Legal reasoning is structural.
Vector retrieval finds passages that look similar. But the questions litigators actually ask — which precedents bind, which decisions overrule, which arguments win — live in the structure of legal reasoning itself, not in the text of any single document.
The problem
Brazilian litigators work across disconnected silos. Their own document corpus lives in one system. Case law sits behind half a dozen tribunal interfaces. Legislation is consulted in a browser tab, detached from the case context. A general-purpose LLM, if it is in the loop at all, has no memory of the matter, no structural sense of the corpus, and no audit trail a reviewer could inspect.
The practical result is a workflow that is slow where it should be fast and loose where it should be rigorous. A litigator spends hours reconstructing context that the system already contains, then ships an argument whose provenance nobody — not even the author — can fully trace.
This is the workflow we are replacing. Not the lawyer, not the case strategy, not the judgment call about which argument to run — the plumbing that today forces every one of those to be done from scratch, under time pressure, with no auditable memory.
What Silo bets on
The first bet is that legal reasoning is structural at every layer that matters. Inside one case, the structure is what separates a holding from an argument and a fact from a procedural claim. Across the body of case law, the structure is what binds, what overrules, what has been distinguished. Across legislation and doctrine, the structure is what grounds the whole thing. Vector retrieval cannot see any of those layers — it only finds passages that look similar to each other. Legal reasoning is the opposite: it is the graph, not the paragraph.
The second bet is that traceability is non-negotiable. A legal answer that cannot be audited is not a legal answer — it is a guess wearing a citation. Every output from Silo has to point back to what was retrieved, what was derived, and what was assumed, in a form a reviewer can follow end to end. Fluency without provenance is worse than silence.
The third bet is that the defensible asset is the pipeline that runs end to end: extracting structure from raw legal documents, maintaining it in a graph a litigator's questions can be asked against, and walking that graph with an autonomous agent that knows what a case analysis is supposed to contain. General search engines can index PDFs. Frontier models can summarize them. Neither can extract the claim-by-claim structure of a petition, resolve a citation across six tribunal syntax variants, and return an auditable case analysis — because each of those requires a practitioner-built taxonomy at a different layer. The moat is the whole stack, and the fact that the taxonomy at each layer was built by someone who practices the law.
Why this bet beats vector retrieval
Consider a concrete question: "Has this line of reasoning been overruled by a later decision?" Vector retrieval can return cases that look similar to the ruling in question — decisions about the same topic, with overlapping vocabulary. What it cannot do is tell you whether any of those later decisions actually overruled the line. Overruling is not a text property; it is a structural relationship between two rulings, expressed in how the second cites and treats the first. A system that does not model citations as edges cannot answer the question. A system that does model them as edges can — and it can show its work by handing the reviewer the exact citation chain that establishes the overrule.
This example generalizes. "Which precedents bind this court?" requires the court hierarchy. "How does this panel tend to rule on this doctrine?" requires decisions grouped by panel composition. "Are there contradictory rulings in the same court?" requires comparing holdings, not passages. Every one of those questions is a structural query. Vector retrieval, at best, returns plausible reading material. It cannot answer.
Where to go next
For the architecture and data flow that delivers on this bet, continue to 02 · System. For the live numbers and a recorded trace, jump to 03 · Proof.