Zwei abstrakte Akteure, Blau und Rot, steigen auf einer Leiter mit neun Eskalationsstufen auf und ab. Tausende simulierte Krisen ergeben eine Verteilung der Ausgänge. Verändern Sie Entschlossenheit, Publikumskosten und vor allem das Wahrnehmungsrauschen. Kernaussage: Fehlwahrnehmung wiegt schwerer als Absicht.
Two abstract actors, Blue and Red, move up and down a ladder of nine escalation rungs. Thousands of simulated crises yield a distribution of outcomes. Change resolve, audience costs and above all perception noise. Key point: misperception weighs more than intent.
Two abstract actors, Blue and Red · nine rungs · thousands of simulated crises · the outcome distribution, not a forecast
Each run is one crisis between two abstract actors, Blue and Red, on a nine-rung ladder: 0 status quo, 1 diplomatic protest,
2 economic measures, 3 military signalling and exercises, 4 mobilisation, 5 limited use of force,
6 strikes on military targets, 7 major conventional war, 8 unrestricted war. A crisis starts with one actor at rung 2 and the other at rung 1.
Every turn, each actor observes the other's rung with Gaussian perception noise (σ in rungs; a hotline halves σ, the AI advisory multiplies it by 0.35 but
replaces 3 % of readings with a confident error of two to three rungs, biased towards threat inflation) and then chooses to escalate, hold or de-escalate.
The three options are scored and drawn from a softmax (temperature 0.27). Escalation is pushed by resolve × stakes and by the perceived gap to the opponent
(matching pressure, the security dilemma); it is held back by the running cost of the next rung and, from rung 6 upwards, by fear of war that is stronger for
low-resolve actors. De-escalation is pulled by low resolve, accumulated cost and a perceived opponent below oneself; it is penalised by audience costs that scale
with how far one has climbed above the current rung. With off-ramps enabled, the penalty shrinks when the opponent is seen to step back and collapses when both step back
at once, which then covers two rungs. At rung ≥ 3 an accident adds a rung with the set probability. Cost per turn is 0.006 + 0.011 · rung² per side.
Terminal states: war once either actor reaches rung 7, settlement when both sit at rung ≤ 1 for two consecutive turns, stalemate after 30 turns.
The Monte Carlo draws N independent crises from a seeded mulberry32 generator, so the same seed reproduces the same figures; work is chunked with
setTimeout to keep the page responsive. The counterfactual table re-runs 2,000 crises with one setting changed. The sweep evaluates P(war) at eleven
noise levels from 0 to 2 with a smaller sample per point and shows a 95 % Wilson interval. Everything runs locally; nothing is sent anywhere.
The sensitivity curve is the point: outcomes hinge less on what the actors want than on how well they read each other, and a modest error in situational assessment can multiply the probability of the worst case. An AI advisory that reduces noise helps most when the fog is thick — but its rare, confident misreads add a tail risk that is invisible in the average, so in clear conditions the tool can make things worse. Mature organisations therefore do not ask whether a model is usually right; they ask how it fails, how calibrated its confidence is, and whether the human in the loop still has the time, information and authority to overrule it.
Research and teaching demo. The model is deliberately abstract: Blue and Red are not real countries, the parameters are not estimates, and no result is a prediction or a claim about any real crisis. Not policy or investment advice. All figures are generated in your browser from synthetic runs.