01What is a permissioned retrieval unit?
An LRU is the smallest useful piece of authorized legal or institutional information that can be retrieved with its source context intact. It may be a statutory provision with its proviso, a judgment passage with its issue and holding, or a contract clause with definitions and exceptions. It is not an arbitrary token window.
02Facts, provenance, confidence, terms
Our proprietary retrieval-unit architecture builds on a simple principle: the smallest piece of data that can answer a question on its own. Facts become searchable fields; provenance resolves to the exact source; confidence describes extraction and review, not a guarantee of legal correctness; terms govern each use. Proprietary software does not transfer ownership of the institution’s underlying knowledge.
03The evidence envelope
The unit carries a stable identity, version, source checksum, precise citation anchor, jurisdiction, legal context, effective dates, graph links and verification history. The source remains resolvable. OCR confidence, uncertain extraction and incomplete legal treatment must be visible to the reviewer.
04The permission envelope
A unit inherits source ownership, license restrictions, tenant and client identity, matter membership, ethical walls and permitted operations. Retrieval, quotation, export, evaluation, fine-tuning and training are separate uses. Permission is evaluated against both the principal and the stated purpose at request time.
05An illustrative decision
An associate assigned to Matter A asks for a clause comparison. The policy engine may allow retrieval of approved Matter A precedents and decline a similar clause from Matter B. Even for the allowed unit, exporting the source or sending it to an unapproved model remains a separate decision. The denial must not leak Matter B’s existence or title.
06Versioning and revocation
Source corrections create new versions rather than rewriting history. Withdrawal removes a unit from active retrieval and invalidates cached grants. Derived embeddings, exports and downstream recipients require an agreed deletion and notification process; a revoked source must not remain available through an old vector index.
07From source to auditable response
Source → Rights → Structure → Enrich → Verify → Permission → Retrieve → Audit. Each transition records an accountable actor and an outcome. A failed rights check stops ingestion; a failed scope check stops retrieval; an unresolved authority issue travels with the answer as a review requirement.
08Prepare once. Reuse under current authority.
The thin film converts approved source material into a structured surface that agents can scan. Expensive interpretation is concentrated at ingestion and revision, instead of repeated indiscriminate processing of the entire lake. Every query still checks current permissions, source validity and policy. Training, fine-tuning and retrieval are possible use classes only where separately authorized; none follows automatically from another.