Main idea

A set of notes linked to one another holds contents and links. It holds nothing of what is needed to use them to move forward on a product opportunity.

Three questions show it, and none has an answer in the graph. Two notes contradict each other: which one holds today, and who settled it? An entire area of the subject is undocumented: how do you see that, since a gap produces no note? A trade-off between two options was settled three weeks ago: where is it written, who made it, and does it still hold?

These three objects — the inconsistency, the state of completeness, the decision and its justification — are not contents among others. They bear on the state of the corpus itself, and they therefore demand places of their own: dedicated files, kept up to date, distinct from the notes. A flat file carries them no better than a network does: what is missing is not the list form or the graph form, it is the layer that speaks about the knowledge rather than about the subject.

A product description meets the same three objects, and houses them the same way: a block of discrepancies per card, separate from the body, carrying the divergences between what the product does and what it should do; a coverage counter per enumeration angle, giving the state of completeness; and, for every object without a card, a written decision — business object, or plumbing. The rejected discrepancy stays there with its reason, and so does the decision. The layer that speaks about the knowledge is therefore not reserved for domains of thought: it takes the same place in a technical reference model.

Layer added by "I wrote a product's ontology. Three times, I thought I was done." (2026-08-11).

Layer added by "Stateful Corpus Research Loop: stop collecting sources" (2026-10-04). This layer isn't only for picking a case back up: in research carried out in iterations, it becomes the input to the next decision. What depends on a single source, what remains uncertain, terminological ambiguities, over-represented or almost absent areas each point to a type of evidence to go and find. A corpus that can describe its own state can steer its research; a corpus that only talks about its subject can only grow.

Why it matters

This separates two things we usually pile into the same tool: what we know, and what we know about what we know. The second layer is the one that lets you pick a case back up three weeks later without rereading everything, and it is the one no note-taking tool provides by default.

It also gives an evaluation criterion: faced with a knowledge setup, ask where the current contradictions, the uncovered areas and the decisions are written down. The absence of an answer is the diagnosis.

Nuances and limits

Keeping these files is work, and work that produces nothing visible. On a short subject, it costs more than it returns.

And a tracked contradiction is not a resolved contradiction: the setup makes visible what remains to be settled, it does not settle it.

Layer added by "Graphify without AI: turning a wiki into a navigable graph for zero tokens" (2026-10-20). A computation on the graph surfaces one kind of gap without anyone looking for it: the isolated note. On a 1,013-note wiki, the report from a graph tool flags just one, a note on the marketing persona that nothing connects to the rest. The network still does not track its undocumented areas, but it exposes the notes it has failed to integrate — a gap of another order, and the only one a graph can see.

Layer added by "Smoothing over contradictory customer feedback is a product mistake" (2026-11-04). A contradiction recorded in the state of the work — what seems settled, what remains open, the next action — reappears every time the case is picked back up. Nobody can forget it for convenience, or decide in committee as if it did not exist; it stays open as long as the sources do not make it possible to settle it. This is the use of the state layer that a network of notes does not provide.

Open questions

• What, in a corpus, makes it possible to detect a contradiction nobody has flagged?