A model that reasons is not enough; a sovereign model must cite. Grounding binds every answer to the nation's own statutes, registries, and records — with the source attached.
Retrieval starts by turning heterogeneous official sources into a structured, versioned corpus.
Statutes, regulations, gazettes, court decisions, land and company registries, and supervisory filings are ingested from systems of record — not scraped copies.
Documents are chunked along their real structure — article, clause, schedule — so a retrieved passage carries its citation coordinates, not just a paragraph of text.
Every provision is stored with its effective dates. The corpus can answer as of a date, distinguishing law in force today from a version since repealed.
Each ingested document's hash is written to the tamper-evident chain, so the corpus can be proven un-altered since a point in time.
Grounding quality comes from combining lexical precision, semantic recall, and reranking.
BM25-style exact matching catches statute numbers and defined terms; dense vector search catches paraphrase and concept. The two are fused before ranking.
A reranker reorders candidates against the actual question, pushing the genuinely responsive clause above merely similar text.
Embeddings and indexes are computed and stored on sovereign infrastructure. No query or document is embedded through an external API.
Multi-part legal questions are split into sub-queries and retrieved independently, so a compound question does not collapse to a single weak match.
The model is constrained to answer from retrieved sources and to show its work.
Answers are assembled from retrieved passages with inline references to instrument and article. A claim without support is withheld rather than invented.
When the corpus does not contain an answer, the model says so and points to the nearest authority, instead of filling the void with plausible fiction.
Each citation resolves to the exact registry record or clause, with its effective date and ledger hash, so a reviewer can verify the ground truth directly.
A user only ever sees grounding drawn from records they are cleared to read.
Access policy is applied during retrieval, not after generation, so restricted records never enter the model's context for an uncleared user.
Chunks inherit the classification and need-to-know of their source. A sealed filing cannot ground an answer for someone without clearance.
Every retrieval — who asked, what was returned — is logged to the ledger, producing a reviewable trail of which records informed which answer.
Grounding is only trustworthy if it reflects the current state of official record.
New gazette entries, amendments, and registry changes are ingested continuously. A repeal propagates so the model stops citing dead law.
Knowledge lives in the retrievable corpus, not baked into weights, so updating the law is a data operation rather than a retraining project.
Every corpus revision is versioned and hash-chained, so an answer given last quarter can be reproduced against the corpus as it stood then.
Talk to us about retrieval and knowledge in a sovereign deployment.