Research / Legal judgment

Legal judgment is the research problem.

We work on the assignments where context changes the answer, experts disagree, and a plausible output can still fail the work.

Models need more than examples of acceptable language. They need the assignment, the decision, the outcome, and a test that can tell the difference.

Research agenda

Three questions we keep coming back to.

01

The Murky Middle

The obvious cases are easy to label. We study the decisions that change with the document, the instruction, the client, and what happens next.

Read the note 02

Train for the Outlier

The judgment worth teaching is not always the answer most people chose. We study how to preserve unusual but defensible expert reasoning.

Read the note 03

Where the Model Fits

We separate model failures from missing context, weak tools, agent logic, and evaluations that reward the wrong thing.

Read the note

Public record

Results and artifacts.

001

Training report / Published / 12 Jul 2026

Post-training a contract-review model

A supervised run on reviewed contract work, evaluated in the same serving lane on seen, unseen, and new tasks.

002

Legal environments / Open / 03 Jul 2026

Prime Legal Environments

Corporate-actions tasks where the agent reads the record, takes valid actions, and completes the required approval process.

003

Open model / Released / 25 Aug 2026

Octen Law 8B v1

A multilingual legal retrieval model released with its training note, evaluation results, data audit, and limitations.

004

Benchmark / External / Current

MTEB(Law, v1)

Eight legal retrieval tasks across case documents, statutes, legal Q&A, and legal summarization in three languages.

005

Design note / Proposed

A services-agreement RL environment

A proposed environment contract for revising one versioned services agreement against a bounded legal playbook.

How we report

Count the decisions.

A lift number needs a denominator. We report the task, serving setup, held-out work, limits, and the number of additional decisions the changed system handled correctly. The bad run stays in the record too.