Evaluating NVIDIA Auto Ontology for CatProof registry questions
A proposed integration for asking better questions about agent ownership, source evidence and identity history.
CatProof editorial · · 5 min read
An agent appears in two systems under different names. One record identifies its creator. Another lists a technical owner. Your team has already decided that both records belong to the same business agent, owned by Finance.
Now someone asks: “Which Finance agents still have stale source evidence?”
Answering that question requires more than finding similar text. The system needs to know what counts as an agent, which records have been explicitly linked, who owns the business record, and what “stale” means.
That is where a semantic layer could help. We reviewed NVIDIA Auto Ontology to assess whether it could make CatProof registry information easier to query while preserving the decisions behind it.
Our conclusion: a focused, read-only proof of concept is worth testing. The integration described here is proposed. It is not an available CatProof feature, and we have no integration benchmark results to announce.
What CatProof records today
CatProof brings source observations together around business agent records. Original source identities remain visible, while teams maintain accountable ownership and make deliberate reconciliation decisions.
A similar name can help identify a pair to review, but it does not authorize a link. CatProof’s documented workflow lets reviewers link records, keep them separate or defer a decision. A stable source ID can retain its binding through a name change; a new ID requires a separate continuity review. Reconciliation guide.
Those distinctions matter when answering questions. A source-reported creator is not necessarily the accountable business owner. A service associated with an agent is not automatically another agent. An old observation does not establish that an agent has stopped running. Connector coverage, source refresh semantics.
What Auto Ontology could add
Auto Ontology provides interfaces for working with business terms, SQL attributes and relationships, alongside a natural-language query workflow. Its API exposes glossary and semantic-model operations as well as a chat endpoint. Its MCP implementation offers tools for inspecting meaning and asking questions. API specification, MCP tool definitions.
For CatProof, the useful possibility is a reviewed vocabulary over registry evidence. “Business owner” would refer to the business decision recorded in CatProof. “Source observation” would refer to what a configured source reported. “Linked records” would refer to explicit bindings.
That could support questions such as:
- Which active business agents have no accountable owner?
- Which agents have linked observations from more than one configured source?
- Which source connections have failed reads, and which agent records depend on their evidence?
These are candidate pilot questions. Their usefulness depends on correct mappings, reliable query results and clear evidence references.
How the proposed integration would work
We would begin with a sanitized snapshot from one test registry, loaded into a separate reporting database. Auto Ontology would query that database using a credential restricted to approved reporting relations.
The reporting model would preserve canonical agent IDs, source identities, explicit bindings and observation timestamps. It would exclude connector secrets, account data and raw provider payloads. Ownership and reconciliation decisions would continue to be made in CatProof.
An answer should include the SQL used, the relevant record IDs and the snapshot time. Auto Ontology’s MCP query implementation already separates answer text, SQL and rows; attaching verified CatProof evidence links and reporting freshness would be additional integration work. Query result implementation.
This design also keeps the two applications operationally separate. Their inspected dependency manifests require different versions of shared Python libraries. A service boundary lets us evaluate the query layer without forcing a dependency migration across the registry. CatProof dependencies, Auto Ontology dependencies.
What a semantic layer must preserve
The central requirement is that meaning cannot silently change authority.
An ontology may describe a relationship, but that description must not turn into permission to merge identities or assign an owner. CatProof’s reconciliation architecture keeps source evidence and business decisions separate. Any query layer should preserve that boundary. CatProof architecture.
Access control needs the same care. CatProof registry membership does not automatically transfer to another application’s database connection. The first pilot should use an isolated reporting database and deployment, with access enforced independently of generated SQL.
Data handling also needs review. Auto Ontology has catalog paths that store sampled values, and its query results contain database rows. A deployment must account for what enters model requests, embeddings, logs and retained conversations. Calling something “metadata” does not settle whether it is appropriate to send to an external service. Catalog candidate implementation.
Auto Ontology’s own code uses Apache-2.0, while its dependencies and model services have their own terms. Any packaged integration would need a distribution review covering the actual components shipped. License, third-party notices.
What would make the pilot worthwhile
We would test the proposed integration against reference SQL and ordinary reports using synthetic registry data. The dataset should include renamed agents, distinct agents with similar names, missing owners, multiple source records and failed source reads.
The evaluation should measure answer correctness, evidence references, isolation, latency and model cost. It should also test whether the system acknowledges unsupported questions instead of inventing a relationship or owner.
The practical question is whether this makes registry evidence easier to understand at an acceptable operating cost. If it does, a semantic query layer could help teams ask more useful questions about the records they have already reconciled.
CatProof’s role would remain clear: preserve the source identities, maintain accountable business records, and keep the decisions behind them available for review.
Explore CatProof reconciliation and its architecture documentation.
