Supreme MindSupreme Mind
1 · Before filingGatekeeper PilotKnow the expert risk before you take the case.2 · Discovery to settlementExpert Agents LiveEach side’s expert analysis and the settlement value, rerun as the facts change.3 · Expert stageReport Builder PilotYour own expert’s report, faster and harder to exclude.
One record, from intake to settlement.Use it in Claude, Copilot or Word, freePilotsAPI for your tools
AntitrustSecuritiesCommercialMass TortPersonal Injury & Med MalAll Practice AreasPlaintiff FirmsDefense TeamsLitigation Funders
Expected GainsBenchmarkSample BriefResearchSecurity & Trust
Expert Agent LibraryRule 702 TrackerGuidesHow It WorksFAQ
PricingConnect
Sign inBook a demo

How Expert Agents work

One agent run covers one class of expert on one fact pattern and returns a brief in four sections. This page sets out what goes in, what comes out, how it is checked, and where the limits are. To see the output itself, read the sample briefs: three real matters, each run for both sides’ experts, complete and free to read without an account.

What a run reads

Your matter. The public case materials you already have, such as the complaint, the opposing expert’s report and key filings, plus the basic facts of the case. Supreme Mind extracts the facts from them: parties, class period, disclosure dates, the issues in dispute. You decide what to upload, and you are never required to upload anything beyond a caption and a docket number. Material produced under a protective order waits for a signed agreement; see below.

Real data, not only text. The analysis is grounded in data the expert would use: an event study on actual market prices around the disclosure dates for securities matters, a cost pass-through analysis on public FRED input-cost series for antitrust, and Daubert and prior-testimony records from public court filings. In a securities matter, that is an OLS market-model event study on the issuer’s prices against the S&P 500, SEC EDGAR filings, and the judicial record on the methodology from CourtListener.

Before and after an agreement is signed

Supreme Mind works on public material from the first run. Material produced in discovery under a protective order is different: the order decides who may see it, so it waits until the paperwork the order asks for is signed.

  • Without a signed agreement (the example cases, on a firm email): public material only. The complaint, public filings and docket entries, expert reports and transcripts already on the public record, and the market and court data a run pulls itself. Nothing designated under a protective order, nothing filed under seal, and nothing else a court order keeps from outside vendors.
  • With a signed agreement (a firm plan or a pilot): a confidentiality agreement and data processing addendum with your firm, and, where the protective order asks it of litigation-support vendors, the order’s own acknowledgment and agreement to be bound, signed before anything designated under it is uploaded. Material designated under the order can then be used on that matter, at the confidentiality tiers the order lets a vendor see. The order sets those limits, not us, and counsel confirms them before upload.

The handling itself does not change between the two: the same US-region processing, no retention by the model provider, no training on your content, and deletion when you delete the matter. The agreement records those commitments in writing and adds what the order requires. See Security & trust and Pricing.

What a run returns

  • Likely opinion: what the expert will conclude, in their own voice, with the methodological basis.
  • Cross-examination weaknesses: ranked, with the question to ask, the deflection to expect, and citations to the case file.
  • Methodology challenges: the Daubert and Rule 702 attack surface, with the authorities looked up (see how each circuit applies the amended rule in the Rule 702 tracker).
  • Settlement range: how the testimony moves it, directional, with the uncertainty stated.

The same matter can then be put to work: rehearse the cross-examination live against the expert agent, stress-test a report before it is served, or ask the case file a question and get an answer cited to the page.

How accurate it is

We gave the same model the same instructions on real expert reports, predicted blind, and changed only what it was allowed to know. The score is the share of the expert’s key findings it reproduced.

General-purpose model on its own28%
+ case data and code execution62%
+ expert-class profile70%
+ individual expert dossier74%
+ dated news timeline78%

+50 points over the model alone80% interval: +37 to +60 points

Every layer of expert knowledge adds accuracy, and together they nearly triple it. An internal blind test, October 2026; what each layer adds is on the Benchmark page.

How every quote is checked

Every cited authority is located by name in public case-law databases at generation time. Every supporting quotation is checked verbatim against the public source, and a quote that cannot be verified is dropped, not guessed. The passage a run relied on can be read in place, so a quotation or citation traces back to a source a reader can open. Figures, calculations and the analysis itself are not checked this way; counsel should verify them before relying on them. The quote and citation check is the part you can verify yourself today, on your own matter; accuracy on the analysis itself is measured on the benchmark.

What a class of expert is

Each expert is a de-identified composite representing a class of expert witness, for example a defense event-study econometrician or a cartel-overcharge damages economist. It is anchored to publicly documented methods and prior testimony and bounded to the public record. It is not any specific named individual, it is not affiliated with, sponsored by or endorsed by any real expert or firm, and it never claims to predict what a named individual will say in a specific deposition.

Where your data goes

Supreme Mind runs on the leading general-purpose models, chosen for legal work on measured accuracy, under zero data retention with inference pinned to the US, so prompts and completions are not stored and are not used to train any model. Matter content is isolated per firm in US-region storage and is never used to train any model, ours included. The detail is on Security & trust; the questions firms ask most are in the FAQ.

See it on your own matter.Start freeRead a sample brief →

Questions about method? Email richard@suprememind.ai.

Gatekeeper→Expert Agents→Report BuilderOne record, from intake to settlement
Book a demo
Supreme MindSupreme Mind
AI expert witness agents for litigation, from the first screen to settlement
Products
  • Gatekeeper Pilot
  • Expert Agents
  • Report Builder Pilot
  • Pricing
  • Pilots
  • API
  • Use in Claude, Copilot or Word
  • Book a Demo
Solutions
  • Antitrust
  • Securities
  • Commercial Litigation
  • Mass Tort
  • Personal Injury & Med Mal
  • All Practice Areas
  • For Funders
Why Supreme Mind
  • Expected Gains
  • Benchmark
  • Sample Brief
  • Research
  • Security & Trust
Resources
  • Expert Agent Library
  • Rule 702 Tracker
  • Guides
  • How It Works
  • FAQ
Company
  • About
  • Why Now
  • The Fourth Institution
  • Contact
© 2026 Supreme Mind AI, Inc. All rights reserved.
Terms of ServicePrivacy PolicyDPA