Skip to content

Matterframe

The unified model of legal superintelligence.

One model of how legal work is read, owed, and done, beneath every practice and every firm. It reads like a solicitor and counts like a machine.

A model that reads the whole file, not what someone pasted in.

A chat session sees what a person gives it. Matterframe's extraction pipeline turns every kind of document into clean text a model can read, with its pages, and its agents work through the entire file: searching it, opening what matters, cross-checking figures, and knowing exactly what they have and have not read.

Reading pipelineHL-0131, Estate of Evelyn Marsh214 of 214 read
The file
Letter from executors.pdf6 pages
Re valuation.msgthread of 4
Witness statement.docx14 pages
Bank statements 2019.pdfscan, 212 pages
WhatsApp Chat with Clara.txt1,840 messages
Estate accounts.xlsx3 sheets
Voicemail from client.m4a0:48
Extraction built for models
Text layer
Email thread, split
OCR, page by page
Chat export, dated
Sheets as tables
Audio, transcribed with timings
One text per document, with its pages
An agent reads the whole file
Searched the file for every mention of the house
Opened the valuation, the accounts, and 3 letters
Cross-checked the figures against the bank statements
Coverage: 214 of 214 read, 0 unreadable

Scans

Every page read, including pages with no text layer.

Recordings

Voice notes, calls, and CCTV, transcribed locally with a time on every line. Faint audio is marked.

Chats and threads

WhatsApp exports and email chains, split into dated messages.

Reads like a solicitor.Counts like a machine.

The reading is done by the strongest models available. Every date, deadline, and figure is then worked out in code against the rules, so each answer can be checked and the same file always gives the same count.

p. 2 of 6
DATE OF GRANTGrant of probate issued 14 May 2026Sealed grant, page 2, line 9

Every page, first hand

Letters, emails, orders, scans, bank statements, medical records, chats, and recordings. Every fact carries the page or time it came from.

Claim must issueInheritance Act 1975, s.414 Nov
Witness statements dueOrder para 5, counted under CPR 2.816 Oct
Reply awaitedOur letter of 14 August46 days
Client update owedFirm standard: every 28 days7 Oct

Every clock, in one register

Limitation, court directions, letters awaiting a reply, and rules of court, counted to the day and each tied to its source.

Was a standstill agreed? Unknown.2 of 214 documents could not be read. Absence is not confirmed until they are.

Honest about what it does not know

Coverage is the product. Where a page is missing or unreadable, the answer is unknown and the gap is named. False comfort is worse than none.

1870Banks v Goodfellow (1870) LR 5 QB 549Checked
2017Ilott v The Blue Cross [2017] UKSC 17Checked
2026Croll v Estate of Thorpe [2026] EWHC 2232 (Ch)Checked
Each authority is quoted from the primary source and checked before it reaches a draft. Nothing is cited from memory.

Grounded in the law

Reasoning is anchored to curated statutes, practice directions, and case law. New judgments are read and curated in as they are handed down.

Citations you can stand behind.

Invented cases and misattributed quotations have reached court filings, and lately even judgments. Caseforth cites only from its own register of authorities, each checked against the primary source. A citation that is not in the register never reaches a draft.

405authorities, each with its citation and holding
386linked to their primary source, not a summary of it
309statutory provisions, from the legislation itself
75pieces of official guidance, from practice directions to regulators' handbooks

A register, not a memory

Models are never trusted to recall a case. They choose from the register, and the register says what the case decided.

Quoted from the source

Every quotation is matched word for word against the stored judgment or Act before it can be used.

Kept current

New judgments are read and curated in as they are handed down, and a change in the law reaches every pack that relies on it.

Runs on the subscriptions you already have.

The heavy reading runs through the flat-rate plans a firm's admin users already hold, not a metered API bill that grows with every page. Where a firm needs it, Caseforth also runs on local models and other compatible ones.

Recommended

Claude Max 20x

For admin users running the agents. Reading, drafting, and review at a fixed monthly cost.

Recommended

ChatGPT Pro

The second lab. Its models review the first's calls, so every decision gets a second opinion from a different mind.

Optional

Local and compatible models

Local deployments, and other OpenAI-compatible and Claude-compatible models, for firms that need them.

One record of every matter, built from the file.

01

The reading

One text for every document, with its source and how it was read. Nothing is truncated on its way to a model.

02

The register

Every obligation on the matter in one place, with the words that created it and the rule that counts it.

03

The position

What is true on the matter now, what is owed next, and why, rebuilt as the file changes.

04

The record

Who read what, which agent did what, and what a person decided. Supervision and audit read the same record.

Read the whole book.Move every case forward.

Caseforth is taking on a small number of firms. If yours could be one, we would like to talk.