Wikilinks are the product

“Wikilinks are the product” means that in a research graph, a [[link]] is a claim about a relationship between two things, and the web of those claims is the real output. When a note links to another note, it is asserting that these two entities are connected and that the reader can follow the connection to see why. The prose inside each note explains and grades the claim, but the structure, the links themselves, is the thing you navigate and the thing you audit. An unresolved link is not a broken reference to clean up. It is the machine pointing at something it has named but not yet investigated, which makes it the next research target.

This is the inversion at the centre of how I build research graphs. It is part of the method, and it is the difference between a document and a map.

Why most tools treat prose as the product

Almost every research tool, AI or otherwise, treats the written answer as the deliverable and any structure as a by-product. You ask a question, you get paragraphs. If the tool draws a diagram or extracts entities, it does that after the fact, by parsing the prose it already wrote. The links are a visualisation of the text, so they inherit whatever the text got wrong, and they cannot tell you anything the text did not already say.

That ordering has a cost you only notice later. The prose reads as one continuous claim. You cannot point at a single sentence and ask “is this part verified, and what does it connect to,” because the connections were never first-class. Six months on you are re-reading paragraphs to reconstruct what related to what. The relationships, the part you actually wanted, were never stored as data. They were implied by sentences and then thrown away.

Why I invert it

I build the graph as the primary artefact. Each entity is a note. Each relationship is an explicit [[link]] from one note to another, written down on purpose because a source supports it. The prose exists to explain and grade each link, not the other way around. So when you open a vault-builder graph you are not reading a report that happens to have a sidebar diagram. You are navigating a structure where every edge is a deliberate assertion you can trace back to a source.

This changes what the output can do. You can ask the graph questions the prose never could: which entities are most connected, which clusters form, which central node is still thin. You can hand it to a colleague who navigates the same structure instead of re-reading your paragraphs. You can audit it edge by edge, because each edge is a discrete claim rather than a clause buried in a sentence.

An unresolved link is the next research target

Here is the mechanism that falls out of treating links as claims. When the machine writes “this person founded that company” and links to a company note that does not exist yet, it has not made an error. It has made a claim and admitted it has not researched the other side of it. That dangling link is a named gap. It is the graph saying “I know this entity matters, I just have not looked at it yet.”

So unresolved links drive the research loop instead of breaking it. The next round picks up the most important dangling links and investigates them, which either creates the missing note with its own evidence grade or surfaces a conflict worth recording. A broken link in a normal document is a defect. In this method it is the to-do list, generated by the research itself rather than by me guessing what to look at next.

This is also why the core invariant is strict: no phantom nodes, and Title-Case filenames so links resolve predictably. A phantom node is a link that points nowhere and will never be picked up, a claim with no home. By the time a graph is finished, every link either resolves to a real, evidence-graded note or has been deliberately closed out. The final audit re-reads every note against its sources and deletes orphaned and phantom files, so what survives is structure that resolves. That requirement is in the code, not in a style guide I hope to follow.

The graph of claims is what you navigate and audit

Because every link is a claim and every claim carries an evidence grade, the structure becomes auditable in a way prose never is. You can walk the graph and at each edge ask: what is this connection, what grade is it, what source backs it. You are not trusting a confident paragraph. You are inspecting a web of individual, traceable assertions.

The numbers from a shipped build show what this looks like at scale. SignalTrace publishes thirteen research wikis built from vault-builder output. Of its markdown files, 866 carry wikilinks, and more than 780 are entity pages that each carry an evidence_strength grade in frontmatter. That is not 866 documents with decorative cross-references. It is a network of more than 780 graded claims about real entities, connected by links that each had to resolve before the graph was called done. The web of links is the product. The pages are where each claim is explained and sourced.

Where this shows up

This is one of the quality gates inside the method, alongside evidence grading and convergence. The engine is the knowledge graph vault builder, and you can see the published result in SignalTrace.

If you want a research surface where the relationships are first-class and auditable rather than buried in prose, read the method or email me at info@devai.co.za and tell me what you are trying to map.