Glossary
The words, defined plainly
Legal AI arrives wrapped in vocabulary that vendors use loosely and buyers sign without unpacking. These definitions are ours — written so you can use each term in a vendor meeting and expect a precise answer back.
20 terms · vendor-neutral definitions
- AGPL-3.0
- The GNU Affero General Public License, version 3 — the licence on Vaultr's public repository. Grants broad use rights; its network clause can attach source obligations if third parties interact with a modified deployment. See the post on what it means for firms.
- Boundary (processing)
- The line a piece of matter data cannot cross during a task. In local mode it is the device itself — enforced by architecture, verifiable with a traffic capture, not by a policy page.
- Checkpoint
- A point where the system stops and waits for a human decision before applying a change. Vaultr's proposals — edits, drafts, answers you rely on — wait at checkpoints; nothing applies itself.
- Citation (span-level)
- A reference from a finding to the exact passage that produced it — page, line, clause — rather than to a whole document. The unit of verifyability in legal AI.
- Data residency
- A contractual or configuration promise about where servers storing data are located. Weaker than sovereignty: a region does not control access, logging or transmission.
- Data sovereignty
- An architectural property: whether data can leave your control at all. Enforced by how the system is built, not by where someone promises to put a server.
- Decision ledger (evidence ledger)
- The per-matter record of what the system read, proposed and cited, and what counsel decided. Persists locally and exports with the matter; deleting the matter deletes the ledger.
- Embedding
- A numeric representation of a passage's meaning used to find related text. Embeddings power retrieval; in local mode they are computed and stored on your device.
- Fallback (silent)
- The undisclosed switch to a different model or provider when the primary one fails. The failure mode legal buyers should fear most. Vaultr does not do it — the task stops with an explicit state.
- Grounding
- Constraining a model to answer from specific retrieved passages rather than from its general training. The discipline that makes citations possible.
- Hallucination
- Fluent output without a basis in the source material. The mitigation is not trust but structure: retrieval, grounding and citations that can be checked, plus refusal when no span supports an answer.
- Inference
- The act of running a model to produce output. Where inference runs — on your device, your firm's host, or a provider's — is the single most consequential fact about a legal AI tool.
- Local model
- A model that runs on your own hardware rather than a remote service. Typically smaller than frontier models, which matters less for grounded reading tasks than vendors of large hosted models suggest.
- Open-weight model
- A model whose trained parameters are published, so it can be run and inspected anywhere — including on your own machine, inside your own boundary.
- Prompt
- The text sent to a model. In external modes, prompts and retrieved passages are what leaves your control — which is why modes name what is transmitted before the first run.
- RAG (retrieval-augmented generation)
- The index → retrieve → ground → cite pattern behind serious legal AI. Explained step by step in the journal; its citation step is the part that matters for law.
- Red team
- Deliberately attacking a system to find its failures before someone else does. Included in a buyer's evaluation as exercises against the tool's citations and refusals.
- Retention
- How long data is kept, where, and deletable by whom. The place confidentiality actually erodes — ask for the table, not the paragraph.
- SOC 2 / ISO 27001
- Attestations about an organisation's controls over a period. Valuable when in scope and current; a red flag when borrowed from a cloud provider to decorate someone else's product.
- Zero data retention (ZDR)
- A provider commitment not to store what you send it. Relevant only to external modes — local mode needs no such promise because nothing is sent.
Terms that pair well with the journal: the RAG explainer for grounding and retrieval, residency vs sovereignty for the boundary vocabulary, and reading a security page for the weasel patterns these definitions arm you against.