How to read the graph
Method
A claim on the canvas is never just a sentence. Its border tells you how well it is grounded, its badge whether anyone could act on it, and its panel which texts and numbers stand behind it. This page defines those encodings and the process that produces them.
Three support states
-
Supported. At least one aligned expert citation backs the claim with a verbatim passage. Drawn with a solid border and a filled band on the canvas.
-
Hypothesized. The claim came from a lead and no aligned expert has grounded it yet. It stays on the map as an open question. Drawn with a dashed border and a hollow fill.
-
Contested. Aligned citations disagree, or an aligned voice disputes the claim outright. Drawn with a double border so the state is legible without colour.
Colour is never the only signal. Each state has its own border style, so the graph reads the same in greyscale and in print.
Fixability
Every claim also says whether the condition it describes can be changed by the actors in the graph's scope: fixable, unfixable, or unknown until someone has made the call. A hammer badge on the claim card carries the verdict. The distinction is what makes a causal map useful for action rather than for description: the fixable branch is the agenda, the unfixable branch is the context the agenda must live with.
Two engines, never mixed
Claims and voices are produced by a language model and gated by a person. Opinion pieces are read for candidate causes, candidates are abstracted into claims, and expert texts are judged passage by passage against the claims they might ground. Each judgement becomes a proposal with a confidence score and the exact excerpt.
Statistics travel the opposite way. Indicators are fetched from official statistical offices through typed adapters, validated against a schema, and attached to the claims they quantify. No language model touches a number on its way into the graph, and institutions on the source blacklist get no adapter at all.
The human gate
No proposal publishes itself. A person reads each proposed citation, checks the excerpt against the source, and accepts or rejects the row. Only an accepted citation can change a claim's support state. The error rate of the model's judgements is being measured on the first few dozen rows before any automatic acceptance is considered.