Verification Community — Member Bonds

A map of who is working on AI compute verification, and how. Every node is a person publishing research in the field; every edge is a paper they've co-written with another node. Colours are clusters found by grouping the co-authorship pattern. As of September 2026, this is roughly the whole field.

Verification Community — Member Bonds

Every node is a member of the verification community under the loose_2026 rule; every edge is co-authorship on a verification-relevant work found by an OpenAlex sweep — not the hand-curated corpus of 46 papers. Edge thickness is the number of shared works. Node colour is the Leiden community of this bond graph (γ=1.0, Q=0.248); same-cluster edges are drawn shorter so the partition separates visually. Click a cluster to isolate it, a work to highlight the members it links, or a node or edge for details.

same-cluster bond
cross-cluster bond
highlighted work
width ∝ shared works (1–10)
node size ∝ √(verification-relevant works)
† = deceased

How to read the graph

  • Nodes are researchers. Colour is the cluster their co-authorship pattern places them in. Node size grows with their number of verification-relevant works.
  • Edges are shared work. Thickness is the number of works two people have written together. Same-cluster edges are drawn shorter so the partition separates visually.
  • Interactions. Click a cluster in the left rail to isolate it; click a work to highlight the members who share it; click a node or edge for details on the right. Drag to pan; scroll to zoom.

Methodology

Membership follows the loose_2026 rule: a person is included if they have at least one verification-relevant work in 2026, or at least three across 2023–2025. Author records come from an OpenAlex sweep, cross-referenced with a hand-curated corpus of 46 papers held in the Canary Institute's verification bibliography.

Clusters come from Leiden community detection on the bond graph (resolution γ=1.0), with the resulting partition modularity Q=0.248. Labels for the seven largest clusters are carried over from an earlier partition by member overlap; the two smallest are named for their dominant organisation.

Limitations

  • Missing people. The rule counts only work that shows up in OpenAlex under a person's own record. Anyone whose verification-relevant contribution is entirely inside an organisation without externally published output is invisible here, as is anyone who publishes under a name OpenAlex has not yet reconciled.
  • Missing edges. Eight members satisfy the rule on their own publication record but share no verification-relevant paper with any other member. They are listed in the left rail but not drawn.
  • Snapshot, not stream. OpenAlex lags recent preprints and grey literature. This is the September 2026 slice; the field is small enough that a handful of papers materially changes the picture, and this page will need refreshes.
  • Cluster labels are lossy. Nine clusters cannot describe the shape of a nine-way multidimensional collaboration. The labels are shorthand for orientation, not ontology.

Citation

The Canary Institute, Verification Community — Member Bonds, canaryinstitute.ai/research/verification-community, September 2026 snapshot.

← All research