Sakana Marlin for Legal Research
Deep, long-form research is where an autonomous research agent fits legal work best — with citation verification that is not optional.
Sakana Marlin is an autonomous research agent from Sakana AI in Tokyo. You submit a research topic in plain language, it runs on its own for up to roughly eight hours, and it returns a finished strategic report — an executive summary, strategic options, citations grounded in primary sources, and presentation slides. It is not a chatbot and not a conversational model.
Of the legal use cases people ask about, deep legal research is Marlin's best fit. Broad precedent surveys, regulatory landscape mapping, and multi-jurisdiction questions are exactly the kind of large, well-scoped research problems its long autonomous runtime and AB-MCTS search process are built for.
This page covers how to use Marlin for legal research, what its output is good for, and the citation-verification and confidentiality obligations that come with it. It also draws a firm line: Marlin produces background research, not an authority check. Confirming that authorities are real, current, and good law still happens on Westlaw or Lexis.
Why research is the best fit
Marlin's design lines up with deep research more than any other legal task. It takes a topic and runs autonomously for hours, using AB-MCTS — the Adaptive Branching Monte Carlo Tree Search method from a NeurIPS 2025 Spotlight paper — to explore many research paths and prune the unproductive ones. The result is a structured report rather than a single answer.
That shape suits questions you would otherwise spend a day or more assembling by hand: how a doctrine has developed across circuits, what a regulatory regime looks like across several jurisdictions, or how a policy area is trending. You trade immediacy for breadth and structure.
- Deep precedent research across a doctrine or issue.
- Regulatory landscape surveys within or across jurisdictions.
- Long-form research memos produced over hours.
- Multi-jurisdiction comparative research.
Using AI for legal research without a verification workflow is a risk. Layer3 Labs helps regulated firms build one that holds up.
Book a ConsultationHow to use it for a research question
Scope the question before you submit it. Because a run can take up to eight hours, a vague prompt is an expensive way to discover you asked the wrong thing. State the jurisdiction, the time frame, the specific issue, and the form of output you want.
Treat the returned report as a first draft of your research, not the finished product. The executive summary and strategic options give you structure; the citations give you leads. Your job is to follow those leads to the primary sources and confirm them.
Background research, not an authority check
Marlin grounds citations in primary sources, but grounding is not the same as verification. It does not replace running a case through a citator to confirm it is still good law, nor does it replace confirming a statute or regulation is current. Those steps belong on Westlaw, Lexis, or the official primary source.
Use Marlin to find the landscape and surface candidate authorities quickly. Use your research platform to confirm each authority exists, is current, and stands for what the report claims.
Confidentiality and privilege
Research questions are often the safest AI inputs precisely because they can be framed without client-identifying facts. A question about how a regulation applies to a category of business rarely needs the client's name or matter details. Prefer that generic framing under ABA Model Rule 1.6.
If a research question genuinely requires confidential facts, review the vendor's data-handling terms first and consider informed client consent. Do not paste privileged material into any third-party service without confirming how it will be stored and used.
Mandatory citation verification
This is the non-negotiable part. Marlin can produce confident, well-formatted output that is wrong — including citations that misstate a holding or that do not exist at all. Courts have sanctioned lawyers for filing briefs with AI-hallucinated cases, most notably in Mata v. Avianca, where fabricated citations led to sanctions.
Before any Marlin research reaches a filing, a memo to a client, or advice, a lawyer must independently verify every citation and material assertion against the primary source. Under ABA Model Rules 5.1 and 5.3, the supervising lawyer is responsible for that verification regardless of how the draft was produced. Marlin is not a lawyer and gives no legal advice.
Frequently Asked Questions
- It is Marlin's best legal fit. Its autonomous, hours-long runtime suits deep precedent surveys, regulatory landscape mapping, and multi-jurisdiction research. It returns a structured report with primary-source citations — which you must then verify.
- No. Marlin produces background research and surfaces candidate authorities. Confirming a case is still good law or that a statute is current belongs on Westlaw, Lexis, or the official primary source.
- Up to roughly eight hours per task. It runs autonomously and returns a finished deliverable. Because of that runtime, it is built for large research questions, not quick lookups.
- Because Marlin can produce confident but false output, including citations that do not exist. Lawyers have been sanctioned for filing AI-hallucinated cases, as in Mata v. Avianca. Verify every citation against the primary source before relying on it.
- Adaptive Branching Monte Carlo Tree Search, the method behind Marlin, described in a NeurIPS 2025 Spotlight paper. It orchestrates multiple AI models to systematically explore and prune research paths during a run.
Put deep legal research on a safer footing
Book a free 30-minute AI workflow audit with Layer3 Labs. We will help you scope research prompts, set citation-verification checkpoints, and keep confidential facts out of third-party tools.
/ai-workflow-audit