a local convenience class? authority

Confidence and Uncertainty

Automated and inferred bindings carry uncertainty that must be recorded explicitly, including the score, its scale and semantics, whether it is calibrated, the acceptance threshold applied, and the runner-up candidates considered. No consulted standard mandates a confidence field, so its semantics must be locally defined and declared rather than assumed comparable across systems.

1. What is the confidence value, on what scale, and does it represent a calibrated probability? measurement
2. What acceptance threshold was applied, and what happens to assignments below it? decision
3. Which alternative terms were considered and rejected, and are they retained? evidence

Coding Quality Measurement

Population-level assessment of whether bindings in a slot are correct: verification sampling, independent dual coding and agreement, error rate against a reference standard, and acceptance thresholds. This is the evidence base for claiming that a body of bindings is fit for the decisions it feeds.

1. How is coding accuracy for this slot measured, over what sample and against what reference standard? measurement
2. What accuracy or agreement level is required, and what happens when it is not met? quality
3. How often is quality re-measured, and what triggers an unscheduled measurement? process

Temporal Frames of a Binding

At least four distinct times matter and must not be collapsed: the time of the subject state that justifies the term; the time the assertion was made; the time the assertion was observed, ingested or recorded by this system; and the interval over which the binding is asserted to hold. All are recorded in RFC 3339 with seconds and an explicit offset or Z.

1. Which temporal frames are recorded for this binding, and what does each mean? temporal
2. Over what interval does the binding assert that the subject bears the term? temporal
3. How is the ingestion or observation time recorded separately from the original assertion time on import? provenance
4. May a binding be asserted with retroactive or future validity, and who approves that? constraint

Scheme Version Drift and Migration

Schemes are revised, terms are deprecated, split, merged or retired, and FHIR's CodeSystem.versionNeeded signals that a code system may not commit to concept permanence across versions. XKOS models version succession (follows, supersedes, variant) and correspondences for m-to-n mapping; migration must be an explicit, auditable operation that preserves the original binding rather than overwriting it.

1. How is it detected that a bound term has been deprecated, split, merged or retired in a newer scheme version? state
2. What is the policy for bindings whose term is no longer valid: freeze, migrate, or invalidate? lifecycle
3. When a term splits into several successors, how is the single correct successor chosen? decision
4. How is time-series or reporting continuity preserved across a scheme revision? interoperability

Binding Lifecycle States

A binding moves through declared states such as proposed, asserted, verified, disputed, superseded, expired, revoked and withdrawn. Transitions must be enumerated and guarded; PROV-O's invalidatedAtTime marks the end of an entity's applicability and the BTI regime supplies concrete end states (expiry, annulment, revocation, cessation of validity).

1. What is the closed set of lifecycle states, and what does each mean for a consumer? state
2. Which transitions are permitted, who may trigger each, and what evidence must accompany them? lifecycle
3. How does a state change interact with the validity interval and with already-published downstream data? temporal

Supersession, Correction and Dispute

Distinguishing a correction of an erroneous binding from a legitimate reclassification following a genuine change in the subject, and providing a route for a party to contest a binding. GDPR Article 16 gives data subjects the right to rectification of inaccurate personal data, which for a person-classifying binding is a hard requirement rather than a preference.

1. Is this change a correction of an error or a reclassification following a real change in the subject? decision
2. How is the supersession chain recorded so that a consumer can reconstruct the binding history? provenance
3. How does an affected party register a dispute, and what does the binding look like while disputed? exception
4. Where the subject is a natural person, how is a rectification request received, decided and propagated? privacy

Binding Validation Rules

A binding is validated against its slot specification: the scheme URI must resolve and be permitted, the version must be resolvable, the term must be a member of the permitted value space in that version, level and selectability constraints must hold, cardinality must be satisfied, and the outcome severity must follow the declared binding strength.

1. What is the ordered set of checks a validator runs, and which are blocking? validation
2. What happens when the scheme or version cannot be resolved at validation time? exception
3. Is a binding to a deprecated or retired term a validation error, a warning, or acceptable for historical records? constraint
4. Is the validation outcome retained as evidence, and for how long? evidence

Conflict and Consistency Detection

Detecting contradictions across a set of bindings on the same subject: mutually exclusive terms bound simultaneously where the scheme level declares exclusivity, ancestor and descendant terms bound as if independent, equivalent codings from different schemes that disagree, and duplicate assertions from competing authorities.

1. Which conflict types are detected, and how is each defined precisely? validation
2. When two authorities assert incompatible bindings, which prevails and on what recorded basis? authority
3. How is a cross-scheme disagreement assessed given that mappings may be inexact? interoperability

Mapping-Derived Bindings

A binding may be produced by translating an existing binding through a mapping. SKOS provides exactMatch, closeMatch, broadMatch, narrowMatch and relatedMatch; ISO 25964-2 refines equivalence into exact, inexact and partial and adds hierarchical and associative mappings; XKOS ConceptAssociation supports m-to-n. Derived bindings must be marked as derived, must cite the mapping and its version, and must carry the fidelity of the mapping path so consumers do not treat them as original assertions.

1. Which source binding and which mapping entry, at which mapping version, produced this binding? provenance
2. What is the fidelity of the mapping used, and what information was lost? quality
3. May derived bindings be chained through multiple mappings, and how does fidelity degrade along the chain? constraint
4. How is a derived binding distinguished from an original assertion in every projection? interoperability

Exchange Projection and Round-Trip Fidelity

The same binding may be projected as a FHIR Coding or CodeableConcept, an RDF triple with dcterms:subject to a skos:Concept, an SDMX code reference within a hierarchy association, a JSON object, a Markdown front-matter field or a MongoDB document. Most projections are lossy: a bare Coding cannot carry status, validity interval, authority or dispute state. Each projection must declare which fields survive and how a round trip is reconstructed.

1. Which exchange projections are supported, and what is the field mapping for each? interoperability
2. Which binding fields are lost in each projection, and how is that loss disclosed to the consumer? quality
3. Can a binding be reconstructed from a projection, and what canonical form is used for comparison? validation

Translations, mappings and uncoded text

A CodeableConcept may carry several Codings as translations of one concept plus text. SKOS mapping properties and GSIM correspondence tables relate items across schemes and are referenced, not duplicated. Uncoded text or dcat:keyword remains a valid degraded form when a URI is not feasible, with recommended later alignment.

1. Which additional codings are translations of the same concept, and which coding is user-selected or preferred for exchange? interoperability
2. What mapping or correspondence, if any, justifies treating two items from different schemes as comparable on this subject? relationship
3. If only a free-text keyword or literal subject is present, what controlled-vocabulary alignment is planned and who owns that gap? interoperability
4. When projecting to DCAT, is dcat:theme used for KOS concepts and dcat:keyword for literals, without assuming a fixed OWL range for theme? interoperability

Sensitive Binding Access, Retention and Erasure

A binding can be more sensitive than the subject it describes. Classifying a natural person by health, ethnicity, religion, trade union membership, sex life or criminal convictions engages GDPR Article 9; automated classification of persons is profiling under Article 4(4) and may engage Article 22. Access, retention and erasure must therefore be governed at binding granularity, and erasure must not silently destroy the audit trail that legal or statistical obligations require.

1. Does this binding constitute a special category of personal data or otherwise heightened-risk information? privacy
2. Who may read this binding, and is read access narrower than access to the subject record? access
3. How long is the binding retained after it is superseded or revoked, and under what obligation? retention
4. When a binding must be erased, what remains, and how is the audit trail preserved without re-identifying the subject? exception