The metastore
Know what your data means, without moving it.
Topos builds a knowledge representation of a data estate. Concepts at the nodes, relationships at the edges, inside a source and across sources. The data never moves, is never federated, and is never remodelled.
Of every source, it states four things.
A source card is not an inventory entry. It is a claim about what the source means, and it is built by reading the source rather than by asking someone to fill in a form.
What it is
The subject, the entities, the grain, and the period it covers.
What it is used for
The questions it already answers, and for whom.
What it could be used for
The questions it could answer, including those needing a second source and a relationship between them.
What it cannot be used for
The questions this source does not hold, with the nearest thing it does hold.
The fourth is the one no other product returns. Every tool in this space answers the question it was asked. None of them says "this estate does not hold what you are asking for, and here is the closest thing it does." Absence is the most valuable thing an estate can tell you.
And a fifth, on every assertion
With what confidence, on whose authority, and revisable in the face of evidence.
A claim is evidence with uncertainty attached. That is what makes it auditable, and what makes it possible to change our mind when the estate changes rather than quietly carrying a stale answer.
Worked example
The answer that is arithmetically correct and invalid.
What is the average visit duration in minutes?
25.5 minutes. Correct arithmetic over the column named duration.
Its own documentation says that column records scheduled duration, not actual. The average is not a measure of how long visits took. The question never mentioned documents.
A lake would have loaded the column. A federation would have returned the number. Neither consults the sentence. Topos does, because the relationship between the document and the table is in the graph. This example runs end to end on a demonstration corpus.
What this is not.
Four approaches are in market. Each is coherent. Each gives up the same thing.
The common failure is that all four return assertions. None carries uncertainty, none distinguishes a declared fact from an inferred one, and none can tell you that the answer it just gave is invalid.
The cost is already on the books
A replacement sale into a known loss.
The last one is the one that matters, because it is not a statement about models. The budget exists, the failure happened, and the diagnosis points at the estate rather than at the model.
How an engagement starts
A design partner engagement. Fixed scope, 90 days, paid.
Read-only access to a real estate and a named sponsor. Three artifacts at the end.
- The bill of materials. What is in the estate, where data of a kind lives, and what the estate cannot answer.
- The contradiction list. Where the estate disagrees with itself.
- The correspondence set. How the sources relate, with provenance and uncertainty on every assertion.
The data never leaves the estate. Nothing is copied, replicated or remodelled.