Grounds

AI whose reasons can be checked

We study how neural networks work inside, and we build a private AI workspace for law firms, insurance brokers and agents, and other businesses, set up around the documents each one already writes. When the AI cites a case or one of your documents, the citation opens the source at the words it relies on.

Discuss your documents Why we study model internals

Please read the policy carefully. The flood insurance provided is subject tolimitations, restrictions, and exclusions.I. AgreementA. This policy insures the following types of property only:1. A one to four family residential building, not under a condominium form ofownership;2. A single-family dwelling unit in a condominium building; and3. Personal property in a building.(paragraph I.B left out)C. We will pay you for direct physical loss by or from flood to your insuredproperty if you:1. Have paid the full amount due (including applicable premiums, surcharges,and fees);2. Comply with all terms and conditions of this policy; and3. Have furnished accurate information and statements.D. We have the right to review the information you give us at any time andrevise your policy based on our review.E. This policy insures only one building. If you own more than one building,coverage will apply to the single building specifically described in the FloodInsurance Application.(paragraphs I.F and I.G left out)II. DefinitionsA. In this policy, “you” and “your” refer to the named insured(s) shown on theDeclarations Page of this policy and the spouse of the named insured, if aresident of the same household. Insured(s) also includes: Any mortgagee andloss payee named in the Application and Declarations Page, as well as any othermortgagee or loss payee determined to exist at the time of loss, in the order ofprecedence. “We,” “us,” and “our” refer to the insurer.(one sentence left out)B. Flood, as used in this flood insurance policy, means:1. A general and temporary condition of partial or complete inundation of twoor more acres of normally dry land area or of two or more properties (one ofwhich is your property) from:a. Overflow of inland or tidal waters;b. Unusual and rapid accumulation or runoff of surface waters from any source;c. Mudflow.
C. We will pay you for direct physical loss by orfrom flood to your insured property if you:1. Have paid the full amount due (includingapplicable premiums, surcharges, and fees);2. Comply with all terms and conditions of thispolicy; and3. Have furnished accurate information andstatements.D. We have the right to review the informationyou give us at any time and revise your policybased on our review.E. This policy insures only one building. If youown more than one building, coverage willapply to the single building specificallydescribed in the Flood Insurance Application.(paragraphs I.F and I.G left out)II. DefinitionsA. In this policy, “you” and “your” refer to thenamed insured(s) shown on the DeclarationsPage of this policy and the spouse of thenamed insured, if a resident of the samehousehold. Insured(s) also includes: Anymortgagee and loss payee named in theApplication and Declarations Page, as well asany other mortgagee or loss payee determinedto exist at the time of loss, in the order ofprecedence. “We,” “us,” and “our” refer to theinsurer.

the policy's definitions GPT-2 small, layer 0, head 5

Point at a word with a blue line to follow it.Tap a label on the right to follow one word. “You” and “your” mean the named insured shown on the Declarations Page, and the named insured's spouse if they live in the same household (II.A). Used 7 times above, 6 of them before the definition. This head looks back to the first “you”, in I.C.“You” and “your” mean the named insured shown on the Declarations Page, and the named insured's spouse if they live in the same household (II.A). Used 6 times above, all of them before the definition. This head looks back to the first “you”, in I.C.“We,” “us” and “our” mean the insurer (II.A). Used 4 times above, all of them before the definition.Flood is defined in II.B, starting with “a general and temporary condition of partial or complete inundation”. Used once above as the defined word, before the definition.Building is defined further on, in II.C.6. Its lines leave the bottom of this excerpt.

A page of a flood insurance policy. Blue lines join uses of you, we and flood to their definitions further down the page; 11 of 12 uses come before the definition. Dark lines show where an attention head in GPT-2 small looked back from the same words, mostly to the first time the word appeared.

The opening of the federal flood insurance policy, with three passages left out. Blue lines join each use of you, we and flood to where the policy defines the word: 11 of the 12 uses come before the definition. Dark lines show where one attention head in GPT-2 small looked back as it processed the same words. It mostly returns to the first time it saw the word. Part of the opening of the federal flood insurance policy, from I.C to II.A, with paragraphs I.F and I.G left out. Blue lines join each use of you, we and flood to where the policy defines the word: all 11 uses shown come before the definition. Dark lines show where one attention head in GPT-2 small looked back as it processed the same words. It mostly returns to the first time it saw the word.

How this figure was made

The text is the Standard Flood Insurance Policy, Dwelling Form (44 CFR Part 61, Appendix A(1)), a public federal form, retrieved from the eCFR on 2 October 2026. Paragraphs I.B, I.F and I.G, and one sentence before II.B, are left out to fit the page; each cut is marked with an ellipsis. GPT-2 small (124 million parameters, released by OpenAI in 2019) processed the excerpt once. We picked the attention head by a rule, as we did for the figure on our research page: of GPT-2 small's 144 heads, the one that most strongly attends to an earlier copy of the same token on random text (layer 0, head 5, score 0.64). For each use of a defined word, its dark line goes to the earlier position the head attended to most, when that weight is at least 0.10. The very first position is left out, because heads park unused attention there. Of the 12 uses drawn, the head went back to the word's first appearance 10 times and to another earlier use of the same word twice. It never went to the definition: a model can only look back, and 11 of these 12 uses come before their word's definition. The one use after its definition also went back to an earlier use of the same word. The blue lines come from matching the defined words in the text.

The text is the Standard Flood Insurance Policy, Dwelling Form (44 CFR Part 61, Appendix A(1)), a public federal form, retrieved from the eCFR on 2 October 2026. Paragraphs I.B, I.F and I.G, and one sentence before II.B, are left out to fit the page; each cut is marked with an ellipsis. The figure on this screen shows the part from I.C to II.A. GPT-2 small (124 million parameters, released by OpenAI in 2019) processed the excerpt once. We picked the attention head by a rule, as we did for the figure on our research page: of GPT-2 small's 144 heads, the one that most strongly attends to an earlier copy of the same token on random text (layer 0, head 5, score 0.64). For each use of a defined word, its dark line goes to the earlier position the head attended to most, when that weight is at least 0.10. The very first position is left out, because heads park unused attention there. Of the 9 uses drawn in this part, the head went back to the word's first appearance 8 times and to another earlier use of the same word once. It never went to the definition: a model can only look back, and all 9 of these uses come before their word's definition. The blue lines come from matching the defined words in the text.

Grounds for Law

Case law, eDiscovery and drafting, in the same matter. When the AI cites a case, the case opens beside your draft at the passage it quoted.

  • The AI searches our database of state and federal case law and statutes, and its citations are checked after each answer.
  • Load a production, review it, then quote its documents in chat or use them as sources for a brief.
  • Motions and discovery start on your firm's own pleading paper, and AI edits arrive as tracked changes.

More about Grounds for Law

A motion in the Grounds editor citing a case, and the cited California Supreme Court opinion open on the right with the quoted passage highlighted.A California Supreme Court opinion open in the Grounds side panel, with the quoted passage highlighted.
A motion in an invented insurance coverage case, with one citation opened beside it at the quoted passage. The opinion is public. The firm, the parties and the motion are made up.The cited opinion, opened at the passage a motion in an invented insurance coverage case quotes. The opinion is public. The firm, the parties and the motion are made up.

More about Grounds for Law

Grounds for Insurance

Brokers, agents and claims teams can map a policy and read it as an outline beside the PDF, with each defined term opening its definition where it's used and endorsements marked on the sections they change.

More about Grounds for Insurance

The definition of “insured contract” open in Grounds' policy map, with a line naming the endorsement that replaced its paragraph f.The definition of “insured contract” open in Grounds' policy map, with a line naming the endorsement that replaced its paragraph f.
A definition in an invented liability policy, opened from the text. An endorsement replaced one of its paragraphs, and the map names it.

More about Grounds for Insurance

Part of a kitchen scope of work in a remodeler's own template, with the grout color line highlighted as a question for the customer.
A kitchen scope for Northgate Remodeling, a made-up company. The grout color wasn't in the notes, so it's asked.

More about Grounds for Business

Grounds for Business

For companies that send the same kinds of documents over and over. We set up your templates and terms. Drafts come out in your formatting, with anything the notes didn't cover listed as a question.

More about Grounds for Business

Each workspace is built around the documents you already send out

I start a setup with the same questions: what does your team write again and again, which past examples are the good ones, and what does each document depend on? Then we load your templates, examples and reference material, build the workflows that draft from them, and use your words for things: matters and clients at a firm, accounts and clients at an agency, jobs and customers at a builder.

A coverage summary on an insurance agency's letterhead in the Grounds editor, with the AI's edits shown as word-level tracked changes and a colleague's named cursor at the title.A paragraph of a coverage summary in the Grounds editor, with the AI's edits shown as word-level tracked changes.
A coverage summary on an invented agency's letterhead, with the AI's edits as tracked changes and a colleague in the document. The agency and the client are invented.The opening of a coverage summary from an invented agency, with the AI's edits as word-level tracked changes. The agency and the client are invented.
  • Documents stay Word files, edited in the browser, with tracked changes, comments, and versions you can compare and restore.
  • Drafts start on your own letterhead, pleading paper or template.
  • The AI's edits arrive as tracked changes, word by word, under the name you choose.
  • When the AI cites one of your documents, the link opens it beside the answer at the passage it quoted.
  • Each matter, account or job holds its files, notes, people and dates, and chat works from that record.
  • A client portal is an add-on. Clients sign in with an emailed link, see what you share and send what you ask for, and don't take a seat.
  • In law and insurance workspaces, long documents can be mapped: an outline, the defined terms and the cross-references, beside the PDF. You choose which documents to map.

Models run on AWS Bedrock under zero-data-retention terms, so the model provider keeps nothing you send and doesn't train on it. Document text is encrypted with your company's own key.

How we handle data

We think neural networks can be understood completely

That is a long way off, and our research is early. It covers interpretability, training methods and continual learning, and we're working to apply interpretability tools to the documents we know best, in law and insurance, where a right reading and a wrong one can be told apart. The figure at the top of this page is one small example: one attention head of a small model, on one public document.

Interpretability as a Guiding Principle

Why understanding the internals of neural networks matters, why it is hard, and what Grounds is doing about it.

  • Grounds goes live

All research notes

Tell me what your team writes most, and what slows it down

I'm Joe Wilbert, the founder. I read every note and answer it myself. Workspaces are set up by invitation, after we've talked through your documents.

Already have an account? Sign in.