LEGALIR / Model Families

Many specialists. One governed workflow.

Orchestrate legal language, taxonomy, graph tools and purpose-specific models.

Sign in to learn more ↗
The permissioned thin filmInstitutional records remain in a lower plane. A thin policy boundary lets only permitted retrieval units reach approved agents above. Restricted information stays below.APPROVED AGENTSSCOPED RETRIEVAL ↑POLICY BOUNDARYPRIVATE RECORDSFIG. 03 / PREPARE ONCE. RETRIEVE UNDER AUTHORITY.
PROPOSED INFRASTRUCTURE / IN DEVELOPMENT
01

Purpose-specific language intelligence

Fine-tuned LLMs may be developed for tax, governance, regulatory obligations, contract language and legal taxonomy when training rights and evaluation evidence support them. Smaller classifiers and deterministic parsers may be better for narrow tasks. The model family describes the capability, not a commitment to a single model architecture.

02

A routing layer with explicit constraints

The orchestrator receives the task, jurisdiction, matter scope, permitted endpoints and budget. It selects a workflow from approved configurations. A model request outside the grant is rejected even when the orchestrator predicts that it would improve the answer.

03

Sequential, parallel and reviewer patterns

Tax research may classify the issue, resolve the relevant assessment period, retrieve applicable authorities, invoke a specialist and verify each citation. Governance review may extract obligations in parallel with entity and policy mapping, then reconcile conflicts for human review. A second model can critique reasoning but cannot certify legal authority by itself.

04

A common evidence contract

All stages exchange unit identifiers, source anchors, policy decisions and structured uncertainty. They do not trade unrestricted data dumps. Intermediate outputs retain their original restrictions; transforming a unit does not make it public.

05

Evaluate the orchestration

Measure task quality, citation support, failure propagation, latency and cost for the complete workflow and each stage. Compare simpler baselines before adding agents. Log which configuration and model versions produced the result so a reviewer can reproduce the decision path.

Your knowledge remains yours.

Enter the reading room ↗