Legora builds legal research on an ontology and citator: what you must be able to check per workflow

You must now be able to demonstrate, per legal workflow, that any AI-powered research tool grounds its citations in consulted sources, documents the scope of its underlying legal ontology, and maintains audit trails of which sources informed each answer. Grounding — whether a citation actually supports the claim — is distinct from citation presence alone, and you are responsible for verifying both.

The prompt is an analysis of 14 September 2026 of how legal AI grounds citations within an ontology and citator, which argues that the structural integration of legal ontology and AI-native citation measurement makes grounding testable and verifiable per matter. The analysis examines a benchmark that measures separately whether claims carry citations and whether those citations match the sources the AI actually consulted. In our assessment, this development means you can no longer treat citation presence as sufficient evidence of reliability; you must now treat grounding — the match between source and claim — as a measurable control point, and you must be able to demonstrate that threshold per task before deployment.

What does grounding mean in practice, and why does it differ from citation presence?

A citation that appears in an answer is not the same as a citation that actually supports that answer. This distinction is measurable and material. When an AI system returns a legal research answer with source references, those references may be present without being accurate — a known and recurring failure mode in legal AI. A benchmark that tests grounding separately from citation presence makes this gap explicit and verifiable. You can then use grounding scores as a deployment threshold, setting a minimum acceptable match between the sources the system consulted and the claims it made.

Which concrete controls must you be able to demonstrate per workflow?

1. Document the ontology scope and currency — record which version of the legal ontology the system uses, which jurisdictions and practice areas it covers, and when it was last updated.
2. Establish a grounding threshold before deployment — define the minimum acceptable grounding score (the proportion of claims supported by consulted sources) for each workflow, and do not deploy below that threshold.
3. Maintain an audit trail of sources per answer — ensure that for every answer returned, you can retrieve the specific sources consulted, the citations made, and the grounding score achieved.
4. Test citations against source material — before relying on an answer in a matter, verify that cited sources actually contain the propositions attributed to them.
5. Record ethical walls and confidentiality controls — document which data access restrictions, client confidentiality rules and regulatory requirements apply to each workflow, and confirm the system respects them.

How does ontology scope affect your control obligations?

The legal ontology underlying a research tool determines which sources, precedents, statutes and regulations the system can access and reason about. A tool built on a comprehensive ontology of law can ground citations across a wider range of legal material; a narrower ontology limits what can be verified. You must therefore know the scope of the ontology you are using — which jurisdictions, which practice areas, which types of source material — and confirm that scope matches the work you are asking the tool to perform. If you are researching a matter that touches jurisdictions or practice areas outside the tool's ontology, grounding becomes impossible for those elements, and you must know this before you deploy the tool.

What governance structures support verifiable legal research?

Structured legal research is not tied to a single supplier or interface. A research layer built on an ontology and native citator can be accessed through multiple front ends — whether a vendor's own interface, a large cloud provider's AI assistant, or a firm's internal workflow system. This means you can evaluate the research layer separately from the interface. The governance question then becomes: which sources does the research layer consult, how are citations grounded, and which audit trails does it maintain? These are properties of the research layer itself, not of whichever interface you use to access it. For workflows involving confidential or high-risk information, you must also confirm that the system respects ethical walls, maintains confidentiality, and logs usage in a way that meets your regulatory obligations.

What can tooling do, and what remains your responsibility?

A verification console can make grounding visible per matter and workflow, displaying which sources informed an answer and which corrections or disagreements occurred between models. Such a tool does not promise correct answers and cannot eliminate the risk of hallucination. What it does is make control possible by rendering verification steps visible. The professional final judgement — whether to rely on an answer, whether to investigate further, whether the grounding is sufficient for the matter at hand — remains yours. Tooling can structure your verification process and make it auditable; it cannot replace your assessment of whether the work is fit for purpose.