Our method
Make the basis for the work visible.
AI can produce a confident draft before the important questions are settled. What sources support it? Which checks were performed? What is missing? Who may decide the work is ready? PriorityLex designs AI-assisted work around those questions, using the four-layer method Thomas Potter set out in his working paper.
Four layers, one record.
Three of the layers follow the structure of the verification problem: some questions can be looked up, some must be judged against the source, and some belong to the person responsible for the result. The fourth, configuration, sits beneath them.
Matter 0417 · Demand letter · Draft 2
| # | Check | Layer | Result |
|---|---|---|---|
| 01 | Citation resolves to a reported case | L1 | PASS |
| 02 | Quotation matches the retrieved opinion | L1 | PASS |
| 03 | Parallel citation found in sourceSource does not index this reporter. Held, not passed. | L1 | UNRESOLVED |
| 04 | Case supports the stated propositionHolding is narrower than paragraph 3 says. | L2 | FLAGGED |
| 05 | Attorney decision on 03 and 04Revised ¶ 3. Citation confirmed in the reporter. T.P., 10:42. | L3 | RECORDED |
Fig. 1. A fictional record showing each layer at work.
Deployment configuration
Before work begins: the permitted information, the approved tools and what the system may do.
Programmatic verification gates
Questions with a checkable answer are resolved against authoritative sources. Unresolved is never passed.
Independent adversarial review
A separate reviewer reads the draft against the retrieved sources and looks for what does not hold.
Concentrated human judgment
A person with authority decides each flagged question. The decision and its reason are recorded.
Layer 0 · Deployment configuration
Define the purpose, permitted information, approved tools and actions the system may take. Identify the authoritative sources and the people responsible for maintaining them. Access arrangements should match the assignment, including whether a system may only read information or also change it.
Layer 1 · Programmatic verification gates
Use programmatic checks for questions with a determinable answer: whether a citation exists, whether a quotation matches, required fields, dates, calculations and document versions. Each check returns pass, fail or unresolved. A claim the source cannot resolve is held, never counted as passed. Whether a source supports a conclusion still requires interpretation, which belongs to the next layer.
Layer 2 · Independent adversarial review
A review distinct from the drafting step examines support, omissions, contradictions and competing explanations against the retrieved sources. Where a separate AI reviewer is used, it does not see how the draft was produced, and its configuration and performance must be tested; a second model’s agreement is not proof that a conclusion is sound.
Layer 3 · Concentrated human judgment
Present the flagged questions to the person authorized to decide, each with its reason and source. Record the decision and its basis. A workflow should make uncertainty visible and provide a usable route to resolve, restrict or stop the work.
Fit the method to the institution
In legal representation, the attorney determines the legal advice, strategy and work delivered to the client.
In law-firm support, the supervising attorney defines the assignment and reviews work within the firm’s professional and operating arrangements.
In municipal government, the workflow sits within wider governance: policy, procurement, delegated authority, staff responsibilities and records requirements. Officials retain authority over official action.
Test it and maintain it
An implementation needs representative tests before acceptance: the defects the checks catch, the errors they miss, the questions they raise unnecessarily and the workload passed to reviewers. Changes in a model, source, policy or assignment can change performance, so continuing oversight assigns responsibility for updates and retesting.
The method specifies functions and leaves the choice of tools open. No model or framework eliminates error; the aim is a process whose checks, limits and human decisions can be examined and improved.