Grounded in sources.
Honest about uncertainty.
We’re investigating whether a retrieval-grounded system can produce useful, inspectable simulations of judicial reasoning.
01 / Start with public evidence
Prioritize original, attributable material: opinions and other public writings. Record provenance, source quality, dates, storage permissions, and dataset splits. Discovery pages help us find evidence; they do not replace the evidence itself.
02 / Retrieve before reasoning
The planned baseline retrieves relevant source snippets for justice-specific agents. A general reasoning model would then use those sources to produce clearly labeled simulated memos. Fine-tuning is a later experiment that must be compared with a properly evaluated baseline.
03 / Keep the test honest
Historical case inputs must not reveal the decision they ask the system to predict. Final opinions, vote labels, related answer material, and post-decision publications must remain outside simulation inputs. Split controls exist in the scaffold; the full evaluation harness remains future work.
04 / Make limitations visible
The proposed deliberation stage would compare simulated arguments and report reasoning, citations, and uncertainty. It will not represent private deliberations or statements actually made by a justice. The current empty-corpus prototype returns undetermined instead of inventing an evidence-backed vote.
Build in public
Our goal is an open-source research project. A public code release and contribution channel are still being prepared. Until then, this journal shares our methods, decisions, and verified progress. Follow the RSS feed or share an article’s URL to follow the work.