Model the Property Pack from first principles with bounded-context ownership

Update 2026-09-03 — the implemented reader pages now live at /development/property-pack/**; this route-only change does not alter the modelling decision.

Context and Problem Statement

ADR-0066 fixes the initial coverage boundary for the new ontology at the 451 data points marked Required in the Property Pack workbook. It deliberately leaves their semantic treatment unresolved. A source path or JSON datatype does not decide whether a data point represents a resource, relationship, attribute, classification, rule or controlled value.

The previous ontology was generated from JSON Schemas and form overlays. That work is useful evidence, but it inherited document-tree structure, uneven definitions and form-support decisions that were never established as domain meaning. Repairing that model in place, or allowing its classes and predicates to seed the replacement, would carry those assumptions into the new work.

The sibling semantic-modelling project provides a broader fourteen-category ontology model in ODR-0071. It was designed for a source-code extraction pipeline and therefore contains concerns that do not belong in this OPDA model, including process modelling, service and enterprise architecture, capability modelling, source-to-ontology mappings and data-product packaging. It also contains concerns OPDA does need, including domain structure, vocabularies, validation, provenance, time, sensitivity and authorisation.

The Property Pack is not itself a bounded context. It is a delivery profile that draws on meanings owned in several domain contexts. OPDA therefore also needs a firm rule for where each modelled resource belongs, how genuinely shared concepts enter the common boundary, and how different context-specific meanings remain connected without being collapsed.

This ADR decides the target architecture for the first greenfield Property Pack ontology. Its implementation may generate isolated review candidates; it neither approves their meaning nor publishes a replacement for the current website.

Decision Drivers

• Cover every one of the 451 required Property Pack data points without turning the workbook or JSON tree into the ontology structure.
• Model identity and meaning from first principles, evidence and established domain practice.
• Give every OPDA-defined resource one accountable semantic home.
• Preserve legitimate differences between bounded contexts while making exchange explicit.
• Keep provenance, validation, time, sensitivity and authorisation as first-class modelling concerns.
• Reuse established linked-data standards instead of inventing OPDA equivalents.
• Keep legacy artefacts available as attributed evidence without granting them semantic authority or allowing them to expand the closed Property Pack scope.
• Use an auditable, escalation-driven agentic authoring method without making the ontology corpus depend on Semantic Builder, one agent substrate or one LLM family.
• Preserve material disagreement for human resolution instead of manufacturing consensus through agent votes.

Considered Options

• Option A — Repair and extend the existing schema-derived ontology. Improve its terms and definitions while retaining its resource and relationship structure.
• Option B — Adopt the complete ODR-0071 pipeline model and Semantic Builder. Use all fourteen categories and the source-extraction architecture that motivated them.
• Option C — Model the 451-item scope from first principles, adopt only the relevant ODR-0071 concerns, and assign every resource to a bounded context or the common boundary (chosen).

Decision Outcome

Chosen option: Option C — a first-principles, context-owned ontology constrained by the 451-item Property Pack coverage boundary.

The ontology will be designed afresh. The 451 required Property Pack items determine what the initial model must be able to represent; they do not determine its classes, predicates, nesting, identities or number of ontology terms. One item may require several semantic constructs, and several items may be represented by one well-founded construct. No source item may disappear without an explicit coverage disposition.

The ontology model and durable corpus are independent of Semantic Builder. The authoring programme nevertheless adopts a pinned OPDA execution profile over Semantic Builder’s accepted contracts for workcells, model routing, source isolation, candidate branches, deterministic conformance, councils, authority and releases. These contracts govern how candidate models are produced and challenged; they do not become ontology dependencies or grant an agent authority over OPDA decisions.

Semantic Modelling ODR-0104 is deprecated and relocated to Semantic Builder ADR-0021; it is historical provenance rather than a governing record. The earlier OPDA ODR-0001 protocol also does not govern this greenfield programme: its voting Queen, vote tallies and selectable consensus machinery are replaced here by the Builder configuration’s non-voting lead, independent priors, bounded cross-examination and escalation of unresolved disagreement. Historical sessions remain evidence, not authority.

AI agents and expert councils assist with evidence analysis and candidate generation; they have no vote or ontology authority. Deterministic tools certify conformance, while real domain participants judge meaning and recorded OPDA decisions remain authoritative.

OPDA authority follows its Constitution and Articles: the General Assembly is the overall decision-making body; the Executive Committee oversees the Trust Framework and may delegate work; and the directors retain management and delegation powers. Within that chain, the affected domain working group reviews local meaning; the Interoperability Working Group reviews common-boundary terms and mappings without overriding local meaning; Technical Review applies change and conformance controls, with Compliance and Risk input for major or sensitive changes; and only the Executive Committee, Board or a recorded delegate may authorise promotion.

The governance SOP inherited with the PDTF schema material is a process input, not a complete release charter or evidence of an endorsed predecessor scheme; its placeholders, working-group decision rules, conflicts, appeals and final release delegation still require recorded approval. Until ratified, outputs remain non-normative candidates: AI agreement, informal consensus or validation cannot authorise release alone.

1. Evidence hierarchy and scope guard

The evidence hierarchy for this work is:

1. Property Pack coverage authority — the 451 required source-item IDs define the closed initial business-data scope.
2. Domain evidence — working-group material, recognised standards, legislation, common industry practice and expert review establish candidate meaning.
3. Legacy OPDA evidence — the existing ontology, business glossary, data dictionary, JSON Schemas, mappings and form overlays may explain history, expose implementation expectations and support migration analysis.
4. Generated proposals — AI-produced definitions, classifications and model structures remain candidates until reviewed under OPDA governance.

Legacy material must never silently introduce an optional field, overlay concern, resource, relationship, cardinality or vocabulary into the new business-data scope. It may be cited when it supports a decision about one or more of the 451 items.

Modelling from first principles may introduce the classes, relationships, policy concepts and other scaffolding required to represent the 451 items correctly. Every new business construct must trace to one or more Property Pack source-item IDs. A supporting technical construct that is not itself a Property Pack item must trace to a retained modelling concern, a named standard and a documented need; it must not enlarge the Property Pack’s business-data scope.

2. Resource-home rule

Every OPDA-defined ontology resource must have exactly one semantic home:

• one of the six property bounded contexts: Finance and Banking, Conveyancing, Estate Agency, Surveying and Valuation, Property Data Services, or Property Technology;
• the DBT Smart Data scheme context, where the resource expresses scheme-level trust, consent, authorisation or participation semantics rather than property-domain meaning; or
• the deliberately small common boundary, governed by the Interoperability Working Group.

This home rule applies to OPDA-minted classes, properties, relationships, shapes, vocabulary schemes and concepts. An externally governed resource keeps its external home; OPDA records which context reuses it and any local constraint or mapping. Semantic home records definition responsibility. ODR-0065 now selects separate flat context namespaces on opda.org.uk for the new governed model, refining this ADR’s earlier namespace discretion. One resource may still be used by several contexts. Policy selection does not migrate existing candidate or schema-derived IRIs.

Each Property Pack source item receives one candidate semantic home—one context or the common boundary—and may name additional consuming contexts. Usage by several contexts does not automatically move a resource into the common boundary. A resource enters the common boundary only when multiple contexts require the same identity criterion and stable exchange meaning, and the Interoperability Working Group approves that shared definition.

Where the same label has materially different meanings in different contexts, the contexts keep distinct resources and the Interoperability Working Group records the relationship between them. Cross-context mappings connect meanings; they do not force one context to adopt another context’s internal model.

The common boundary is therefore an explicit semantic home, not a miscellaneous shared folder or a default for unresolved ownership.

3. ODR-0071 adoption profile

ODR-0071 is adopted as a taxonomy of modelling concerns, not as a wholesale pipeline architecture. OPDA retains eight of its fourteen categories:

The eight retained categories are a completeness check, not a demand for eight separate ontology files. Working-group presentation lenses are not formal ontology categories. Governance remains an operating concern for authoring, review and release, but is not an ontology output or a source of property-domain terms.

The access-control category models the meaning of sensitivity, consent, roles, purposes and authorisation. It does not by itself implement a runtime security system or grant access to data.

Coverage links from Property Pack source-item IDs to model constructs are requirements traceability and provenance. They are not executable RML/R2RML source mappings and do not reinstate Category 13.

4. Required semantic package

The first candidate must keep the following outputs aligned:

• the context-owned RDF 1.2 and OWL 2 domain ontologies and small common-boundary ontology;
• a plain-language business glossary;
• a data dictionary tied to the 451 source items;
• SKOS controlled vocabularies and taxonomies;
• SHACL 1.2 data shapes and validation rules;
• the DDD context map and cross-domain mappings;
• provenance, quality, temporal, sensitivity and authorisation metadata;
• a coverage matrix from all 451 source-item IDs to their semantic home and resulting semantic constructs.

Generated JSON Schemas, forms, documentation and other familiar representations may be produced from the governed model later. They are projections of the semantic agreement and must not become independent sources of meaning.

5. Semantic Builder execution profile

The reference baseline is Semantic Builder commit b64e4288bc07277198abad83bd7978db5c938b6b. OPDA inherits behaviour, not a binary dependency:

This ADR does not import the complete fourteen-category target, code-analysis and RML pipeline, Data Product packaging, H&M-specific governance bodies or role names, gold/training/ Darwin machinery, unattended autonomy, exact Jena version, signed-ledger implementation or deployment topology. Excluding those mechanisms does not exclude OPDA governance. Semantic Builder’s Fable/Sol build-collaboration policy is not an ontology-runtime model policy.

5.1 Bounded work orders and untrusted evidence

The 451 items are grouped into evidence-backed concept packages, not treated as 451 predetermined ontology properties. Each immutable work order records its item IDs, candidate semantic home, allowed read/write graphs, evidence snapshot and citations, retained concerns, target standards profile, competency questions and hard cases, chosen mechanism, model route and budget, expected outputs, validation plan and human decision owner.

All uploaded files, emails, transcripts, standards, spreadsheets, schemas and legacy artefacts are untrusted data, even when supplied by OPDA. They enter prompts only as delimited evidence with source identity, media type, digest, source span and limits; they cannot select instructions, tools, models, policy or destinations. Baseline checks cover decoding, secrets, personal/prohibited data, instruction-like content, archives, licensing and resource limits. Findings are quarantined for human disposition rather than silently obeyed or discarded. Source provenance and taint follow every derivative.

Agents write only a candidate delta in an isolated workspace. They receive no authority to edit the accepted corpus, common boundary, policy, evaluator or release. Initially, reviewed Git history and recorded OPDA approval remain the durable authority; stronger ledger, signing and fencing mechanics need their own implementation ADR.

5.2 Agents, reviews and councils

Every work order uses the least costly sufficient mechanism:

1. deterministic tools for extraction, normalisation, coverage, graph queries and validation;
2. one typed specialist for routine bounded authoring;
3. a linked pair or small panel for a known dependency;
4. a council only for underdetermined, high-lock-in, cross-profile or failed-specialist cases; and
5. the responsible working group or Interoperability Working Group for durable meaning.

The initial typed roles are evidence reader, ontology router, ontology author, conformance repairer and council panellist. OPDA excludes Builder’s Category-13 source mapper and any generic model judge: deterministic gates assess conformance, while domain experts and OPDA governance judge meaning and promotion.

There is no generic AI “quality reviewer” and no second council over every result. Verification consists of deterministic conformance, targeted evidence-based challenge, real domain review and the distinct cross-context harmonisation concern. Review agents may flag contradictions, unsupported claims, likely duplication and impact; they cannot certify truth or overrule human governance.

A creation council seats only the four to six ADR-0021 lenses needed for a named crux, such as identity, class-versus-value, reuse-versus-mint, constraints, lifecycle, provenance, authorisation or context ownership. Specialist guests cover SHACL, SKOS, PROV-O, DPV/ODRL, classification or time when required. Named experts are simulated, citation-grounded methodological lenses—not claims of participation or endorsement—and at least one real domain expert reviews every material package.

The lead is a non-voting protocol officer. The Devil’s Advocate is selected because a published methodology genuinely opposes a load-bearing premise, never because it uses another provider. Councils preserve: R1 clean-context independent priors; R2 bounded cross-examination with reasons for every change and no visible tally; and R3 re-poll, held dissent and escalation. Infrastructure consensus and model agreement never decide ontology meaning.

The Interoperability Council is a separate ownership mode, not a creation-model review. It compares two or more context candidates using identity, SKOS/SSSOM and DDD lenses, plus the affected human domain stewards. Its default is map before merge; only the Interoperability Working Group may approve durable mappings or common-boundary promotion.

5.3 Role-aware multi-model routing

Semantic Builder defines the portfolio roles and a provisional development roster; its manifest labels that roster development-hypotheses and its catalogue marks the named routes unavailable pending refresh and its prices as placeholders. OPDA therefore adopts that design, but pins an active OPDA roster before the first governed run. claude-fable-5 through Claude Code and gpt-5.6-sol through Codex/ChatGPT are the first routes to qualify. Routine work uses one role-qualified route; high-risk, uncertain or calibration work uses both model families under one work order, followed by blinded comparison and whole-candidate revalidation.

Expert-lens diversity and model-family diversity remain separate: opposed published lenses create methodological variety; providers provide a robustness check. A model route receives no vote. Role qualification records the exact provider, model and returned identity; underlying independence group; supported context and structured output; privacy/residency; latency/cost; evidence; reviewer and expiry. Mutable aliases, automatic routing and silent model/provider fallback are inadmissible evidence routes.

An outage may produce a clearly marked single-route draft, but cannot satisfy a dual-evidence gate. Replacement or changed routes require requalification. Calibration will assign champion, least-cost routine, independent challenger and declared fallback per role; it will not hard-code provider specialisms before measurement.

5.4 Deterministic validation, promotion and releases

Agents may propose RDF 1.2, SKOS, SHACL 1.2 and repairs; they never run, suppress or certify the validation gate. A trusted deterministic attempt pins the candidate, base graph, standards/tool profile, applicable OPDA concerns and evidence, then parses RDF 1.2, checks write scope and provenance, builds the exact graph view, runs SHACL 1.2, OWL 2, SPARQL 1.2 and competency queries, and seals the results. Each required validation cell is one of pass, fail, not-applicable, not-run or infrastructure-error; missing output, unsupported semantics or an empty report cannot become pass.

The standards family is RDF 1.2, SPARQL 1.2 and SHACL 1.2. Each candidate declares the exact conformance level it actually uses and tests: the initial Property Pack candidate uses valid RDF 1.2 Basic syntax, targets SHACL 1.2 Core, and tests a portable SPARQL 1.2 subset. Full RDF 1.2 and the SHACL 1.2 Union Profile are not claimed until a qualified engine and positive and negative feature probes demonstrate them. Jena/Fuseki is an implementation, not the standards boundary; unsupported or inert required behaviour blocks the run. Six excluded concerns are predeclared not-applicable.

Only a human disposition authorised through the OPDA authority chain above can move a validated candidate into the governed corpus. Published releases are immutable named cuts. Later changes supersede rather than rewrite history, classify compatibility for generated schemas and other consumers, retain deprecated terms for a declared window, and emit a machine-readable change record. Publication remains separately authorised.

5.5 Calibration and implementation boundary

The calibration harness pins 24 difficult Property Pack items across eight cases and compares the GPT and Claude routes under the same evidence contract. It measures conformance, unsupported claims, identity/resource-home errors, human correction, defensibility, reliability, latency and cost. Councils remain escalation-driven unless they demonstrate material value for the work class.

The first dual-provider execution produced all 16 case candidates and divergent fingerprints for every pair, but remains incomplete: Claude returned an attested model identity, the Codex CLI exposed only the requested route, and semantic RDF/SHACL validation was not run inside that calibration. The result is evidence for human disposition, not a route qualification or winner.

Semantic Builder has substantial typed contracts and tested foundations but no production-composed ontology-build command, qualified active runtime portfolio or council executor. OPDA may initially orchestrate this profile manually or through another substrate, provided work-order isolation, route/evidence receipts, clean priors, deterministic validation and human authority are preserved. No OPDA build or validation step may require Semantic Builder to be running.

6. Modelling and review gates

Before the first ontology candidate is eligible for working-group review:

1. all 451 source items have an explicit candidate semantic home;
2. every proposed OPDA resource has exactly one home and evidence for that choice;
3. every proposed business resource or rule traces to at least one source-item ID;
4. every supporting construct traces to a retained concern and named standard;
5. every source item maps to one or more model constructs, or has an explicit unresolved or challenged disposition—never silent omission;
6. candidate definitions clearly distinguish evidence, machine drafting and human approval;
7. common-boundary candidates have a multi-context justification and are reviewed by the Interoperability Working Group;
8. context mappings preserve local meaning and avoid unjustified owl:sameAs or premature merging;
9. the eight retained concerns each have a recorded disposition: model here, reuse shared, boundary contribution, or not applicable with rationale;
10. the six excluded concerns have not re-entered the ontology by accident;
11. every evidence snapshot has passed the source-policy boundary and retains source, digest, span and taint through its candidate derivatives;
12. every model-produced candidate records its exact route and work-order binding, and no agent has written directly to an accepted or common graph;
13. the deterministic validation matrix has no required fail, not-run or infrastructure-error, with conditional and repeatable structures preserved; and
14. no legacy path, nesting decision, predicate or definition has been adopted without an attributable semantic decision.

The field-to-term identity, split/consolidation rules and final context assignments must still be ratified in the fresh ontology decision record required by ADR-0066 before any semantic release. This ADR supplies that work with its architectural constraints; it does not pre-approve the resulting ontology terms.

Consequences

• Good, because the new model is accountable to a finite, stakeholder-recognisable scope without copying the JSON document tree.
• Good, because every resource has a named semantic owner and cross-context differences can remain explicit.
• Good, because the common boundary stays small and governed rather than becoming a universal property model.
• Good, because authorisation, privacy, provenance, quality, time and validation remain first-class even though process, enterprise architecture and source mapping are out of scope.
• Good, because the existing ontology, glossary and dictionary remain useful evidence without controlling the new design.
• Good, because OPDA reuses Semantic Builder’s work-order, routing, isolation, conformance and release contracts without making its corpus depend on that runtime.
• Good, because uploaded evidence cannot become agent control and AI-generated changes remain isolated proposals until deterministic checks and human approval.
• Good, because routine modelling is not burdened with council ceremony, while high-lock-in decisions receive opposed, citation-grounded review.
• Good, because the Interoperability Council maps context meanings without becoming a central authority over each bounded context.
• Good, because agent convergence cannot silently replace working-group and OPDA governance.
• Good, because immutable releases and explicit compatibility changes give generated schemas and other consumers a stable surface.
• Neutral, because correct modelling may produce more or fewer than 451 ontology terms.
• Neutral, because some Property Pack items will be represented through common resources or relationships rather than one direct datatype property.
• Bad, because every resource-home and split/consolidation decision now requires explicit evidence and review.
• Bad, because the new candidate will initially diverge structurally from the current published ontology and legacy JSON Schemas.
• Bad, because a small common boundary and separate context meanings require ongoing mapping and version governance.
• Bad, because screening, immutable work orders, route qualification, candidate isolation, validation receipts and release records add operational work.
• Bad, because the inherited Builder profile must initially be orchestrated without a production council executor.

Confirmation

• ADR-0066’s deterministic catalogue continues to report exactly 451 required source items.
• A generated coverage report accounts for every source-item ID and flags duplicate, missing and unresolved dispositions.
• A resource register demonstrates exactly one home for every OPDA-defined resource.
• The Interoperability Working Group reviews every common-boundary term and every cross-context mapping before release-candidate status.
• Each bounded context can review its glossary, dictionary, model diagrams, vocabularies and constraints without needing to edit RDF or JSON Schema.
• The ontology and SHACL corpus pass the project’s deterministic validation gates.
• A provenance audit can distinguish Property Pack source facts, external standards, legacy OPDA evidence, AI-generated proposals and human approvals.
• Adversarial source fixtures cannot select instructions, tools, models, policy or destinations, and quarantined evidence cannot reach a candidate unnoticed.
• Model receipts identify the actual provider, family and route; high-risk comparison uses independent families, and no silent fallback counts as evidence.
• Every council activation is traceable to a documented escalation reason or the calibration experiment, and its lead has no vote or durable write authority.
• Council records preserve clean-context priors, reasons for changed positions, unresolved disagreement and the responsible human disposition.
• A calibrated comparison demonstrates whether a strong agent, typed specialists or a council is the proportionate mechanism for each modelling work class.
• Only recorded human authority can promote a validated candidate; releases are immutable, and later changes carry compatibility, supersession and deprecation data.
• No build or validation step requires Semantic Builder.
• Publication or replacement of the current website remains separately authorised.

More Information

• ADR-0066 — the 451 required Property Pack data points as the greenfield seed scope
• ADR-0039 — linked data as the SPDTF foundation
• ADR-0063 — proposed domain-led bounded-context working groups
• ADR-0065 — proposed AI-assisted evidence-to-model workflow
• OPDA Constitution and Articles of Association
• PDTF technical governance process and website status summary
• Property Pack 451 evidence validation
• ODR-0071 — Ontology Modelling Category Framework
• Pinned Semantic Builder baseline
• Builder ADR-0003 — typed agent topology and ADR-0006 — model routing
• Builder ADR-0007 — harmonisation, ADR-0008 — escalation, ADR-0009 — termination and ADR-0021 — expert protocol
• Builder ADR-0011 — untrusted sources and ADR-0012 — stable releases
• Builder ADR-0026 — candidate branches and authority and ADR-0039 — replaceable Ruflo capabilities
• Builder ADR-0040 — semantic standards plane, ADR-0041 — deterministic validation and the development model portfolio

Amendments

• 2026-09-07 — New-model namespace policy selected by directing authority. ODR-0065 refines §2 with flat opda.org.uk/ns/{bc}/ namespaces, separate kernel/common/meta/facet responsibilities and the /id/ instance convention. Existing context codes are reused, not new boundaries discovered; term approval, candidate migration and publication remain separate.
• 2026-08-03 — Accepted by operator. Selected a first-principles Property-Pack scope; rejected a Builder dependency and five non-OPDA concerns; retained the other nine; and required one bounded-context or common-boundary home for every OPDA resource.
• 2026-08-03 — Pinned Builder execution profile adopted. Inherit Builder’s accepted authoring and assurance contracts without its runtime or excluded target categories; human working groups retain meaning and promotion authority.
• 2026-08-03 — OPDA governance clarified. Constitutional authority and the technical change process govern human review and promotion; incomplete release mechanics must be ratified before a normative ontology release.
• 2026-08-04 — Governance excluded from ontology scope. Category 6 remains an operational Builder and OPDA control; the ontology retains eight concerns and excludes six.
• 2026-08-04 — Semantic standards target fixed. Generate RDF 1.2, SPARQL 1.2 and SHACL 1.2 output with explicit conformance levels and fail-closed qualification; do not claim full RDF or SHACL Union support without feature evidence.
• 2026-08-04 — Candidate implementation recorded. Generated an isolated 451-item, context-owned review candidate and dual-provider calibration evidence. Deterministic candidate gates pass; route qualification and every semantic disposition remain pending human review.
• 2026-08-19 — Accelerated determination and canonical route adopted through ADR-0075. The current candidate is the Property Pack ontology component of SPDTF. The Technical Working Group reviews it as-is and makes the initial determination by the end of September 2026; prior review by every bounded-context group is not a prerequisite. Domain, scheme and Interoperability groups subsequently review the meanings they own and may propose controlled, versioned changes. The complete public corpus moves to /spdtf/property-pack/** without compatibility routes. This does not turn technical validation into semantic approval, conflate the Technical Working Group with Property Technology, or authorise an unrecorded release.
• 2026-08-22 — Chair-authority terminology correction. SPDTF is the first collaboratively authored scheme draft, not a numbered successor to an endorsed predecessor. The PDTF schema and schema-derived ontology remain distinct, attributed evidence inputs. This decision governs an accelerated component within the schema-to-scheme continuation; stable /pdtf/** identifiers and historical technical evidence remain unchanged.