LegalReasonBench (micro1): Sample Release

LegalReasonBench measures how well AI agents carry a legal matter through a sequence of assignments the way a practicing lawyer would: reading a case file, researching the law, applying it to the facts, and producing court filings and memoranda, where each assignment builds on the work before it. Every deliverable is graded against attorney-authored rubrics that reward required legal points and penalize specific errors.

Published report: https://www.micro1.ai/benchmark/legal-reason-bench

Citation

Please cite as: "LegalReasonBench (micro1). Available at: https://www.micro1.ai/benchmark/legal-reason-bench"

Scope of this release

This is a sample of 5 matters (10% of LegalReasonBench), released for evaluation and benchmarking. The remaining matters are held out. To request access to the full dataset, contact maria@micro1.ai or fill out the enquiry form at https://www.micro1.ai/.

Matter	Practice area	Assignments	Criteria (penalties)	Research
lrb-certiorari-quasi-judicial-land-use · Petition for Writ of Certiorari for Quasi-judicial Proceeding	Real estate and land use	2	37 (5)	open
lrb-dcca-challenge-police-discipline · Viability of a DCCA Challenge	Employment and labor	5	54 (11)	closed
lrb-interim-cmo-appointment-governance · Interim CMO Appointment: Governance and Compliance Analysis	Corporate governance	4	42 (8)	closed
lrb-vicar-trial-jury-instructions · OTB Gang VICAR Trial Jury Instructions	Criminal and investigations	2	35 (7)	closed
lrb-forum-non-conveniens-response · Plaintiff's Response, Forum Non Conveniens	Litigation and dispute resolution	2	53 (5)	closed

15 assignments, 221 criteria, 40 attachments (PDF and DOCX).

Research universe. In an open-universe matter the agent may research outside authority on the web. In a closed-universe matter the prompt restricts it to the supplied case file.

Contents

A tasks/ tree of 5 runnable Harbor multi-step tasks, one per matter, plus a Parquet index for the Dataset Viewer.

tasks/lrb-dcca-challenge-police-discipline/
├── task.toml                     # metadata, timeouts, one [[steps]] entry per assignment
├── environment/Dockerfile        # sandbox: Python, pandoc, poppler-utils, LibreOffice
└── steps/
    ├── assignment-1/
    │   ├── instruction.md        # role prompt + the assignment, shown to the agent
    │   ├── workdir/attachments/  # case-file documents for this assignment (copied to /app/attachments/)
    │   ├── tests/                # verifier, NOT visible to the agent
    │   │   ├── fixtures/rubric.json      # the weighted criteria for this assignment
    │   │   ├── raw_correctness/*.toml    # one judge criterion per rubric item
    │   │   ├── m1_judge_panel.json       # the three-judge panel
    │   │   └── test.sh                   # verifier entrypoint
    │   └── solution/solve.sh     # no oracle solution ships
    ├── assignment-2/
    └── ...

Assignments run in order in one persistent /app sandbox. Each assignment writes its deliverable to /app/response<N>.md, and later assignments can read earlier deliverables and attachments.

Dataset Viewer columns

The assignments/test split has one row per assignment.

• matter_id, matter_title, practice_area, research_universe
• assignment, num_assignments, assignment_title
• prompt: the full assignment prompt (the instruction file adds the shared role prompt)
• attachments: paths of the documents introduced in this assignment
• rubric_count, penalty_count, positive_points
• rubric: JSON list of criteria with criterion_id, points (negative for penalties), polarity, irac_section (Issue, Rule, Application, Conclusion), criterion, source (the authority or record cite the criterion rests on) and justification
• instruction_path, rubric_path, source_task_path: paths into the Harbor bundle

Scoring

• Criteria. Each criterion is judged pass or fail. A required criterion earns its points when met. A penalty criterion subtracts its points when the deliverable commits the described error.
• Assignment reward. Signed points earned over available positive points, clamped to [0, 1].
• Matter reward. Points-weighted across the matter's assignments: total signed points earned over total positive points, clamped to [0, 1]. scripts/score_matter.py computes it from a finished Harbor trial. Harbor's built-in multi_step_reward_strategy = "mean" is diagnostic only.
• Judges. A strict majority of three LLM judges decides every criterion: GPT-5.6 Terra, Claude Sonnet 5 and Gemini 3.8 Flash, each at medium reasoning effort. Each criterion is judged in its own call, and the judge reads only the deliverable.

Running

uv tool install harbor
export OPENAI_API_KEY=... ANTHROPIC_API_KEY=... GEMINI_API_KEY=...   # the judge panel
harbor run -p tasks/lrb-dcca-challenge-police-discipline -a opencode -m <provider/model>
harbor run -p tasks -a opencode -m <provider/model>                    # all five matters
python scripts/score_matter.py <trial directory>

All three judge keys are required. The panel fails the assignment if any judge cannot be reached, rather than scoring with fewer votes. The published results ran the OpenCode agent with a three-hour limit per assignment.

Important notes

• Sources. The case files combine public legal authority (court opinions, statutes, regulations and pattern jury instructions), public court records reproduced as filed, and matter documents that legal experts wrote or adapted for the benchmark. Some matters are built on real, publicly filed cases, so public records can name real parties.
• No professional use. This dataset does not provide legal advice and should not be used for production decision-making without independent validation.
• No endorsement. Use of this dataset does not imply endorsement by micro1.

License

CC-BY-4.0.