Reference model v0.1
A working reference for discussion: shared vocabulary between legal, risk and engineering. Version 0.1 is intentionally compact; extend it in your own templates and policies.
Use this page as a checklist: define route policy, set gate thresholds by destination and reliance, and confirm that matter state plus audit are captured before output is used.
Design commitments
A compact set of design commitments that anchor the model.
- Legal AI should be designed around coordinated work, not isolated answers.
- The firm should own the routing layer between legal intent, model execution and human judgement.
- Controls should sit at the point of reliance, not only at the point of generation.
- Matter state should be treated as a live operational record, not a folder of documents.
- Evaluation should measure rework, variance and degradation over time, not just first-pass accuracy.
What is orchestration?
Orchestration is the layer between legal intent and execution. It decides what work is being requested, what context is required, which path (model, tool, human, policy check) should run, whether the output may proceed, what is logged and how matter state updates. It is not a synonym for automation. It implies sequencing, coordination, judgement and control. The distinction matters even more for agentic AI, where systems may plan, call tools, update records or trigger follow-on actions rather than only produce text.
Orchestration loop
Between isolated Q&A and coordinated work: intent moves through routing and execution, then gates, state, audit and monitoring before the next task.
Each cycle connects intent to execution, then gates reliance, updates matter state and the audit trail, and feeds monitoring back into routing.
Control point
Primary flow
Demos, playbooks and writing all use this backbone; treat it as the shared spine for design and policy discussions.
- A legal task is requested.
- The system identifies the type of work.
- The system gathers the right context.
- A routing decision is made.
- The task is executed by a model, tool or human.
- The output is checked against the intended use.
- An execution gate determines whether it can proceed.
- The matter state is updated.
- The audit trail records the decision.
- Monitoring continues where needed.
Terminology
Short definitions you can lift into architecture reviews, gate workshops and vendor questionnaires.
- Route
- The chosen path for a task: model tier, tool, environment, retrieval posture, human steps, cost controls and governance band, selected under a named policy version.
- Gate
- A required pause before an output reaches a destination: review, approval, block, logging or escalation. Gates encode policy at the point of use.
- Reliance
- How strongly the organisation or client is expected to treat the output as authoritative, from draft assistance through to binding use.
- Destination
- Where the output goes next: internal notes, working group, client channel, filing system, regulator, court bundle, and so on. Risk follows the destination.
- Mandate
- The scope of authority for automation or AI assistance on a matter or work type: who may trigger it, within which boundaries, and under which supervision.
- Matter state
- The live picture of phase, open tasks, pending decisions, obligations and risks, not a static file list.
- Audit event
- A durable record of who ran what, on which policy version, with which inputs, evidence, gate outcome and reviewer actions.
Concepts
How the pieces fit together in operating models and tooling discussions.
How to use this model
Practical entry points by function, without needing the whole mental map on day one.
Downloadable assets
Markdown files you can version in Git, paste into a risk pack or print to PDF from your browser.
- Reference model (Markdown)Portable copy of this page’s core definitions and flow.
- Routing policy templateSkeleton for task attributes, route dimensions and decision tables.
- Execution gate matrixDestination × reliance starter matrix with audit field ideas.
Open questions
- How should routing policies change across matter type, client terms and jurisdiction?
- What evidence is enough before AI-generated work can leave the firm?
- How can firms use multiple AI tools without fragmenting audit, policy and accountability?
- When should matter state trigger automation, and when should it only guide human judgement?
See the Routing Simulator, Matter State Viewer and Execution Gates demos for concrete examples. New to the site? Start here.