Atropos
Name something and break it. Atropos follows the recorded dependencies outward and shows what stops, in what order, and roughly how soon. A trace through evidence — not a forecast.
What it actually does
Choose a subject — a strait, a supplier, a commodity, a power station — and a severity, from a partial reduction to a total loss. Atropos walks the dependency graph outward from it and reports what is exposed.
Every hop shows the link it travelled and the figure behind it. If a step claims one country's mineral imports run through another, it shows the share and the year it was measured. You can disagree with any individual hop — that is the point. The output is an argument you can audit, not a number you have to accept.
The cascade. Each branch is one recorded dependency; thickness is the share carried, colour the channel. Hover any node in the real thing for the figure behind it.
The reach. The same run drawn geographically — no basemap, because the shock's own spread is the picture.
Fig. 02 — Atropos, running. Hormuz closed, two degrees.
How the graph is built
Links are ranked by how they were established, and the rank travels with them everywhere.
A statistical link is never presented as a flow of anything. Two things moving together is a correlation, not a pipe, and the interface says so.
How a shock travels
Four channels, named here as Atropos itself names them, kept separate because they behave differently and collapsing them would be a lie about the world.
| Channel | Meaning | Speed | Weight |
|---|---|---|---|
| Dependency | Something you rely on stopped — a supplier, an input, a passage | Immediate | full |
| Demand | You lost a buyer — a revenue shock, not a halted line | Quarters | well below dependency |
| Containment | A place was hit, so what it holds was hit — walked from the place to its contents, never the other way | Immediate | by relation |
| Security | A standing commitment was engaged | Variable | damped onward |
Security deliberately does not chain. An attack draws in a guarantor; it does not then draw in everyone that guarantor knows. Allowing it to chain turns one invasion into a world war on every run, which is the most seductive wrong answer this model could give.
Each step weakens the effect by how much of the target's dependence actually runs through that link, and accumulates delay along the way. Arriving later does not additionally weaken a path — the horizon already encodes the timing claim, and charging for it twice was tested and made the ordering measurably worse.
Honestly
A score means more exposed than, not a percentage of anything. The ordering carries information. The magnitude is not calibrated and should not be read as though it were.
The ordering has been scored against what actually moved after real historical shocks,
against a random ordering and a "big things move most" baseline, and the rank correlation is
published beside both baselines rather than the better of the two — including when it is
negative, as the aggregate currently is. The sample is a handful of shocks and is not a test that
proves anything either way; the count and the figures are at /api/v1/graph/stats
under accuracy, and the sample grows as shocks are scored.
Atropos models what is recorded. A dependency nobody publishes is one it cannot see, and those are frequently the ones that matter most.