The problem was not simply finding documents.
Cross-country legal research requires analysts to locate relevant instruments, determine whether a source covers the right jurisdiction and period, interpret provisions against a structured question, and preserve enough evidence for another researcher to review the judgment.
A single-pass assistant would collapse these activities into one opaque response. The design problem was to automate selected steps while preserving the team’s responsibility for evidence quality and final coding decisions.
My contribution
I translated the research process into a staged workflow, built the orchestration and agent logic, defined evidence and confidence states, implemented deterministic routing between stages, and developed the review outputs used by analysts.
I also designed the validation approach around two separate questions: whether the software behaved as specified, and whether the resulting evidence and analysis were suitable for domain review. Passing the first did not imply the second.
Three design choices defined the system.
Separate retrieval from substantive judgment
Discovery agents identify candidate sources. Later stages evaluate eligibility and support, preventing retrieval confidence from being treated as legal confidence.
Route evidence through explicit states
Structured status fields and rule-based routing determine what may advance, what needs more work, and what must be held for review.
Make human approval part of the architecture
The workflow produces a review queue with traceable evidence. Human review is a defined operating step, not a disclaimer added after generation.
Validation was layered rather than summarized by one score.
Software checks covered deterministic behavior, state transitions, expected schemas, routing rules, and edge cases. Pilot outputs were then compared with manually coded reference material and prepared for analyst review.
These checks support claims about pilot completeness and system behavior. They do not establish legal accuracy, organization-wide adoption, or measured productivity impact.
What the pilot established
The work produced a functioning multi-stage pilot, structured evidence and review artifacts, documented operating logic, and a handover package. It demonstrated how a broad research task could be decomposed into controlled, inspectable steps.
Planning assumptions anticipated lower manual processing effort and eventual integration into a recurring team workflow. Those are expected benefits, not observed results, so they are not presented here as performance outcomes.
Limits and public boundary
- Analyst review of substantive legal conclusions remains necessary.
- The pilot does not establish adoption, recurring operational use, or measured time savings.
- Internal code, prompts, datasets, reports, screenshots, and detailed test results are not public.
- The architecture shown above is a simplified reconstruction of top-level design choices.