One Vocabulary, Every Model: Governing Meaning Across the Semantic Estate

Every growing data estate develops the same quiet disease, and the disease rarely announces itself as a data problem. Finance models revenue one way, the sales dashboard models revenue another way, and a third definition survives inside the board pack.

Each model can pass every local test and the company can still argue about the numbers in the same meeting. The argument was never mainly about arithmetic. Meaning drifts wherever models multiply without a shared vocabulary above them, and the drift stays invisible until two certified answers disagree in public.

The ontology direction in the Datus and Dosi October work points the other way. One governed vocabulary, mapped onto many models and versioned like code, gives teams a place to settle meaning once and let every dashboard, model and agent inherit the answer.

01 · The fragmentation problem

Every growing data estate develops the same quiet disease. Finance models revenue one way; the sales dashboard models revenue another; a third definition lives inside the board pack, reconciled by hand each quarter. Each semantic model is internally consistent, and the company still argues about the numbers — because the argument was never about arithmetic. Meaning drifts wherever models multiply without a shared vocabulary above them, and the drift stays invisible until two certified answers disagree in the same meeting. Agents inherit the disease at machine speed: an agent reasoning across models that define the same word differently will produce fluent, confident, contradictory work.

The cost of that drift is easy to underestimate because the cost arrives as ordinary friction. A meeting spends its first half reconciling totals instead of deciding anything. An analyst maintains a private translation between two dashboards. A careful answer carries a footnote explaining which revenue was meant. None of those moments looks like a system failure, yet together those moments tax every decision the estate was built to support.

02 · One ontology, many models

The architectural answer emerging in the Datus and Dosi October work is a separation of concerns. Ontology Mapping, split into the Apache Ossie standard, divides the world into three documents — the ontology, the semantic model, and the mapping between them — so that one ontology can span many semantic models. The vocabulary of the business (customer, order, account, and the relationships among them) lives once, governed once, and each underlying model declares how local fields and tables express those shared concepts. Models keep their local structure; meaning stops being local. New models join the estate by writing a mapping, not by renegotiating the language.

The three-document split also clarifies ownership. The ontology holds the language leadership agrees to govern. A semantic model holds a local structure built for a platform, a team or a workload. The mapping holds the honest translation between the two, where compromises and gaps can be seen rather than hidden inside a query. When a local model changes shape, the mapping absorbs the change. When the business changes meaning, the ontology records the change where every mapping can respond.

The shared language is not documentation about the system; the shared language is part of the system.

03 · Governance is a versioning problem

Treating vocabulary as a governed artefact changes the daily work of data leadership. A definition change becomes a versioned event — proposed, reviewed, released — rather than an edit inside one team's file. Derived concepts, such as the rule that makes a customer a Premier Customer, carry versions of their own, so historical reports can be read against the rule that was in force when the numbers were produced. Disputes move up a level, from whose SQL is right to what the business means — a shorter, more honest argument, held once, in the open, with the decision recorded where every model can inherit the answer.

Good governance here stays deliberately light. A proposed change needs a reason, a reviewer with authority over the concept, and a record that downstream owners can find. Heavy ceremony would push teams back toward private definitions, defeating the purpose. Light, visible versioning does the opposite: changing the shared language becomes normal, auditable work, and keeping a private dialect becomes the option that needs explaining.

04 · What agents inherit

Agents raise the stakes on all of this, because agents consume meaning literally and at scale. An ontology spanning the estate gives an agent a bounded business world to reason in: concepts that resolve consistently across models, relationships that can be traversed safely, rules that execute the same way in every tool. The OntologyAnalyst pattern — a specialised subagent working the concepts and mappings directly — shows the consumption model forming already. Context engineering, as practised at Ovia since December 2022, treats that vocabulary as core context: the shared language is not documentation about the system; the shared language is part of the system.

Bounded worlds matter for safety as much as for accuracy. An agent that can traverse named relationships inside a governed vocabulary has a clearer edge to stop at than an agent improvising joins across raw tables. Permissions, review and audit all become easier to reason about when the concepts an agent used are the same concepts a person can inspect. The vocabulary does not make an agent wise, but a governed vocabulary makes agent work legible.

05 · What to watch

Watch whether the three-document split — ontology, model, mapping — becomes the standard shape other engines adopt, whether mapping maintenance proves cheaper than definition drift (the bet the architecture makes), and whether governance ceremonies stay light enough that teams actually hold them. The estates that win the agent era will not be the ones with the most models. The winners will be the estates where every model, dashboard and agent means the same thing by the same words — and can prove the lineage of that meaning on demand.

The cost side deserves honesty as well. A governed vocabulary is not free: mappings must be written, changes reviewed, and someone must own the language with authority to settle disputes. The architecture's bet is that this visible, bounded cost replaces an invisible, unbounded one — the reconciliation meetings, the duplicated definitions, the agent errors traced back to two models meaning different things by the same word. Estates already pay for meaning; the ontology simply moves the payment onto the books, where leadership can see the line and judge the return.

A sensible starting point stays small. Pick the words that cause the most expensive arguments, govern those words first, and let the practice earn the right to expand.