Project: Shinshō-27B (codename)
Tl;dr: An Experimental Judicial Reasoning Model.
Shinshō-27B is our experimental quasi-reasoning model built on Gemma 3. Trained on approximately 20 million judicial opinions from the United States, United Kingdom, Canada, New Zealand, and Australia.
The aim is not another general-purpose legal language model. The aim is something narrower and more difficult: a model that captures how judges actually reason.
The Training Corpus
Twenty million opinions across five common-law jurisdictions. Appellate courts. Trial courts. Majority opinions, concurrences, dissents. The full texture of judicial decision-making, not a curated highlight reel.
We are interested in the reasoning itself — how judges weigh competing authorities, distinguish precedent, apply standards of review, and articulate holdings. This is different from knowing what the law says. It is knowing how law thinks.
Why This Matters
Most legal AI treats judicial text as content to be retrieved or summarized. But judicial opinions are not just information. They are performances of reasoning — structured arguments that follow particular conventions, reflect particular psychology, and produce particular kinds of authority.
A model that understands this distinction can do things a retrieval system cannot:
- Identify the actual ratio of a decision, not just its subject matter
- Distinguish holding from dictum, analysis from advocacy
- Recognize when reasoning is being extended, limited, or distinguished
- Anticipate how a court might reason about a novel question
Current Status
Shinshō-27B is not a productionised system. It is an active R&D project.
We are developing the model to test a hypothesis: that specialized training on judicial reasoning produces capabilities that general-purpose models and retrieval-augmented systems lack. The work is ongoing.
The Name
Shinshō (審証) — Japanese, meaning approximately “judicial proof” or “examination of evidence.” The 27B denotes the parameter count of the base architecture. We chose the name because the work is about understanding how courts weigh and evaluate — not just what they conclude, but how they get there.
What Comes Next
If successful, Shinshō-27B becomes infrastructure: a reasoning layer that can be applied to novel inputs, helping practitioners understand not just what courts have said but how courts are likely to think.
We’ll say more when there’s more to say.
Citational is an AI research and development lab dedicated to law. We build verification infrastructure and specialized models for legal reasoning.