Why MOCA?

OpenMOCA is an open standard for portable grounding for AI. A MOCA package carries the knowledge an AI system answers from, how that knowledge is organised, and the original sources it came from, sealed by one digest and optionally signed. Open Readers in TypeScript, .NET and Python read, verify, search and cite it the same way, on a server or offline, whatever model, vector store or agent framework you use.

MOCA stands for Modular Ontology, Content & Attestation: knowledge in modular packages that can be assembled into larger ones, organised by an optional ontology, with attestations, signatures that prove who published it and who checked it.

It fits wherever an AI answers from your knowledge: RAG with citations you can check, trust and provenance, grounding for AI agents, ontology for AI retrieval, and sovereign and offline AI. Questions people often ask are in the FAQ.

The problem

Knowledge that feeds AI systems moves: a vendor ships product documentation to customers, a platform team serves policies to many product teams, a regulator publishes guidance that thousands of assistants quote. At every hop, five questions go unanswered:

1. Which content is this, exactly? Nobody can prove an answer came from version 4.2 of the knowledge base and not a stale copy.
2. Who stands behind it? “Verified” is a label anyone can type. Nothing ties it to the person who checked, or to the text they checked.
3. Is it still in force? Superseded policies and expired procedures stay in the index, and nothing tells the pipeline to stop serving them.
4. Where is the original? An answer cites a source, but the source stayed behind on the sender’s side, so nobody downstream can check the quote.
5. How is this knowledge organised? Procedures have steps in order, regulations have clauses that depend on each other, courses have prerequisites. Plain retrieval flattens all of it, so an assistant answers out of order or out of scope.

And more and more, knowledge is read where there is no service to ask: inside a mobile app, on an embedded device, or in a client network that allows no outside calls. The checks above have to run there too.

What already exists

The Open Knowledge Format (OKF) gives knowledge a common file format: Markdown with YAML frontmatter, a required type, and, since v0.2, trust fields such as generated, verified, status and stale_after. It is the right content format, and MOCA uses it as is.

OKF stops at the file on purpose. It has no package identity or version, no integrity, no signing, and no way to compose bundles or relate versions. OKF can reference sources, but because it defines no packaging, nothing guarantees that the originals travel with the knowledge or arrive unchanged. That is the layer MOCA provides.

In one sentence: OKF describes knowledge; MOCA ships it with its structure and its evidence, sealed when signed.

MOCA has two purposes, and every part of it serves one of them:

1. Moving knowledge between platforms securely and verifiably: identity, a digest, signatures, reviews, pinned members, the structure, and the original sources inside the package.
2. Reading and integrating packages locally, on edge devices or on infrastructure a client owns: an in-process MOCA Reader in the host’s language, offline search, structure operations, and bindings to the host’s framework.

What MOCA adds

• Identity and version for a whole bundle, with version numbers that say how much the meaning changed.
• A digest computed from the content, the same for a folder, an archive or a registry artifact, so any copy can be checked.
• Publisher attestations: who published exactly these bytes, verified with an Ed25519 key or a Sigstore identity the host trusts (signing and review).
• Review attestations: a reviewer signs that they checked the exact bytes of a node; the review stops counting the moment the node changes.
• Members pinned by digest and typed relations, so bundles compose reproducibly and a newer version can retire an older one.
• The original sources inside the package, covered by the digest. A Reader checks each quoted passage against its source and says whether it matched, with no network.
• Evidence, validity and audience on nodes: cite the exact sentence of a source, stop serving content outside its validity window, filter internal content before retrieval.
• An optional structure layer in RDF (Turtle). Every Reader understands a small core of SKOS and Dublin Core terms (hierarchy, parts, order, what requires what, what replaces what) and answers questions about it the same way. The same file can carry any other vocabulary, such as OWL, SHACL, PROV-O or your own domain terms. It travels in the package, covered by the digest, and its meaning is left to the application (structure guide).
• Ontology-guided search, through pluggable hooks. Before a search, an ingress hook can narrow it to a topic and everything under it. After it, an egress hook can reorder the results or add what they build on. Every Reader ships a default, and an application can supply its own for its domain, such as law or procedures. The Reader’s trust checks run after both hooks, so a hook can never return content they would remove (write your own hooks).
• One search entry point over whatever backend a host uses, from the package’s own text to a client’s vector store, with the same trust rules applied every time.
• A Reader contract and conformance corpus, so every implementation reaches the same conclusions.

Why not just use something else?

…OKF? Use OKF for content; MOCA is built on it. Add MOCA when bundles cross a boundary and someone needs to know which version, who published it and who checked it.

…a folder of Markdown in git? Git gives history within one repository. It does not give a reader outside that repository a way to verify who published what they received, which parts a named person reviewed, or that a policy is no longer in force.

…a vector database? A vector index is derived data for one model and one chunking scheme. Keep it; feed it from packages. MOCA gives it verified input and metadata to filter and cite with, and the Reader’s search applies the trust rules to what the index returns.

…a knowledge graph or GraphRAG? GraphRAG usually means a graph an LLM extracts from text. MOCA’s structure is curated with the content, from standard vocabularies, and travels with it; any graph tooling can load structure.ttl.

…OCI artifacts and cosign? Use them for distribution: a package is a set of files, which any registry can carry. OCI digests identify archive bytes; MOCA’s digest identifies content across archives, and is the hash of a standard BagIt manifest. And OCI has no notion of a reviewed node or a superseded policy.

What MOCA is not

• Not a content model, a vector database, an embedding format or an agent framework.
• Not configuration: nothing in a package can make a Reader contact anything.
• Not proof of truth: attestations say who published and who reviewed, and matched says a quote is in its source. Content is still untrusted input to a model, and structure narrows what is retrieved, not what a model says.

Honest status

MOCA is at 0.6.0, before 1.0. The specification, three Readers, the tools and the conformance corpus are complete and tested together, but nobody has yet measured how much MOCA metadata improves answers. That evaluation comes next, and fields that do not earn their place will be removed.