Design-partner pilot · litigation firms · enquiries open Q4 2026 · pilot Q1 2027
MatterGraph ties every proposition in a research map to an exact, page-referenced passage in a court-published authority, freezes the whole map under a hash, and records which lawyer reviewed it against the sources. It refuses to state a current-law conclusion it cannot evidence. That refusal is the product.
01 · The product
A research map is not an answer. It is a set of propositions, each with the passages that were retrieved for it, the status the evidence earned, and the warnings the system could not clear. The lawyer reads it against the sources and decides.
Retrieval returns the literal text of the source, split at the document's own structure, with page range and judgment paragraph numbers where the judgment supplies them. The model never rewrites a passage; the stored text is what the lawyer verifies against the PDF.
A proposition is unsupported until a retrieved authorised passage independently supports it. Authorities carry a currency status that starts as unknown and changes only when a named lawyer certifies it, with the date of the check shown on every review surface.
If controlling-authority recall is unverified, if any cited authority's currency is unknown, or if any proposition is unsupported, the snapshot carries current_law_claim_blocked. A lawyer must clear each warning by review; the software cannot.
The review workspace is a separate, human-only surface. Approving requires confirming that every cited source was opened, a written note, and the exact snapshot hash. Rejection records the corrections required against the same hash. Chat cannot sign.
Relations between authorities (follows, distinguishes, overrules) enter the graph only when a lawyer verifies them against a specific supporting passage. Unverified treatment is never shown as fact.
02 · Boundaries
A research tool for litigators is judged by the claims it refuses to make. These are enforced in software and tested, not stated in a disclaimer.
03 · Where it runs
The privilege boundary is the firm's own deployment, not a row filter in a shared database. What leaves that boundary is narrow, approved, and logged.
Documents, passages, research maps, approvals, deletion receipts, and the hash-linked audit log live in the firm's own database. They work with the network cable pulled.
Document parsing and search indexes are derived from content the firm holds and can be rebuilt. They run on the firm's hardware, or on a node inside the firm's own cloud account that holds nothing persistent.
Only the specific passages retrieved for a question are sent to a model, and only to a provider the firm has approved in writing with zero customer-content retention, bound to an exact model name. Cross-border processing is the firm's information officer's decision, recorded, never a default.
04 · The pilot
One practice group. One recurring research task. A bounded corpus of court-published authorities. A lawyer-labelled benchmark that the build must pass before it is called evidence-grounded. Enquiries and preliminary onboarding run in the last quarter of 2026; the pilot itself runs in the first quarter of 2027.
Together we choose the practice area and the ordered court hierarchy. Every authority in the pack must belong to it; the research-scope check fails otherwise.
Operator agreement under POPIA, provider approval, the authorised pack with its source-rights record, and the evaluation plan. One manifest binds them; the software checks it before ingest and again before release.
A lawyer in the practice group marks the controlling authorities and accepted passages for a defined research task, anchored to paragraph numbers so the labels outlive any re-indexing. The build runs blind against them.
The same lawyer completes the task with the firm's current method and with MatterGraph, reviews the snapshots, and records which they prefer and why. Preference is recorded by the lawyer, never inferred from clicks.
Passing them is not a claim of production fitness or legal correctness. A South African lawyer supplies the labels and remains responsible for the legal work.
| Gate | Required | Measured against |
|---|---|---|
| Citation correctness | At least 95% | Cited document, neutral citation, court, and date match the lawyer's label |
| Passage support | At least 90% | The cited passage is in the lawyer's accepted set for that proposition |
| Issue and proposition coverage | 100% | Every labelled issue and proposition appears in the map |
| Silent current-law claims | Zero | No snapshot permits a current-law claim while any evidence is incomplete |
| Snapshot reproducibility | 100% | Recomputing the hash over the stored map reproduces it |
| Controlling-authority recall | Reported | Becomes a gate only when the design partner sets the target |
05 · Fit
Boutique and mid-sized South African litigation practices. A design partner shapes what gets built next, and is asked for honest, specific criticism in return.
06 · Who we are
Platform engineer, Johannesburg. Through ReactorJet, Tim builds cloud-native systems where scale matters and local-first systems where confidentiality has to be enforced by the architecture rather than promised in a policy. MatterGraph applies that discipline to South African privacy and professional-conduct obligations, an area with no technical standard yet, by making evidence, review, and deletion mechanical and holding the implementation to the strictest reading of the law.
A trading name of the founder, 35 1st Avenue, Illovo, Sandton, 2196, with salegalintelligence Pty Ltd being incorporated to carry the business. MatterGraph is the product. The core is proprietary; the pinned open-source components it deploys beside are listed in the deployment notes given to every design partner.
Decisions about the product, including what it refuses to do, are written down with their reasoning and shared with design partners. If we change one, you see the change and the date.
07 · Start
Tell us the research task you would most want to trust. The first conversation is to work out whether there is a bounded pilot worth running, and to give you the operator agreement, the pilot handoff kit, and the threat model to read before deciding.
A named person replies within two working days. No newsletter, no sequence.
Professional context. The Legal Practice Council's practical vocational training guide on information and communication technology for practice (issued under Regulation 6(10)(n), March 2026) tells candidate attorneys that "AI and AI tools are having a disruptive effect on the legal profession" and that they are expected to use AI in the interests of the client and the firm. On research it is direct: a lawyer must still apply their mind to the facts and the law and reach a reasoned solution, and "the current AI models cannot do that." It records two 2025 judgments in which cited cases did not exist and the practitioners were held negligent, and it instructs that a citation which cannot be found in the usual places is a hallucination and must not be used. MatterGraph is built for exactly that duty: nothing it shows is more than a passage a lawyer can open and check.
On the claims on this page. Everything described in the present tense is built, tested, and can be shown running on synthetic data. Anything gated is named with its gate. We publish the rules we hold our own copy to; ask for them.