Knowledge Graph

Books, papers, statutes and code, each split into chunks and embedded. A document belongs to one of 140 subject areas, from category theory and nonlinear science to law, economics and network ecology. The bar chart shows the largest sixteen; five further areas are private and never leave the machine. The daily news stream is a separate graph with its own page: it changes by the hour, this one by the book. Formulas are pulled from the text layer and given a structural signature; what that signature is good for has its own page.

What comes out

One question, several fields

The point of putting textbooks, statutes, papers and source code into one graph is not storage. It is that a question asked in one language gets answered out of subjects that were never filed together.

“how do small local rules produce large scale order without central control”

One query, 6 different subject areas, shown one source per area. Nobody tagged these books as related; the only thing they share is what the sentences mean.

What happens overnight

Reading is cheap. Keeping order is not. A graph that only ever grows turns into a landfill: the same fact under three names, two entries that quietly contradict each other, and thousands of links nobody will follow. So a maintenance pass runs at night, in five stages, each with an output that can be counted.

Every tenth run, the pass attacks the graph

A maintenance routine that only ever confirms what is already there is not maintenance. So every tenth run, stage 2 is replaced: instead of tidying the graph, an engine is pointed at it with the job of finding what is wrong with it. It looks for three things.

What that looks like: a tool the pass wrote for itself

Not a description. This is 16 lines of graph query that came out of the run on 2026-08-29, written to hunt exactly the problem above. Its own note on what it is for:

Circular Validation Detector - agent-scale echo chamber at graph level (node 5863bdcf). For each MetaInsight, compute the internal density of its evidence subgraph: insights whose supporting sources heavily reference each other are confirmation structures, not independent corroboration.

MATCH (i:MetaInsight)-[e]->(s)
WHERE type(e) IN ['SYNTHESIZES_FROM', 'GROUNDED_IN', 'EXEMPLIFIES_PATTERN',
                  'PROVIDES_EVIDENCE', 'SYNTHESIZED_FROM', 'MERGES_STRUCTURE_FROM']
WITH i, collect(DISTINCT s) AS sources
WHERE size(sources) > 1
UNWIND sources AS a
UNWIND sources AS b
WITH i, sources, a, b WHERE a <> b
OPTIONAL MATCH (a)-[r]-(b)
WITH i, size(sources) AS n_sources, count(DISTINCT r) AS internal_links
RETURN i.name AS insight,
       n_sources,
       internal_links,
       toFloat(internal_links) / toFloat(n_sources * (n_sources - 1)) AS internal_density
ORDER BY internal_density DESC, internal_links DESC
LIMIT 25;

The point of the query is the distinction it draws: a conclusion supported by five sources that mostly cite each other is not five-fold corroboration, it is one source counted five times. That is a structure a graph can be asked about, and this asks.

It does not run every night either. A gate measures pressure first: how many open questions the research side has piled up, and how far the graph has drifted since the last pass. Below the thresholds the run is skipped, and the skip is written down together with the numbers that caused it.