The Great Rewiring Two resource grabs, one network: modern mercantilism reprices the edges while the AI buildout re-values the nodes.

The Slate: ten forecasts across four layers

We trace one shock through four layers of the same network. Layer A covers states deleting and re-pricing edges; layer B, the graph re-routing and re-healing at two speeds; layer C, the AI resource grab colliding with physical constraints; layer D, labor, the slowest stratum, repricing last. Each card carries the paper's printed probability, its resolution source, and, where a comparable series exists, how far today's reading sits from the resolution threshold. Of note, no card invents a series it does not have. Where none exists, the card says so.

Score It: what the slate earns on a Brier score

A forecast slate is worth only what it scores. Each question below takes a YES or a NO, set either as we expect it to resolve or as it actually resolves once the dates land. From this, the panel computes the paper's Brier score live against two baselines: a forecaster who answers 50% to everything, scoring exactly 0.250, and one who answers NO to everything. Lower is better; 0 is perfect. The skill score reports the fraction of the coin-flip baseline's error the slate avoids. Moreover, the bar under each question depicts its divergence from that coin flip. Right of centre in green indicates the printed probability earned its keep; left in red indicates it cost something.

Chokepoints: the sector-layer multigraph, and what a single removal costs

The aggregate graph is too dense to fragment. However, the strategic layers are not the aggregate, and they afford a far narrower aperture. Each layer is the top of its 2024 BACI HS6 product-layer graph, and in every one a single origin holds 21–63% of world exports. Layer toggles sit below, and any node is removable on click, whereupon the readout recomputes stranded value and component count live. That renders the paper's Exhibit X1 inset executable rather than merely printed. Of note, removing China alone strands approximately 58% of critical-minerals top-edge value and approximately 58% of semiconductor value, and the three strongest nodes together take out roughly 85%.

Attack Lab: topology holds while the value structure fails

We run percolation on the real 2024 world trade web. Topology is robust: the giant component retains essentially all surviving nodes even at 30% targeted removal. Conversely, value is fragile: removing the strongest 10% of nodes leaves only approximately 8.4% of original trade weight connected, against roughly 83.7% under random failure. That gap is the robust-yet-fragile signature. The left panel illustrates precomputed curves on the full 226-node network with 100-seed random bands. Lastly, the right panel runs the same removal live on the real top-40 subnetwork.

Calibration Lab: the test that broke our own null

Before trusting a counterfactual ensemble, we test whether it constitutes a calibrated null. We run every ordinary, untreated dyad through it and examine where the actual value lands in the band. Under a well-calibrated null those percentiles are uniform, yielding a flat histogram. They are not. Approximately 28.3% of dyads fall below the ensemble's 5th percentile and roughly 27.5% above its 95th, because the generator's within-dyad dispersion of 0.104 dex is approximately one quarter of the real cross-sectional spread of 0.421 dex. Simply put, the bands are 4.1× too tight, so a raw percentile of 12 is unremarkable. Narrow bands are no boon when they are wrong. Therefore, the fix is Efron's: we build the null from the cross-section of deviations rather than the model's own bands. Both views appear below; the paper prints the corrected reading.

Rewiring Residual: actual against the no-rewiring counterfactual

We initialize the published synthetic WTW generator, validated against its own paper to within ±3.5%, on the actual 2020 network, then evolve it to 2028 along actual GDP paths across 300 runs. That ensemble therefore describes a world where only GDP moves trade. The gap between it and the actual network is the incremental rewiring since 2020. However, the 2018–19 trade war already sits inside the initial condition, which is why a 2017-fitted gravity benchmark reads a larger total displacement, at China–US 0.35× and Vietnam–US 16.2×, than the generator does. The raw percentile is read against the Calibration Lab above, never on its own.

Method & Provenance