Methods and ethics

01

Source hierarchy

Primary text first. Search tools are retrieval aids, not authorities.

• Use current statutes, regulations, rules, practice directions and full decisions.
• Treat snippets, headnotes, AI summaries and commentary as leads until the primary text is checked.
• Record source ID, version or retrieval date, pinpoint, query and inclusion decision.
• Keep adverse authorities and negative findings.
• Label participant account, sourced fact, system inference and unresolved question separately.

SourcesBC Human Rights Tribunal · BC Human Rights Tribunal

02

Data minimization

A person should not have to publish a private case file to prove the access problem.

Public journeys remove names, contacts, medical information, employer identity, file numbers and distinctive facts. They preserve only the causal sequence needed to understand the barrier.

The prototype will collect the minimum needed for routing. A funded inference layer needs documented retention, encryption, access logging, deletion controls and a threat model before handling private facts.

SourcesCanLII · CanLII

03

Evaluation before claims

A fluent answer is not the success metric.

• Citation validity: does the source and pinpoint support the proposition?
• Retrieval coverage: which reviewed decisions were found and which were missed?
• Difference detection: does the output identify facts that weaken the analogy?
• Procedural accuracy: is the rule tied to the correct forum and version?
• Abstention: does the system ask for missing facts instead of filling gaps?
• User comprehension: can a stressed user understand and export the result?
• Equity and language: does quality change across languages and fact patterns?

SourcesStanford Law School · Stanford Legal Design Lab

04

Corrections

Material changes stay visible.

Each output has an update date. A material error receives a dated correction naming the old claim, the new source and the effect. Corpus builds receive version identifiers so an earlier search can be reproduced.

05

Human role

The product organizes research; it does not decide the case.

The exported pack should make later work with a lawyer, advocate, clinic or institution more productive. Yale and Stanford point toward narrow tools, partner distribution, user research and measurable quality rather than a general chatbot presented as a lawyer.

SourcesYale News · Stanford Legal Design Lab · Stanford Legal Design Lab