What are the non-negotiable design rules for legal AI?
Reliability designed into the architecture, not inspected in after the fact.
14 pieces · 6 theses · 1 frameworks · Program RSS

Most software is allowed to degrade gracefully. A search engine that returns a slightly worse result is still useful. Legal AI does not have that luxury: an output that sounds right and cannot be defended will be filed, cited and relied on, and the cost falls on a client who never saw the machine. This program asks what the non-negotiable design rules are for systems whose outputs end up in front of judges, and it answers with architecture rather than aspiration.
The Four Laws of System Design for Computational Law state the rules in order of precedence. The Zeroth Law forbids presenting outputs with unwarranted confidence. The First requires that every result trace back to its inputs, reasoning and authorities. The Second requires that every legal proposition rest on validated authority rather than on whatever a model absorbed in training. The Third makes reasoning components independently testable and version-controlled. The program's other work applies those laws to concrete builds: why retrieval over chunked documents is not legal reasoning; why a benchmark must ask whether a system surfaced the proposition that does the legal work, at the right stage and with confirmed status, rather than whether it avoided a fake citation; how a small company of narrow agents behaved once certain decisions sat behind red lines reserved for humans, and why an off-switch needs as much design as a start button; and what happened when a blind eDiscovery review ran with no humans in the loop and its failed gates were published alongside its results.
Why this matters: reliability cannot be inspected into a legal AI system after the fact. A verification layer bolted onto a generator catches some errors, misses the structural ones, and can teach the generator to cite less rather than reason better. Guarantees belong in code, not in prompts, and a system must be able to halt without producing a confident answer. The Institute publishes its design rules so that any vendor, its own sponsor included, can be held to them.
Theses from this program
Reading order
- 01The Four Laws of System Design for Computational LawFour ordered design rules for legal AI
- 02Why Retrieval Is Not ReasoningInputs, Rules, Inference and Outputs in the Architecture of Legal AI
- 03What a Filing-Grade Benchmark Must MeasureDispositive recall under doctrinal and procedural constraint, and why fabrication rate is no longer the headline
- 04I Built a Tiny Company Out of Robots, Gave Them House Rules, and Nearly Finished a Hard ThingFive narrow agents, red lines, and the off-switch nobody designed
- 05We Ran a Blind eDiscovery Review with No Humans in the LoopOpen-sourcing a TAR engine, and publishing where it failed
- 06isResponsiveAn Open TAR Methodology for Court-Defensible Responsiveness Review