A prediction is only as good as the evidence behind it.
Why we’re starting with sources, uncertainty, and an open research notebook—not a leaderboard.
Opening the notebook
SCOTUS Seer is a research project exploring whether AI can help model Supreme Court reasoning. Our aim is an open-source system whose predictions can be inspected, challenged, and improved. This journal is where we’ll share what we build, what we test, and what doesn’t work.
The work before the forecast
The current project is in its source-inventory phase. The repository contains 54 source metadata records, including nine official opinion document links. These records describe where evidence may be found; they are not a processed training corpus. No documents have been processed, no retrieval chunks are populated, and we have no measured prediction accuracy.
Evidence first, simulation second
Our planned baseline retrieves relevant public writings before producing simulated reasoning. Fine-tuning is a later experiment, not an accomplished milestone. The existing command-line scaffold returns undetermined when it has no evidence. That restraint matters: a convincing sentence is not the same thing as a supported claim.
What you’ll find here
Research notes will explain design choices in plain language. Changelog entries will record concrete changes and checks. Data snapshots will distinguish source links from processed documents. When experiments eventually produce results, we’ll publish the method, limitations, and failures alongside the numbers.
An invitation to follow the evidence
This is a public notebook for a project in progress, not a forecast service. The source release and contribution channel are still being prepared. For now, follow the journal or subscribe through RSS to see the research develop.