A Manifesto · v0.3

Working document · Updated 2026-05-20

The Ontology Project · A Manifesto

The problem we name

Every organization building on shared meaning today inherits a quiet dependency: the knowledge structures beneath its data belong to someone else. Data can be exported. Ontologies cannot. They are shipped as borrowed catalogs inside commercial platforms, seed schemas a customer can use, extend, and depend on, but never take with them. When the contract ends, the data comes out in tables. The knowledge stays behind.

This was the original market failure. A new one is now being layered on top of it.

In May 2026, Anthropic announced a $1.5 billion joint venture with Blackstone, Hellman & Friedman, and Goldman Sachs to deliver enterprise AI transformation services. OpenAI matched it the same week with TPG and Bain Capital. Palantir Technologies validated the pattern at over $325 billion in market capitalization. The model provider, the engineering capacity, and the deployment relationship now sit inside one external entity. The pitch undercuts the strategy houses on slides and the systems integrators on code. The financial sponsors supply the capital and the first clients from their portfolio companies.

The trade press calls it consulting rebuilt from the model up. The press releases call it additional operating capability for the ecosystem. Both readings miss the structural shift. Vertical integration of model provider with implementer with deployment relationship compresses what were previously four dimensions of governance sovereignty (semantic, enforcement, access, container), named by Frédéric Verhelst, PhD in his Agentic AI Capability Stack framework, into one counterparty. We name a fifth that the joint-venture pattern makes unavoidable: provenance independence. Without it, the other four can be reasserted on paper but not reconstructed in evidence.

"If organizations are going to govern agentic AI safely, they must own their semantic infrastructure. Community-driven, open-source initiatives like T.O.P. are exactly the kind of federated infrastructure the industry needs to prevent vendor lock-in and make semantic sovereignty operationally viable."

Frédéric Verhelst, PhD, Author of The Agentic AI Capability Stack

This is the procurement event most CDOs will not recognize until eighteen months in. The transformation contract is not what it appears to be. It is a transfer of the operational graph (the underlying knowledge structure your AI runs on top of) at enterprise scale, executed across private equity portfolios at speed.

"Speed has a price. The price is sovereignty."

Frédéric Verhelst, PhD

The cold-start tax (the time, money, and risk of building meaning infrastructure from scratch for each new deployment) was always real. Every new knowledge-graph deployment paid it, one customer at a time. That tax is steepest in the domains where getting meaning wrong costs the most: in healthcare and life sciences, where it is measured in patients; in energy and process industries, where it is measured in safety; in manufacturing, claims, and regulatory, where it compounds into exposure that no remediation budget can absorb. Independent reviews put the failure rate of knowledge-graph projects above 70%, with enterprise-scale implementations reaching $10 to $20 million and three years before self-sufficiency.

The cold start is the lever the rebundled providers now pull. An organization that cannot start its own ontology accepts whatever ontology arrives bundled. The ontology that arrives bundled is owned by the counterparty. The two market failures interlock: the cold start created the demand for borrowed meaning; the rebundling captures it.

What we owe the human

For four decades, enterprise software taught its users that the cost of automation is the user. Oracle, SAP, PeopleSoft, SAS, and the platforms that followed shipped on the assumption that the database was the artifact and the human at the keyboard was a shape that would have to bend to fit. Users learned the screens. They translated their mental model into the system's vocabulary. They reconstructed by hand what the system would not connect, and they did it every day, in every shift, in every audit, in every regulatory inspection. We have a name for this: digital archaeology. It is the toll the operator pays so the database can keep its assumptions.

Knowledge graphs are the first chance enterprise software has had to repay that debt. An ontology that begins with the operator's mental model, what a clinical research coordinator carries in their head when they walk into a site, what a process engineer carries when they read a batch record, closes the gap that no UX layer alone has ever closed. This is why T.O.P. starts at the conceptual layer that no standards body owns: the working vocabulary of the operator. Object-oriented elicitation methods, including OOUX, are useful tools for surfacing that vocabulary in workshop settings. They are not the ontology. They are raw material that the pipeline then formalizes through controlled vocabulary, taxonomy, SHACL property shapes, SKOS-XL thesaurus, OWL classes with PROV-O and BFO alignment, and operational projection through NGSI-LD. The operator names the entity. The pipeline makes it rigorous. The commons makes it sharable.

We believe AI's purpose is to be a human exoskeleton, not a human replacement. A pulley system for cognition. The thing that lets a clinical scientist who is exhausted from chasing eight signed delegation logs find the ninth in a quarter-second instead of a quarter-day. The thing that tells an auditor the chain is intact before they have to walk it. The thing that lets a process engineer compare a deviation against three prior matching deviations without standing up an entire data warehouse to do it. AI without grounding hallucinates. Grounding is what turns a pattern-matcher into an exoskeleton. The reference graph is the grounding.

For the practitioners who make this work, the ontologists, the information architects, the data engineers who have spent careers arguing for their seat at the table, TOP will change the day-one job. Once a reference graph ships under the project, the first conversation in a knowledge-graph project will no longer be a defense of the discipline. The reference graph will exist. The cross-walks will work. The operator-grade meaning will be there. The practitioner will walk in adding value to an institutional asset, not justifying the existence of one. That value will compound with every subsequent contribution, attributed and traceable. Their customers will stop being researchers of the past and become architects of the future.

This is what we mean by sovereign meaning. Not just that an organization owns its ontology. That the humans inside the organization, the operators and the practitioners both, get back the cognitive load that decades of mismatched systems have been silently extracting.

What we value

We have come to value:

• Shared reference graphs over borrowed catalogs.
• Exportable knowledge over rented schemas.
• Community-governed canon over single-vendor gravity.
• Interoperable interfaces over proprietary APIs.
• Domain fidelity over abstraction theater (models that work over models that look good).
• Versioned commons over snapshot generosity (knowledge whose curation is the work, over knowledge that is shipped and goes stale).
• Sovereign meaning over compressed counterparty stacks (control of your own knowledge structure over single-vendor bundles you cannot exit).
• Governance expressed in architecture over governance bolted on (rules that live in the machine-readable artifacts over rules that live only in policy documents).

The items on the right have value. The items on the left are what we are building.

What we commit to

• Open licensing. Every reference graph we publish is licensed for reuse, redistribution, and commercial application without permission.
• Versioned releases. Reference graphs ship with semantic versions, changelogs, and migration notes. A commitment not to break downstream users silently.
• Governance with rotation. Stewardship is a role, not a tenure. Domain working groups elect and rotate editors on a published cadence.
• Domain-native working groups. Clinical research first. Adjacent healthcare and life-sciences domains follow. Energy, process industries, and manufacturing in parallel as practitioners step forward. The domain decides the ontology; we provide the process.
• A public query interface. Reference graphs are not just downloadable files. They are queryable from day one, by anyone, without an account.
• Interoperability over ownership. We commit to open standards for representation and federation before we commit to any implementation.
• Credit that is verifiable. Contribution is signed, attributed, and traceable across versions. We will not let the record be rewritten.
• A door out. Everything we host can be taken and run elsewhere. If we ever become a chokepoint, we have already failed our own test.
• Industry agnosticism by design. The foundation, the federation pattern, and the commons concepts work for any domain that has paid the cold-start tax. The first reference graph is healthcare; the architecture is not.

These commitments are not novel. Norway's Digitalisation Agency maintains a framework called Orden i eget hus (order in one's own house) that runs the same governance pattern at national scale across the public sector: three formal roles (data owner, data steward, data coordinator), a concept catalog published as machine-readable RDF through Felles datakatalog at data.norge.no, a five-point maturity model that lets organizations measure where they are, and legal anchoring in legislation that makes participation non-optional for public organizations. The Norwegian and European application profiles of SKOS (SKOS-AP-NO, SKOS-AP) describe how the vocabularies federate across catalogs. The question of whether governance like this can work at scale is already answered. T.O.P. is the same pattern, industry-anchored rather than government-anchored, applied to domains where the cold-start tax is steepest.

What we ship first

The founding reference asset, the first commitment this manifesto stakes the project to, is a clinical research reference graph anchored in operator vocabulary and formalized through the full ontology pipeline: controlled vocabulary, taxonomy, SHACL property shapes, SKOS-XL thesaurus, OWL classes with PROV-O and BFO alignment, and NGSI-LD operational projection. When built, it will name the entities, relationships, and attributes that clinical-research work actually turns on (sponsor, study, site, participant, visit, investigational product, oversight body, event), in a form that can be rendered as a knowledge graph, extended by the community, and consumed by any platform.

This is the first reference graph TOP commits to building. We will build it so that no one starting work in this domain has to begin from a vendor's seed. Building it will be our proof that we ship, not just argue. The next reference graphs already in the queue: in healthcare and life sciences, CMC, drug discovery, regulatory, commercial, claims, and healthcare delivery; in energy and process industries, the analogues to ISO 15926 and CFIHOS; in manufacturing, batch processing and supply chain. Each ships under its own working group with its own domain-native vocabulary, importing the same TOP commons.

How to query

We are building toward a knowledge graph of knowledge graphs, a federation where reference graphs from different domains remain their own artifacts, governed by their own communities, but can be discovered, cross-walked, and queried through a shared interface.

The proposed foundation is NGSI-LD, an open standard for context-aware, linked-data exchange. It is the candidate because it treats entities as first-class, speaks JSON-LD natively, and federates without requiring a central authority. We will validate it by using it.

We will describe the catalog itself using DCAT, the W3C standard for data catalog vocabulary, with its European application profile DCAT-AP for federation alignment. DCAT is used by data.gov, the European Data Portal, and thousands of national and institutional catalogs worldwide. Describing the T.O.P. catalog as DCAT means the catalog and the reference graphs will live in the same RDF foundation: metadata and data, queryable through the same interface, with provenance recorded in the same provenance ontology (PROV-O) that grounds the graphs themselves. We do not invent a catalog format. We use the one the open-data ecosystem has already converged on.

What T.O.P. is not

T.O.P. is the commons. It is not an implementation platform, a deployment tool, or a vendor product. The reference graphs we publish are inputs to many downstream systems. Some of those systems will be commercial. Some will be in-house. Some will be other open-source projects. The architecture deliberately separates the commons from the implementations that sit on top of it. If we ever become the only place a reference graph runs, we have failed our own test.

An organization adopting a T.O.P. reference graph still needs the people, the tooling, and the operational discipline to maintain, validate, version, and deploy it inside their environment. The commons reduces the cold-start tax. It does not eliminate the steady-state cost of running an ontology in production. We say this plainly because the alternative is to oversell, and overselling is what landed the industry where it is.

Invitation

v0.1 was deliberately unsigned. The document had to come first. v0.2 stands open for endorsement.

What we ship next, the first reference graph, the governance charter, the contributor process, will be shaped by the people who show up to this draft and mark it up. The first reference graph is clinical research. It has not shipped, and the build has not begun in earnest; the architecture and this manifesto sit ahead of it. If you believe knowledge is the work and data is only what the work enables, read this and argue with it. Where we are wrong, say so. Where we are vague, push.

We are not asking for followers. We are asking for the first edits.