Open Ontologies Plan a change to a production ontology. See every consequence before you apply it. Then give the reviewer a proof that they can check without trust in you. Open Ontologies is written in Rust. It ships as one binary. tesseractsemantics.com English · 简体中文 Put in one spreadsheet. Get one ontology, with the evidence for each line. Then change one cell, and see the shape find the error. The page runs the…
Plan a change to a production ontology. See every consequence before you apply it.
Then give the reviewer a proof that they can check without trust in you.
Open Ontologies is written in Rust. It ships as one binary.
Put in one spreadsheet. Get one ontology, with the evidence for each line. Then change one
cell, and see the shape find the error. The page runs the released binary.
426 asserted, 259 certified, 1 rejected. The engine derived the green edges. A Lean 4
checker then proved them. One certificate holds all 259, and one run of the checker accepts it.
The red edge is a forged line. The same checker refused that line, gave exit 1, and named the rule.
The run supplies each count. Nobody writes a count into the caption.
The same machinery, on a file from a different group, two times. HQDM ships as two files.hqdmTop/hqdmFramework publishes an RDFS file. MagmaCore keeps a copy of that file, byte
for byte. gchq/HQDM publishes an OWL file. The two files fail different checks. The RDFS
file has no owl: term and no disjointness axiom. Thus no named class in that file can be
unsatisfiable. But the file has 23 terms that it uses as a class and never declares. It has
12rdfs:range declarations that name a relation and not a class. It also has 13
pairs of names that differ by one final underscore. Three of those pairs have the same domain and the
same range. The OWL file passes all three checks, and the OWL file is not coherent. The tableaux of
this engine finds 190 of its named classes satisfiable. The tableaux finds 39 unsatisfiable.
HermiT calls the same 39 unsatisfiable, and HermiT is an opinion in the vocabulary of this
repository. This figure proves nothing, and the figure does not claim a proof. A test computes each
count again, and the test also computes that intersection. For the source and the method, read
docs/assets/hqdm/PROVENANCE.md.
One triple. Nothing added, nothing removed, blast radius zero. 901 consequences that were not there before.
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 checks what a
change means and what it breaks, not only its text. 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
An ordinary reasoner
Open Ontologies
The answer
Northwind needs enhanced due diligence
the same answer
Why the answer holds
“the reasoner says so”
a certificate that names each rule and each premise
Who can check the answer
a second implementation can agree, and some reasoners give an explanation. No verified checker accepts either one
anybody, with a checker that shares no code with the engine
If the engine has a defect
a second implementation can disagree. Then you know only that one of the two is wrong
the checker refuses the answer, exit 1
If a person edits the output
you cannot find the edit
the checker refuses it, and names the line and the rule
If a rule was yours, not the standard’s
the report is the same
a different verdict word, and a test holds that word
What an auditor receives
a screenshot
a file that the auditor can check again
Guarantee on an unsatisfiability answer
asserted
none, and the tool says so
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
Capability
What you get
Reason over OWL and RDFS
Materialised inferences and a derivation certificate that a proved checker accepts
Use your own rules
SWRL, RIF Core or a Horn table, evaluated, with a verdict word that says the rules were yours
Validate against SHACL
A report from an evaluator with a measurement against the W3C suite, not an assertion of success
Ask if something is satisfiable
A finite model, replayed and checked, and not only a yes
Ask if something is inconsistent
A refutation, if one is certifiable. If not, an honest opinion from the engine
Retrieve a slice for RAG
Entailment preservation for each claim, because 99% coverage can still lose the one triple that mattered
CSV, JSON, XML, YAML, XLSX, Parquet, PostgreSQL and DuckDB into RDF
Give the problem to a prover
TPTP, CLIF, SMT-LIB and LADR from one translation. The tool names and counts what it cannot export
Work from an assistant
An MCP server, so Claude or Cursor operates all of it in conversation
Run the checker yourself
The repository holds the three files. The output shows only the important fields.
$ 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.
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.
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.
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:
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
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:
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.
Who
When
What they run
You, in the loop
at each run, before you trust an answer
lake exe oo-cert, next to the reasoner
A reviewer
when a change lands
the same command in CI, on the artefact that the run wrote
An auditor, months later
long after the engine moved on
the same command, on the archived files
A different agent
when it receives a claim from an agent that it does not trust
the same command, before it acts on the claim
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.
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.
The check compares where two terms sit in their hierarchies. It does not compare what the terms mean. Two sibling terms under one mapped parent look the same to it. So it can refute a match type, and it never makes a claim stronger. A parent that the mapping does not carry is a gap, and the report names it. A row with nothing to compare says so and is not certified.
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. Run it with crosswalk-probe.
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.
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.
The Docker image is for linux/amd64 only, so on Apple silicon run it under
emulation with
docker run --rm -i --platform linux/amd64 ghcr.io/fabio-rovai/open-ontologies:latest.
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:
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 release binaries have embeddings and plugins, so
they lack only the two tools that need postgres or duckdb. The GHCR image also
has sql, which turns on postgres and duckdb, so it carries all 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.
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.
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.
评论