The graph maps our attention, not the world

Summary

Any large system contains patterns, including a random one, so finding structure in a corpus graph proves nothing. And our graph mostly records what we chose to write about. We registered one narrow question that uses that bias instead of hiding it: does an article’s novelty in the graph predict whether its claim later resolves correct?

Two traps

The first is mathematical. Ramsey theory guarantees that structure is unavoidable in any large enough system, random ones included. “We found a pattern” is therefore not evidence of anything.

The second is ours. Two places co-occur in dozens of our pieces because we wrote dozens of pieces about them. Mining that for market signal would build a very convincing predictor of what we write next.

The question we registered

Not “does the graph predict markets”, but: does an article’s position in the corpus graph at the time it was written predict whether its claim later resolves correct? That question is meant to study our own attention, so the bias stops being contamination. The answer is calibration about which of our readings to trust.

Fixed in advance

Five features, each computed only from what existed on the article’s publication date: how new its pairings of actors were, how familiar its actors were, whether it bridged two unconnected clusters, how dense its neighbourhood was, and how large the corpus was at the time.

One primary test: the Spearman correlation between novelty and resolution. No test until at least 60 claims have resolved. A threshold of p < 0.01, because five available features make fishing tempting. The other features are exploratory and will be labelled that way.

The work needed a time axis the archive had lost in a migration. We rebuilt it from each article’s own citations, on the rule that a piece cannot cite a source published after it: 195 of 195 articles dated, spanning 17 April to 22 August 2026.

What counts as a refusal

Fewer than 60 resolved claims when we look; p at or above 0.01; or a correlation that reverses between the first and second halves of the corpus. A refusal will be published like any other result.

Limitations

This studies ARCANE’s own reliability, not markets, and the answer will not transfer to other publications.
Resolution depends on the editor-reviewed ledger described in the method card.
No result is reported here.

What would change this

Any of the three refusal conditions above.

Source boundary

Published: the question, features, test and refusal conditions. Withheld: entity extraction and the claim ledger contents.

Changelog

23 Aug 2026Registered before any claim resolved.
22 Sep 2026First public edition on Labs.

Cite

@techreport{arcane2026graphmaps,
  title       = {The graph maps our attention, not the world},
  author      = {{ARCANE Labs}},
  institution = {ARCANE Intel},
  type        = {Pre-registration},
  year        = {2026},
  url         = {https://arcaneintel.net/labs/graph-maps-our-attention}
}