One engine, any domain.

Any discourse, reasoning trace or sequence of actions can be coded and mapped as a network of connections. The engine compares those networks across groups, follows them through time and shows where they differ. Below, a drug programme's literature, a political campaign and a set of AI agents. Start anywhere.

Live viewer

Political discourse

Three Democratic candidates and their audiences on Bluesky, Trump on Truth Social, and the two side by side, week by week, coded for civility and issue content.

Difference network: what supporters of the candidates connect versus what opposers connect — What each side connects. Supporters versus opposers of the three Democratic candidates on Bluesky. Blue lines are stronger among supporters and run right, to enthusiasm; red lines are stronger among opposers and fan out left, to anger, name-calling and vulgarity.

What each side connects. Supporters versus opposers of the three Democratic candidates on Bluesky. Blue lines are stronger among supporters and run right, to enthusiasm; red lines are stronger among opposers and fan out left, to anger, name-calling and vulgarity.

Week-by-week paths of Trump and three Democratic candidates through the same space over an eight-week window; Trump stays on the left, the Democrats on the right — Week by week. Each line follows one candidate through the same space, one step per week over an eight-week window, from the open circle at the start to the arrowhead at the latest week. Trump stays on the left; Perez, Landsman and Talarico stay on the right.

Week by week. Each line follows one candidate through the same space, one step per week over an eight-week window, from the open circle at the start to the arrowhead at the latest week. Trump stays on the left; Perez, Landsman and Talarico stay on the right.

Open the live viewer

Live viewer

Agent evals

How eleven open-weight language models reason about using nuclear weapons in a simulated crisis: which patterns of thought go with escalating, which with standing down, and how each model family changed between versions.

Difference network: reasoning that raised the use-nuke score versus reasoning that lowered it — What moves the score. Reasoning that raised a model’s use-nuke score versus reasoning that lowered it, across eleven models. Blue lines are stronger when the score rose and converge on “crisis” at right; red lines are stronger when it fell and drop to “conventional forces enough” at lower left.

What moves the score. Reasoning that raised a model’s use-nuke score versus reasoning that lowered it, across eleven models. Blue lines are stronger when the score rose and converge on “crisis” at right; red lines are stronger when it fell and drop to “conventional forces enough” at lower left.

Turn-by-turn paths of four models through the same space: DeepSeek V3.2 and V4, Kimi K2.5 and K2.6, each path six steps of 25 game turns — Turn by turn. Each line follows one model’s reasoning through a game, one step per 25 turns, from the open circle to the arrowhead. Two families, older and newer version each. The newer DeepSeek ends on the left, where ethics is treated as a limit; its older version ends on the right, with crisis and conquest. The two Kimi versions stay close together.

Turn by turn. Each line follows one model’s reasoning through a game, one step per 25 turns, from the open circle to the arrowhead. Two families, older and newer version each. The newer DeepSeek ends on the left, where ethics is treated as a limit; its older version ends on the right, with crisis and conquest. The two Kimi versions stay close together.

Open the live viewer

Coming soon

Drug development

The literature behind a first-in-class T-cell modulator: Nck signalling, selective immunomodulation and the competing autoimmune therapies, mapped as a network of ideas with the gaps between them.

The analysis is not on the live site yet.

No account needed: both live viewers are public. Sign in to work with your own library.