Step by step
The methods, one record at a time
The rest of the site shows what the records say. These pages show how the computed layers are made, by following one real record through every step and letting you change the record. Nothing here is a textbook example: the titles, the component values and the distances are the release's own, recomputed in your browser so each number can be checked against the published one.
Page 1
From words to geometry
One real headline or bill followed from its title to its vector, its nearest records in four languages, the fold to two dimensions and the map of every embedded record.
Page 2
Data as geometry
One country-year of the influence index followed from six raw measurements to a point, the composite as a direction, distances to other country-years, the two sub-indices as axes and the whole record folded to two.
Page 3
A classifier, step by step
The pipeline's stance head trained in your browser: one layer of weights, the forward pass on a real record, the loss, the gradient, four hundred steps of descent and the decision it reaches.
Page 4 · planned
The money as a network
The finance and trade records as a graph of countries, lenders and projects: degree, paths and communities on the real ledger.
Page 5 · planned
Reading the same text twice
The hand-coded validation sample against the model's labels: agreement, where the two disagree and what that means for the stance figures.
Pages 4 and 5 wait on records that are not yet in the release: the network needs the finance graph export and the second reading needs the owner's coded validation sample.