Open Ontologies

One triple. Nothing added, nothing removed, blast radius zero. 901 consequences that were not there before.

Copy code
$ printf 'load base.ttl\nplan proposed.ttl\n' | open-ontologies batch -
#   the whole change:  ex:hasParent rdfs:domain ex:Person

added_classes         0
removed_classes       0
blast_radius          0 triples affected
risk_score            low
                      ────────────────────────────────────────────
conservativity        not_conservative_under_rule_table
new consequences      901          rule table owl-rl, in 0.04s

Each number from a shape diff tells you that this change is safe. But the change gave a new type to each individual that the property already had. The tool closes this gap. A text diff cannot show you the gap. The change is one correct line, and the text diff is one line long.

Then give the reviewer the proof. The run writes a certificate. A different person re-verifies months later. That person needs no instance of this software and no network. The command is oo-cert asserted.tsv derivations.tsv. The exit code is 0, and the theorem OOCert.certificate_sound covers the result.

The checker also refuses a forged proof. Write a false conclusion into the derivation file. The same checker exits 1 and names the rule that does not hold. This is the red edge in the figure above. This refusal, and not the headline, is the part that survives examination.

This is not an ontology editor. Use Protégé to draw class hierarchies. You run Open Ontologies on the change, before the change goes to production.

The tool is like Terraform, and this is on purpose. But the plan is semantic, not syntactic. A text diff is git diff, and you have git diff already.

You do not need a JVM. You do not need Protégé. The engine speaks MCP to Claude, to Cursor, and to other clients of that protocol.

See the engine at work

Three triples go in, and three triples come out. A person asserted only that ex:Northwind is in a sanctioned jurisdiction. The engine derived the need for enhanced due diligence. A different person can check that derivation. That person does not have to trust you, or this engine, or the model that wrote the ontology.

Watch the last seconds. Somebody forges one conclusion, and leaves the two premises exactly as they were. The same checker refuses the conclusion and names the rule. oo-horn printed each line in that terminal for the fixtures in tests/fixtures/horn/supplier/. A test runs the checker again. The test fails if the figure and the checker do not agree.

With a proof, and without a proof

If the tool measures a property, the tool says measured. If a prover gives an opinion, that opinion never takes the vocabulary of the checker. Read what the tool proves, and what it does not prove.

What the tool does

Run the checker yourself

The repository holds the three files. The output shows only the important fields.

bashCopy code
$ cd lean && lake build            # builds the checkers, core Lean 4, no Mathlib
$ F=../tests/fixtures/horn

$ lake exe oo-horn check $F/builtin_rules.tsv $F/asserted.tsv $F/good.tsv
{"ok":true,"verdict":"entailed","theorem":"OOCert.entails_of_builtin_horn",
 "means":"every conclusion is true in every model of the asserted graph"}

$ lake exe oo-horn check $F/builtin_rules.tsv $F/asserted.tsv $F/bad_conclusion.tsv
{"ok":false}                        # one IRI in the conclusion changed. exit 1.

$ lake exe oo-horn check $F/user_rules.tsv $F/asserted.tsv $F/good.tsv
{"ok":true,"verdict":"entailed_under_supplied_rules","theorem":"OOCert.horn_certificate_sound"}

The third line is the important one. The inference is the same. But you wrote one of the rules. Thus the rule is an assumption that the certificate carries, and it is not a fact that the certificate establishes. The verdict word changes. A test fails if that word stops changing.

mermaidCopy code
flowchart LR
  E["Untrusted engine<br/>Rust, or the pure-Python one"] -->|certificate| C["Verified checker<br/>core Lean 4"]
  I["Isabelle/HOL<br/>independent second kernel"] -.->|same bytes| C
  C -->|built-in rules| A["entailed"]
  C -->|your rules| B["entailed_under_supplied_rules"]
  C -->|forged| X["refused, exit 1"]

Run the tool on your own ontology

The repository ships those fixtures. Now do the same steps with a file that you write. These steps take one minute.

bashCopy code
mkdir /tmp/oo-demo && cd /tmp/oo-demo
cat > coffee.ttl <<'EOF'
@prefix ex:   <http://example.org/> .
@prefix rdfs: <http://www.w3.org/2000/01/rdf-schema#> .

ex:Espresso rdfs:subClassOf ex:Coffee .
ex:Coffee   rdfs:subClassOf ex:Drink .
ex:myCup    a               ex:Espresso .
EOF

export OPEN_ONTOLOGIES_STORAGE_MODE=persistent          # in-memory by default, see below
open-ontologies --data-dir /tmp/oo-demo/store load coffee.ttl
open-ontologies --data-dir /tmp/oo-demo/store reason --profile rdfs --certificate ./cert

Three triples go in, and three triples come out. The cup is a Coffee. The cup is a Drink. Espresso is a subclass of Drink. Each RDFS reasoner can do this much. The directory that the run wrote is the difference.

bashCopy code
lake exe oo-cert /tmp/oo-demo/cert/asserted.tsv /tmp/oo-demo/cert/derivations.tsv
{"ok":true,"asserted":3,"derivations":3,"theorem":"OOCert.certificate_sound"}

Now tell the checker a lie. Keep the premises, and forge one conclusion. The forged conclusion says that the cup is a Beer:

bashCopy code
cp -r /tmp/oo-demo/cert /tmp/oo-demo/forged
sed -i '' 's|example.org/Drink>\t<http://example.org/myCup>|example.org/Beer>\t<http://example.org/myCup>|' \
  /tmp/oo-demo/forged/derivations.tsv     # GNU sed: drop the '' after -i
lake exe oo-cert /tmp/oo-demo/forged/asserted.tsv /tmp/oo-demo/forged/derivations.tsv

jsonCopy code
{"ok":false,"asserted":3,"derivations":3,"first_rejected":2,"rule":"rdfs9",
 "conclusion":"<http://example.org/myCup> <...#type> <http://example.org/Beer>",
 "premises":["<http://example.org/myCup> <...#type> <http://example.org/Coffee>",
             "<http://example.org/Coffee> <...#subClassOf> <http://example.org/Drink>"]}

The exit code is 1. The output gives the number of the bad line. It names the rule. It also shows the premises, so that you can see that the premises do not support the conclusion.

What the proof contains

The proof is two tab-separated files. For the run above, the two files are 1.3 KB. asserted.tsv holds your claims:

Copy code
<ex:myCup>      <rdf:type>          <ex:Espresso>
<ex:Coffee>     <rdfs:subClassOf>   <ex:Drink>
<ex:Espresso>   <rdfs:subClassOf>   <ex:Coffee>

derivations.tsv holds one line for each step. Each line gives the rule, then the conclusion, then the premises for that conclusion.

Copy code
rdfs9    <ex:myCup> <rdf:type> <ex:Coffee>              <ex:myCup> <rdf:type> <ex:Espresso>        <ex:Espresso> <rdfs:subClassOf> <ex:Coffee>
rdfs11   <ex:Espresso> <rdfs:subClassOf> <ex:Drink>     <ex:Espresso> <rdfs:subClassOf> <ex:Coffee> <ex:Coffee> <rdfs:subClassOf> <ex:Drink>
rdfs9    <ex:myCup> <rdf:type> <ex:Drink>               <ex:myCup> <rdf:type> <ex:Coffee>          <ex:Coffee> <rdfs:subClassOf> <ex:Drink>

That is the full proof. It needs no model, no network and no vendor.

A checker reads each line. For each line, the checker derives the conclusion again from the premises of that line, under the rule that the line names. The checker then confirms that each premise is asserted, or that an earlier line concluded it. Anybody can write such a checker. This checker has a soundness theorem.

Who checks the certificate, and when

The certificate is a file. Thus the person who holds the file can check it, at any time that they choose.

The tool streams nothing, and the tool calls no home server. The engine and the checker share bytes on a disk, and they share no protocol. This is what makes the last two rows possible. An auditor who checks a claim next year needs the two files and a Lean build. That auditor does not need an instance of this software.

The certificate records a digest of the assertions that the run used. The file asserted.sha256 holds that digest, next to the two other files. A holder of a store runs certificate-check <dir>. The command reads the digest, computes the same digest from the store, and reports whether the two agree.

That answer has a limit, and the command states the limit. A digest binds a certificate to bytes. It does not bind a certificate to a state of the world. A store that changed and then changed back gives the same answer. The type-level form of this work is decision 0010, and that work is open.

One case needs a word, because you will meet it. reason writes its inferences into the store by default. A check after such a run finds more triples in the store than the certificate lists. The report names that cause, and it does not call the store a different graph. It says different graph only when a triple in the store is not a conclusion of the run.

Two defaults can cause you trouble. First, storage is in-memory, unless you set OPEN_ONTOLOGIES_STORAGE_MODE=persistent. Thus a load and then a reason starts from an empty store, and certifies nothing. The tool gives a warning, and you can miss that warning easily. Second, --data-dir is a flag and not an environment variable. Thus a demonstration without that flag writes into ~/.open-ontologies, next to your real work.

The discipline behind this work has a cost, and the discipline has earned that cost: what the rules are, and what each rule caught.

Three claims that used to travel on trust

A crosswalk states a match type. Nothing checked that statement. The engine now reasons over each side alone. It compares what each side entails, through the mapping itself. It then reports the tightest match type the evidence supports.

An exactMatch that the entailments do not support comes back downgraded. The report gives the reason. It also names every term the crosswalk does not carry. That second list is the one that disappears from most crosswalk files. The output is valid SSSOM, so your tools read it today.

A second question is sharper than a downgrade. The engine carries the translated claims into the target and reasons again. A clash means the target denies what the mapping carried in. That is a disagreement, and a person must settle it.

Contexts can disagree. Birds fly. Penguins are birds, and penguins do not fly. One graph that holds both is inconsistent, and this engine finds the clash. Each module reasons alone instead. A fact earns "true everywhere" when k modules of n entail it.

A number can now carry a certificate. oo-matcert recomputes a matrix product from the definition. It prints MatCert.mul_of_check when it accepts. Integers only, and that is the condition for the sentence to hold. Freivalds costs less and gives a probability, so it stays an opinion with its bound printed. Floating point reports a tolerance, because a proof over the real numbers says nothing about IEEE-754.

bashCopy code
open-ontologies batch plan.json   # crosswalk-certify, modules --threshold k, matcert

What the tool proves

A retrieval slice with 99% coverage can lose the one triple that an answer needs. A slice with 60% coverage can keep each claim that matters. Coverage is a proxy, and the proxy rises as the slice grows.

Thus a retriever that you tune on coverage learns to fetch more, and not to fetch the correct triples. Entailment preservation is the property that you want. The tool can decide that property here, and it gives one certificate for each claim. See decision 0007.

A measurement of the loss is the second-best answer. The best answer is a subset that can lose nothing. onto_module_extract computes such a subset. It computes a syntactic locality module over a signature. Each entailment of the full ontology over those terms is also an entailment of the subset.

That guarantee is a theorem of Cuenca Grau, Horrocks, Kazakov and Sattler, JAIR 31 (2008). The tool CITES that theorem, and no machine checks it here. No file under lean/ is about locality, and the report says exactly that.

The report names no theorem of this project. It offers a measurement instead. It reasons over the ontology and over the module to a fixpoint. It then reports each conclusion over the signature that the module does not reach. On the pizza ontology of this repository, the module is 238 of 1,345 axioms.

The run lost zero conclusions out of 2,583 differences. onto_conservative_check uses the same machinery for the lifecycle. It answers one question: does the addition of these axioms change any consequence over the names that the ontology already used? See decision 0011.

Install

bashCopy code
# macOS (Apple Silicon)
curl -LO https://github.com/fabio-rovai/open-ontologies/releases/latest/download/open-ontologies-aarch64-apple-darwin
chmod +x open-ontologies-aarch64-apple-darwin && mv open-ontologies-aarch64-apple-darwin /usr/local/bin/open-ontologies

# Linux (x86_64)
curl -LO https://github.com/fabio-rovai/open-ontologies/releases/latest/download/open-ontologies-x86_64-unknown-linux-gnu
chmod +x open-ontologies-x86_64-unknown-linux-gnu && mv open-ontologies-x86_64-unknown-linux-gnu /usr/local/bin/open-ontologies

# Docker
docker pull ghcr.io/fabio-rovai/open-ontologies:latest

# From source (Rust 1.85+)
cargo build --release --features embeddings,plugins,sql

For Intel macOS, for native Windows and for other systems, read docs/quickstart.md and docs/windows.md.

The serve command starts an MCP server. That server speaks JSON-RPC on stdin and stdout. Thus, at start, the server looks as if it stopped, while it waits for a client. This behaviour is correct. From a terminal, use the CLI subcommands instead, for example open-ontologies validate <file.ttl>.

Connect the tool to Claude

For Claude Code, add this block to ~/.claude/settings.json. For Claude Desktop, add the block to ~/Library/Application Support/Claude/claude_desktop_config.json:

jsonCopy code
{
  "mcpServers": {
    "open-ontologies": {
      "command": "/path/to/open-ontologies",
      "args": ["serve"]
    }
  }
}

Start the client again, and the onto_* tools are available. For Cursor, for Windsurf, for Zed and for VS Code, read docs/quickstart.md.

Stars

What the box contains

One loop: plan a change, apply the change, watch for drift, certify what the engine derived, and rollback when the change was wrong. That loop is the whole of this front page, and that loop is the purpose of the tool.

The other parts have their own documentation and their own code. They are alignment, embeddings, PDDL planning, clinical crosswalks, the plugin marketplace and CIVeX. docs/tool-reference.md lists them one time, and this page does not list them. A large list of features is not the argument.

Some tools need an optional Cargo feature, and they return an error without that feature. Four tools need embeddings. Two tools need plugins. Two tools need postgres or duckdb. The published binaries and the GHCR image use the default feature set. Thus they do not carry those eight tools.

The Python package open-ontologies-lite also reasons now. It reasons in pure Python, and it needs no Rust toolchain. The same Lean binaries check its certificates. The package is a second engine. The lack of trust in that engine costs nothing, because the warrant was never in the engine.

tools/horn_differential.py runs both engines and the Lean checker over each RDF document in the repository. The two engines are not independent. They run the same algorithm over the same rule table, and the comments in the Python cite the Rust by file and by line. Thus their agreement is strong evidence against a transcription error, and it is almost no evidence against a shared misreading of a W3C rule. The independent leg is the Lean checker. The tool prints that caveat next to its agreement count at each run. docs/lean-certificates.md also states the caveat as a limitation.

With those parts, the repository holds a marketplace of 33 standard ontologies, clinical crosswalks, semantic embeddings and a lineage audit trail. It also holds a desktop Studio, with a virtualized ontology tree, an AI chat panel and an inspector in the style of Protégé. You need no JVM. You need no Protégé.

Documentation

Open Ontologies for teams

The engine in this repository has an MIT licence, and it keeps that licence. The engine does not give you a place for the evidence. You need a place that keeps certificates. You need a review of each change to an ontology, before that change ships. You need an auditor who can check an answer again, months later, with no installation.

We build tesseractsemantics.com for that purpose. If a wrong answer from your ontologies has a cost, a conversation is worth your time.

Stack

Rust edition 2024, one binary, no JVM. Oxigraph 0.5 for RDF and for SPARQL 1.1. rmcp for MCP over streamable HTTP. SQLite for state, for lineage and for feedback. Lean 4 v4.33.1 for the checkers, core Lean only, no Mathlib.

Tauri 2, React 19 and Tailwind 4 for the Studio. The full table is in docs/architecture.md.

Citation

• Open Ontologies: Tool-Augmented Ontology Engineering with Stable Matching Alignment. Fabio Rovai, 2026. arXiv:2605.09184
• CIVeX: Causal Intervention Verification for Language Agents. Fabio Rovai, 2026. arXiv:2605.09168

CITATION.cff holds machine-readable metadata. It also operates the "Cite this repository" button of GitHub.

Language of this page

This page follows the writing rules of ASD-STE100 Simplified Technical English. The rules are one idea for each sentence, the active voice and simple tenses. A descriptive sentence has a maximum of 25 words. A paragraph has a maximum of 6 sentences. The same term always keeps the same meaning.

tests/readme_simplified_english_test.rs measures those rules. That test fails the build if this page breaks them. The same test holds the translated page to the structure of this one.

The official STE dictionary of approved words is a licensed document, and this project does not hold a copy. Thus this page claims the writing rules only. It does not claim approved-word compliance. The technical names and technical verbs of this domain stay in use. Examples are ontology, triple, reason and certify. STE permits such terms for a technical domain.

License

MIT. Fabio Rovai maintains this project at Tesseract Semantics. If this project is useful to you, you can support it through GitHub Sponsors.