Semantic Library

The Semantic Library extension records the controls that govern what an AI system may say, assert, recommend or emit, and ties each one to the obligation it enforces and the part of the business it applies to. It has two record types:

• Obligation: a requirement captured at clause level from its governing source: a statute, prudential standard, regulatory guidance, industry code, contract, clinical standard or internal policy.
• Semantic constraint: a typed, governed constraint on AI output, with the obligations it enforces, the enforcement tier it is routed to and why, its stewardship and lifecycle, and the calibration or conformance evidence for its tier.

Schemas: obligation.schema.json, semantic-constraint.schema.json
Specification: JSON Schema Draft-07

Why it belongs in the metamodel

Authoring a single constraint well is a technical problem. Keeping a library of them coherent is a governance problem, and most of what it needs is context the metamodel already describes:

The library needs	The metamodel provides
A faceted taxonomy to author against, so a constraint is written once and applied in many contexts	Typed domains and the shared, versioned vocabulary in common.schema.json
Traceability from source clause to control to evidence	Obligation records, linked from each constraint's obligationMappings
An inventory of AI use cases, with owners	Model Risk Management model profiles, referenced by governedModelIDs
Where AI output reaches customers	Channels, services, products and value-stream stages, referenced by appliesTo
Who owns, writes and approves each control	Organization units and roles, in stewardship
A record of why the library changed	Triggers whose affectedEntities name the changed obligations
Coverage of obligations by controls	Coverage expectations and status by domain

The extension does not hold the semantic machinery itself. Reference vectors, projections and thresholds are produced by the authoring method and stored as artifacts; the record keeps references to them, the model they were calibrated for, and the evidence that they passed validation.

Routing a constraint to a tier

Every constraint is routed to one enforcement tier, and the reason is recorded in tierRationale:

Tier (enforcementTier)	Use when	Example
rulesTier	Compliance can be decided from the form of the output alone	The response must contain the approved warning; the output must validate against a JSON Schema
semanticTier	Compliance depends on what the output means; two outputs with identical form can differ	Is this general or personal advice? Is this sustainability claim substantiated?
humanAdjudication	Compliance needs facts outside the output, or is genuinely contested	Is this recommendation in the member's best interests?

When a requirement has a part decidable by form and a part that depends on meaning, split it into two constraints and link them with splitFromConstraintID, so the residual is visible rather than lost. The worked example below does this.

Obligation

Attribute	Type	Description	Required
obligationID	String	Unique identifier for the obligation, e.g. OBL-001	✓
title	String	Short name of the obligation	✓
statement	String	The obligation in plain language: what must or must not be done	✓
obligationType	Enum	Kind of source the obligation comes from. recognisedStandard = a published standard or framework from a standards body (for example ISO/IEC 42001 or the NIST AI RMF): statutory, prudentialStandard, regulatoryGuidance, industryCode, recognisedStandard, contractual, clinicalStandard, internalPolicy, other	✓
source	Object	The governing source, cited precisely enough to find the clause	✓
source.authority	String	Body that issues or enforces the source, e.g. Parliament, APRA, ASIC, a standards body, the organization itself (required)
source.instrument	String	The instrument, e.g. Corporations Act 2001 (Cth), CPS 230, RG 271, an internal policy (required)
source.citation	String	Section, paragraph or clause, e.g. s766B(3) (required)
source.instrumentVersion	String	Version, compilation or amendment of the instrument the obligation was captured from
source.url	String	Link to the authoritative text
source.externalOrgID	String	External Organization record for the authority, when one exists
source.provision	String	Short heading of the provision cited, e.g. 'Material service providers'
jurisdiction	String	Jurisdiction in which the obligation applies, e.g. AU
sourceTier	Enum	Tier of the source the obligation comes from: binding (law or prudential standard in force), expected (regulator expectation, guidance or recognised standard) or anticipated (made but not yet in force, under consultation, or a pending bill). Describes the source only: binding, expected, anticipated
tierBasis	Array[Object]	Other obligations this one rests on, with their tiers. Lets a record carry a second tier, as when a regulator's expectation rests on a binding standard
tierBasis[].obligationID	String	The obligation relied on (required)
tierBasis[].sourceTier	Enum	Tier of that obligation's source: binding, expected, anticipated (required)
tierBasis[].note	String	How the one rests on the other
entityTypes	Array[Enum]	Kinds of entity the source is addressed to: authorisedDepositTakingInstitution, generalInsurer, lifeInsurer, privateHealthInsurer, superannuationTrustee, financialServicesLicensee, creditLicensee, listedEntity, bank, insurer, capitalMarketsIntermediary, paymentServiceProvider, financialInstitution, anyOrganization, other
roles	Array[Enum]	Roles within an entity or an AI value chain that the source addresses: regulatedEntity, board, seniorManagement, accountablePerson, provider, deployer, serviceProvider, importer, distributor
useCaseIDs	Array[String]	Standardised AI use cases (Use Case records) the obligation may reach
effectiveDate	Date	Date from which the obligation applies
effectiveDates	Array[Object]	Key dates for the source and the provisions it contains, such as when it was made, when it commences, when particular provisions apply, or when a consultation closes. effectiveDate keeps the main date
effectiveDates[].dateType	Enum	What the date marks: made, published, commences, applies, transitionEnds, consultationOpens, consultationCloses, reviewDue, repealed (required)
effectiveDates[].date	Date	The date (required)
effectiveDates[].provisions	String	Provisions the date applies to, when not the whole instrument, e.g. 'sections 5-6'
effectiveDates[].note	String	Context for the date
status	Enum	Whether the obligation currently applies: proposed, pending, inForce, amended, superseded, repealed, withdrawn	✓
supersededByObligationID	String	The obligation that replaces this one, when status is superseded or amended
appliesTo	Array[Object]	Context the obligation applies to: products, services, channels, customer segments, value streams, capabilities or organization units
appliesTo[].domainType	Enum	Domain of the referenced entity, from the shared domain list (domainType in common.schema.json) (required)
appliesTo[].entityID	String	Unique identifier of a record within its domain, for example "CAP-001" or "MKT-ENTSW-001" (required)
appliesTo[].title	String	Name of the referenced entity, for readability
implementingPolicyIDs	Array[String]	Internal policies (Policy domain) that implement the obligation
orgUnitTitle	String	Organization unit accountable for meeting the obligation
orgUnitID	String	ID of the organization unit named in orgUnitTitle
accountablePerson	String	Named role personally accountable for the obligation, e.g. an accountable person under an accountability regime
riskLevel	Enum	Consequence of failing the obligation: low, medium, high, critical
complianceStatus	Enum	Current assessed compliance with the obligation: compliant, partiallyCompliant, nonCompliant, underReview, exempt, notApplicable
relatedObligationIDs	Array[String]	Obligations that overlap with or depend on this one
monitoringFrequency	Enum	How often the source is watched for change: continuous, realTime, hourly, daily, weekly, monthly, quarterly, semiAnnually, annually, eventDriven, onDemand, adHoc
lastReviewDate	Date	Date the obligation was last checked against its source
lastVerified	Date	Date the record was last checked against the official text by its publisher. Unlike lastReviewDate, which is the organization's own review, this is the date a reference dataset confirmed the record
provenanceRecordIDs	Array[String]	Provenance records, issued through the governance interface, that trace how this record was sourced, extracted, reconciled and approved

An obligation that is superseded or amended must name its successor in supersededByObligationID. A binding obligation cannot be proposed, pending or withdrawn, and an anticipated one cannot be inForce, amended or repealed.

Obligations as a reference dataset

Obligations can also be published as a reference dataset, such as a register of AI governance obligations by jurisdiction. In that use each record carries a source tier, the entity types, roles and standardised use cases it may reach, its key dates, and when its publisher last verified it against the official text. The records can be governed through the governance interface, which adds provenance and a named approver to every change of status, date or tier.

Source Tier (sourceTier)

A tier describes the source, never the position of anyone the source might apply to. It does not say that an organization is compliant or in breach, that it must meet the source, or that an anticipated source will apply.

Value	Source	Examples
binding	Law or prudential standard in force	A prudential standard's paragraphs on material service providers; EU AI Act provisions already applying
expected	Regulator expectations, guidance and recognised standards	A regulator's letter setting minimum expectations; ISO/IEC 42001; NIST AI RMF
anticipated	Made but not yet in force, under consultation, or a pending bill	Guidelines whose later sections apply from a future date; EU AI Act provisions with later start dates

A record can carry a second tier through tierBasis, where an expectation rests on a binding standard: a regulator's expectation about board risk appetite, for example, resting on the risk management and operational risk standards.

Semantic Constraint

Attribute	Type	Description	Required
semanticConstraintID	String	Unique identifier for the constraint, e.g. SC-001	✓
title	String	Short name of the constraint	✓
statement	String	The constraint in plain language, as an approver would read it	✓
constraintType	Enum	Kind of constraint: prohibition, boundedRange, requiredProperty, structuralForm, closedVocabulary, gradedSeverity	✓
obligationMappings	Array[Object]	Obligations the constraint enforces	✓
obligationMappings[].obligationID	String	The obligation enforced (required)
obligationMappings[].mappingRole	Enum	Whether this constraint is the obligation's primary control or supports another: primary, supporting (required)
obligationMappings[].note	String	How the constraint covers the obligation, and any part it leaves to other controls
enforcementTier	Enum	Tier the constraint is enforced in: rulesTier, semanticTier, humanAdjudication	✓
tierRationale	String	Why the constraint is routed to this tier: whether compliance is decidable from form, depends on meaning, or needs facts or judgement outside the output	✓
splitFromConstraintID	String	When a requirement was split across tiers, the constraint holding the other part
appliesTo	Array[Object]	Business context the constraint governs: products, services, channels, customer segments, value streams or capabilities. Constraints are authored against one distinction with this context as metadata, not re-authored per context.
appliesTo[].domainType	Enum	Domain of the referenced entity, from the shared domain list (domainType in common.schema.json) (required)
appliesTo[].entityID	String	Unique identifier of a record within its domain, for example "CAP-001" or "MKT-ENTSW-001" (required)
appliesTo[].title	String	Name of the referenced entity, for readability
governedModelIDs	Array[String]	AI systems the constraint governs, as Model Risk Management model profiles (the AI use-case inventory)
severityScale	Array[Object]	Graded-severity ladder, lowest first; required for gradedSeverity constraints
severityScale[].level	Integer	Position on the ladder; higher is more severe (required)
severityScale[].label	String	Short name of the level (required)
severityScale[].description	String	What a statement at this level does, with the distinction from the level below (required)
severityScale[].disposition	Enum	What happens to a statement at this level: allow, allowWithControl, escalate, block
contestedDisposition	Enum	What the constraint does in a contested or uncertain region: enforce, abstainAndEscalate, advisoryOnly
exemplarSets	Array[Object]	Curated example sets the constraint is authored, calibrated and regression-tested from
exemplarSets[].exemplarSetID	String	Unique identifier for the exemplar set (required)
exemplarSets[].setType	Enum	Role of the set: canonical, adversarial, calibration, regression, mustAccept, mustReject (required)
exemplarSets[].itemCount	Integer	Number of labelled items in the set
exemplarSets[].version	String	Version of the set
exemplarSets[].location	String	Where the set is stored, e.g. a repository path or URI (required)
exemplarSets[].provenance	String	Where the items came from: source clauses, adjudicated incidents, promoted novel cases, red-team runs
semanticEnvelope	Object	Calibration and validation record for a semantic-tier constraint. Records references to artifacts, not the vectors themselves.
semanticEnvelope.embeddingModel	Object	The exact model a calibration is valid for (required)
semanticEnvelope.embeddingModel.modelName	String	Name of the embedding model (required)
semanticEnvelope.embeddingModel.modelVersion	String	Version identifier of the embedding model (required)
semanticEnvelope.embeddingModel.modelHash	String	Weight or version hash, where the model can be attested
semanticEnvelope.methodologyVersion	String	Authoring methodology and version the envelope was validated under, e.g. PACE 2.0 (required)
semanticEnvelope.divergenceMetric	Enum	Divergence measure between the reference and the governed output: cosine, euclidean, mahalanobis, klDivergence, jensenShannon (required)
semanticEnvelope.threshold	Number	Calibrated divergence threshold above which the constraint fires (required)
semanticEnvelope.alpha	Number	Conformal significance level the threshold was calibrated to (target error rate) (required)
semanticEnvelope.calibrationDate	Date	Date the threshold was last calibrated (required)
semanticEnvelope.referenceArtifact	String	Location and hash of the reference vector and projection artifacts
semanticEnvelope.validation	Object	Result of the constraint validation harness, the ship gate for publication
semanticEnvelope.validation.validatedDate	Date	Date the harness was run (required)
semanticEnvelope.validation.passed	Boolean	Whether the constraint cleared every registered floor (required)
semanticEnvelope.validation.separationAUC	Number	Separation of compliant from non-compliant exemplars
semanticEnvelope.validation.recall	Number	Recall at the operating point
semanticEnvelope.validation.falsePositiveRate	Number	False positive rate at the operating point
semanticEnvelope.validation.monotonicityPassed	Boolean	Whether divergence increases with authored severity
semanticEnvelope.validation.beatsBaseline	Boolean	Whether the constraint outperforms the registered baseline
semanticEnvelope.validation.shiftRobustnessPassed	Boolean	Whether performance held under the distribution-shift test
semanticEnvelope.validation.leakagePassed	Boolean	Whether the leakage check found no concept leakage
semanticEnvelope.validation.evidenceLocation	String	Where the harness evidence pack is stored
grammarArtifact	Object	Specification and conformance record for a rules-tier constraint
grammarArtifact.formalism	Enum	Formalism the rule is expressed in: jsonSchema, regularExpression, contextFreeGrammar, closedVocabulary, other (required)
grammarArtifact.artifactLocation	String	Where the grammar, schema or vocabulary is stored (required)
grammarArtifact.artifactVersion	String	Version of the grammar artifact
grammarArtifact.compiledArtifactHash	String	Hash of the compiled automaton or enforcement artifact
grammarArtifact.methodologyVersion	String	Authoring methodology and version, e.g. GRACE 0.1 (required)
grammarArtifact.conformance	Object	Result of the conformance harness, the ship gate for publication
grammarArtifact.conformance.validatedDate	Date	Date the harness was run (required)
grammarArtifact.conformance.passed	Boolean	Whether the grammar cleared every registered floor (required)
grammarArtifact.conformance.mustAcceptPassRate	Number	Share of must-accept items accepted
grammarArtifact.conformance.mustRejectAcceptances	Integer	Must-reject items wrongly accepted (the gate requires zero)
grammarArtifact.conformance.mutationKillRate	Number	Share of grammar mutants the corpus detects
grammarArtifact.conformance.evidenceLocation	String	Where the harness evidence pack is stored
stewardship	Object	Who owns, writes, calibrates and approves the constraint. The author and approver must be different people (separation of duties).	✓
stewardship.policyOwner	String	Accountable for the intent the constraint enforces (required)
stewardship.policyOwnerOrgUnitID	String	Organization unit of the policy owner
stewardship.author	String	Semantic librarian or constraint engineer who drafted the constraint (required)
stewardship.calibrator	String	Person who set and validated the threshold (semantic tier)
stewardship.approver	String	Editorial board member or delegate who approved publication; not the author
stewardship.approvalDate	Date	Date the current version was approved
monitoring	Object	How the constraint is watched once published, and what prompts a review
monitoring.monitoringFrequency	Enum	How often coverage and drift are reviewed: continuous, realTime, hourly, daily, weekly, monthly, quarterly, semiAnnually, annually, eventDriven, onDemand, adHoc
monitoring.coverageFloor	Number	Registered minimum coverage; falling below it triggers recalibration
monitoring.reviewTriggers	Array[Enum]	Events that put the constraint back through review: coverageDegradation, guaranteeInvalidation, suspectedLeakage, obligationChange, modelChange, novelCasePromotion, incidentAdjudication, scheduledReview
monitoring.lastReviewDate	Date	Date of the last review
status	Enum	Lifecycle state. Constraints are never deleted or edited in place: a constraint that no longer reflects its obligation is deprecated with a pointer to its successor: draft, validated, published, deprecated, inactivated	✓
version	String	Semantic version of the constraint	✓
supersededByConstraintID	String	Successor constraint; required when deprecated. An inactivated constraint (e.g. its obligation was repealed) may have none.
relatedConstraintIDs	Array[String]	Constraints that overlap with or complement this one; overlaps are reviewed so each obligation has one primary constraint

Rules the schema enforces

• Every constraint maps to at least one obligation.
• A gradedSeverity constraint has a severityScale of at least two levels.
• A semanticTier constraint has exemplarSets; once validated or published it has a semanticEnvelope with its harness validation result.
• A rulesTier constraint has a grammarArtifact; once validated or published it has its conformance result.
• A published constraint names its approver and approvalDate.
• A deprecated constraint names its successor in supersededByConstraintID. Constraints are never deleted or edited in place.

Separation of duties (the approver is not the author) cannot be expressed in JSON Schema; check it in the approval workflow.

Worked example: the advice boundary

A superannuation fund's member-services assistant may give factual information and general advice, but its chat channel is not authorised to give personal advice. All names, numbers and dates are illustrative; no calibration has been run, and the obligation statements summarise the cited provisions for the example, not as legal advice. The full example, with a model profile for the assistant, is in the repository: examples/semantic-library-advice-boundary.md.

Obligations

{
  "obligationID": "OBL-001",
  "title": "No personal advice through unauthorised member channels",
  "statement": "Member channels operating without a personal-advice authorisation must not give personal advice: a recommendation or opinion that takes into account the member's objectives, financial situation or needs.",
  "obligationType": "statutory",
  "source": {
    "authority": "Commonwealth Parliament; administered by ASIC",
    "instrument": "Corporations Act 2001 (Cth)",
    "citation": "s766B(3) (personal advice); s911A (need for a licence)",
    "url": "https://www.legislation.gov.au/C2004A00818/latest"
  },
  "jurisdiction": "AU",
  "status": "inForce",
  "appliesTo": [
    { "domainType": "channel", "entityID": "CH-MEMBER-CHAT", "title": "Member services chat" },
    { "domainType": "product", "entityID": "PROD-ACCUM", "title": "Accumulation account" }
  ],
  "implementingPolicyIDs": ["POL-ADVICE-001"],
  "orgUnitTitle": "Member Advice",
  "orgUnitID": "OU-ADVICE",
  "accountablePerson": "Chief Member Officer (accountable person)",
  "riskLevel": "critical",
  "complianceStatus": "underReview",
  "relatedObligationIDs": ["OBL-002"],
  "monitoringFrequency": "monthly",
  "lastReviewDate": "2026-10-01"
}

{
  "obligationID": "OBL-002",
  "title": "General advice warning",
  "statement": "When general advice is given to a retail client, warn the client that the advice does not take into account their objectives, financial situation or needs.",
  "obligationType": "statutory",
  "source": {
    "authority": "Commonwealth Parliament; administered by ASIC",
    "instrument": "Corporations Act 2001 (Cth)",
    "citation": "s949A"
  },
  "jurisdiction": "AU",
  "status": "inForce",
  "appliesTo": [
    { "domainType": "channel", "entityID": "CH-MEMBER-CHAT", "title": "Member services chat" }
  ],
  "implementingPolicyIDs": ["POL-ADVICE-001"],
  "orgUnitTitle": "Member Advice",
  "orgUnitID": "OU-ADVICE",
  "riskLevel": "high",
  "relatedObligationIDs": ["OBL-001"]
}

The advice-boundary constraint (semantic tier)

A graded-severity constraint: factual information is allowed, general advice is allowed with the warning, borderline responses escalate to a person, and personal advice is blocked.

{
  "semanticConstraintID": "SC-001",
  "title": "Advice boundary for member chat",
  "statement": "The member services assistant must not make a recommendation or state an opinion that takes into account the member's objectives, financial situation or needs.",
  "constraintType": "gradedSeverity",
  "obligationMappings": [
    { "obligationID": "OBL-001", "mappingRole": "primary" }
  ],
  "enforcementTier": "semanticTier",
  "tierRationale": "Whether a response is personal advice depends on its meaning: two responses with identical form can differ in whether they consider the member's circumstances. A keyword list would be evaded by paraphrase.",
  "appliesTo": [
    { "domainType": "channel", "entityID": "CH-MEMBER-CHAT", "title": "Member services chat" },
    { "domainType": "valueStream", "entityID": "VS-MEMBER-ENQ", "title": "Resolve member enquiry" }
  ],
  "governedModelIDs": ["MODEL-CHAT-001"],
  "severityScale": [
    { "level": 0, "label": "Factual information", "description": "States facts about the fund, products or rules without recommending anything", "disposition": "allow" },
    { "level": 1, "label": "General advice", "description": "Recommends or gives an opinion without reference to the member's own circumstances", "disposition": "allowWithControl" },
    { "level": 2, "label": "Borderline", "description": "General advice framed around details the member has shared, where it is unclear whether their circumstances were taken into account", "disposition": "escalate" },
    { "level": 3, "label": "Personal advice", "description": "Recommends a course of action for this member based on their objectives, financial situation or needs", "disposition": "block" }
  ],
  "contestedDisposition": "abstainAndEscalate",
  "exemplarSets": [
    { "exemplarSetID": "EXS-001-CAN", "setType": "canonical", "itemCount": 240, "version": "1.0", "location": "library/advice-boundary/canonical", "provenance": "Written by the advice team against s766B and the fund's advice policy; each item labelled with a severity level" },
    { "exemplarSetID": "EXS-001-ADV", "setType": "adversarial", "itemCount": 600, "version": "1.0", "location": "library/advice-boundary/adversarial", "provenance": "Paraphrase red-team runs" },
    { "exemplarSetID": "EXS-001-CAL", "setType": "calibration", "itemCount": 1500, "version": "1.0", "location": "library/advice-boundary/calibration", "provenance": "Labelled sample of historical chat transcripts" }
  ],
  "semanticEnvelope": {
    "embeddingModel": { "modelName": "example-embedding-model", "modelVersion": "2026-06" },
    "methodologyVersion": "PACE 2.0",
    "divergenceMetric": "cosine",
    "threshold": 0.31,
    "alpha": 0.05,
    "calibrationDate": "2026-10-15",
    "validation": {
      "validatedDate": "2026-10-16",
      "passed": true,
      "separationAUC": 0.94,
      "recall": 0.95,
      "falsePositiveRate": 0.04,
      "monotonicityPassed": true,
      "beatsBaseline": true,
      "shiftRobustnessPassed": true,
      "leakagePassed": true,
      "evidenceLocation": "library/advice-boundary/evidence/2026-10-16"
    }
  },
  "stewardship": {
    "policyOwner": "Head of Advice",
    "policyOwnerOrgUnitID": "OU-ADVICE",
    "author": "Semantic librarian, Member Advice",
    "calibrator": "Model validation analyst",
    "approver": "Editorial board chair"
  },
  "monitoring": {
    "monitoringFrequency": "weekly",
    "coverageFloor": 0.95,
    "reviewTriggers": ["coverageDegradation", "guaranteeInvalidation", "suspectedLeakage", "obligationChange", "modelChange", "novelCasePromotion", "incidentAdjudication"]
  },
  "status": "validated",
  "version": "1.0.0",
  "relatedConstraintIDs": ["SC-002"]
}

The general advice warning (rules tier)

The other half of the split: once a response is classified as general advice, whether it carries the approved warning is decidable from the text.

{
  "semanticConstraintID": "SC-002",
  "title": "General advice warning present",
  "statement": "A response classified as general advice must include the fund's approved general advice warning.",
  "constraintType": "requiredProperty",
  "obligationMappings": [
    { "obligationID": "OBL-002", "mappingRole": "primary" }
  ],
  "enforcementTier": "rulesTier",
  "tierRationale": "Once the response is classified as general advice, the presence of the approved warning is decidable from the text alone.",
  "splitFromConstraintID": "SC-001",
  "appliesTo": [
    { "domainType": "channel", "entityID": "CH-MEMBER-CHAT", "title": "Member services chat" }
  ],
  "governedModelIDs": ["MODEL-CHAT-001"],
  "exemplarSets": [
    { "exemplarSetID": "EXS-002-ACC", "setType": "mustAccept", "itemCount": 120, "location": "library/general-advice-warning/must-accept", "provenance": "Approved warning wordings in context" },
    { "exemplarSetID": "EXS-002-REJ", "setType": "mustReject", "itemCount": 80, "location": "library/general-advice-warning/must-reject", "provenance": "Truncated, reworded and missing warnings" }
  ],
  "grammarArtifact": {
    "formalism": "regularExpression",
    "artifactLocation": "library/general-advice-warning/warning.regex",
    "artifactVersion": "0.1",
    "methodologyVersion": "GRACE 0.1"
  },
  "stewardship": {
    "policyOwner": "Head of Advice",
    "policyOwnerOrgUnitID": "OU-ADVICE",
    "author": "Semantic librarian, Member Advice"
  },
  "status": "draft",
  "version": "0.1.0",
  "relatedConstraintIDs": ["SC-001"]
}

When a source changes

A trigger records the change; its affectedEntities name the obligations, and the constraints that map to them are reviewed.

{
  "triggerID": "TRG-101",
  "label": "Regulator updates guidance on general and personal advice",
  "description": "Updated regulatory guidance changes the examples of where general advice becomes personal advice in digital channels",
  "primaryCategory": "regulatoryCompliance",
  "subCategory": "Advice guidance",
  "origin": "External",
  "timeHorizon": "shortTerm",
  "impactLevel": "high",
  "detectionMethod": "Regulatory watch",
  "detectionDate": "2026-10-20",
  "affectedEntities": [
    { "domainType": "obligation", "entityID": "OBL-001", "title": "No personal advice through unauthorised member channels" },
    { "domainType": "obligation", "entityID": "OBL-002", "title": "General advice warning" }
  ]
}

Related

• Model Risk Management: the AI use-case inventory the constraints govern
• Strategic Response Model: Trigger: recording why the library changes
• Cross-Domain Relationships: obligation and semanticConstraint are domains that relationships can target