LEGALIR / Model Families

One legal system. Many forms of intelligence.

Proposed specialist capabilities, supported by evidence and expert evaluation.

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

Orchestrate capabilities, not one universal model

The proposed model system combines general language models, purpose-specific fine-tuned LLMs, taxonomy classifiers, graph tools and deterministic policy services. A tax question may use a tax-specialist model and time-filtered graph retrieval; governance work may use an obligation extractor, entity resolver and review workflow. The orchestrator selects the sequence, while policy remains outside the model.

02

Sovereign models for institutions

Search finds the authority. For an institution that needs more, a bespoke sovereign model can reason over its own approved knowledge, inside its own boundary, with no pooling into any shared model. It is a proposed capability, described on its own page.

03

Modularity before model dependence

LegalIR’s model families describe legal capabilities and workflows. They do not claim that seven independently trained models exist. Retrieval-enhanced foundation models may be appropriate first; fine-tuning is considered only when the evidence, rights and evaluation results warrant it.

04

Evaluation is part of the product

A family needs a defined task, a rights-cleared reference set, expert-reviewed outputs and an abstention policy. Citation fidelity, applicable law, treatment history and access correctness matter alongside usefulness and latency.

05

Private institutional context

The Institutional family is scoped to a particular customer’s approved knowledge and workflows. A successful workflow for one firm does not justify reusing its source material, prompts or outputs for another customer.

06

Four orchestration patterns

A sequential pipeline classifies, retrieves, reasons and verifies. A parallel pattern asks distinct authorized specialists to examine a question and reconciles their source-linked findings. A routing pattern selects a purpose-specific model by task and jurisdiction. A reviewer pattern has a separate verifier challenge citations and assumptions before human approval. These are proposed patterns to evaluate, not deployed performance claims.

Your knowledge remains yours.

Enter the reading room ↗