Ein synthetisches soziales Netz mit sechs Gemeinschaften, in dem sich ein abstraktes Narrativ ausbreitet, organisch und über wenige koordinierte Konten. Wählen Sie Zeitpunkt und Art der Gegenmaßnahme: Kennzeichnung, Drosselung, Richtigstellung oder Sperrung. Achten Sie darauf, wie schmal das Zeitfenster für Früherkennung ist.
A synthetic social network with six communities in which an abstract narrative spreads, organically and via a few coordinated accounts. Choose when and how to intervene: labelling, rate-limiting, debunking or takedown. Note how narrow the window for early detection is.
Information operations and intervention timing · 1,500 synthetic accounts in six communities · "Narrative X" is abstract, no real platform or event
The network is a stochastic block model with 1,500 synthetic accounts in six communities (A–F, 170–330 accounts each). Every account carries a heavy-tailed "fitness" weight, so a few hubs collect 50–80 connections while the median account has 7; dense links inside communities, about one in nine links across them, plus two explicit bridge accounts per community. The layout is a seeded circle placement followed by a short force relaxation, computed once at load. Nothing is drawn from a real platform.
Spread follows an SIR-style process on the graph with five states: unaware → exposed → sharing → fatigued (silent), and corrected (immune). Per tick, each sharing account exposes every unaware contact with probability β × susceptibility(community) × susceptibility(account); account susceptibility is exponentially distributed (mean 1), so many accounts are hard to reach. Exposed accounts start sharing with a complex-contagion rule: adoption probability 0.35 per tick once the sharing share of their contacts reaches θ (or at least four contacts share), otherwise one fifth of it; homophily weights how much same-community contacts count. Exposed accounts that do not adopt lose interest at 12 % per tick and fall silent; sharers go quiet after a geometric number of ticks (mean 4). The organic scenario starts from 16 low-degree accounts in communities C and D. Coordinated accounts are extra nodes with a 1.5× posting rate, 12 links (60 % into their home community, weighted towards hubs, 20 % each into two other communities), no fatigue, and a sustained search for new audiences — every tick each one re-points one of its links to a fresh target with probability 0.5 — until a takedown removes them.
1 − 0.8 × strength from the start tick.5 × (1 − strength), at least 1). It bites hubs, not a swarm of low-degree accounts — which is the point.0.2 + 0.6 × strength (12 % of that if they never saw the narrative — corrections travel badly among the uninterested), and 40 % of the corrected accounts pass the correction on for a few ticks. Coordinated accounts ignore corrections.k × (0.3 + 0.75 × strength) coordinated accounts at the start tick (strength stands for detection recall).Attribution is a counterfactual: the same seed is replayed with the coordinated accounts absent and the organic seeds unchanged; the KPI shows how much of the current reach would not have happened. Monte Carlo repeats the scenario across 100 random seeds; the window-of-opportunity curve re-runs the chosen intervention for start ticks 0, 5, …, 60 with 25 common seeds each and plots the median and interquartile band against a no-intervention baseline. Limits: the graph is synthetic and uncalibrated, there is no recommendation algorithm, no content, no demographics; parameters are illustrative, not measured. The demo exists to make timing effects visible, not to predict a real campaign.
Detection is rarely the bottleneck: a classifier can flag coordinated behaviour within a few ticks, but the window curve shows that reach is governed by when the organisation acts far more than by which measure it picks — decision latency eats the gains of detection latency. Mature organisations therefore pre-authorise response playbooks, so a detection can trigger action without a committee, and they measure effect against a counterfactual (what would have happened without the seeds, or without the response) rather than against raw counts. Reporting "corrections delivered" without "share of the exposed who saw one" is the kind of vanity metric this model is built to expose.
Research and teaching demo. Fully synthetic network and behaviour; no real platform data, no real accounts, no political content — "Narrative X" is a placeholder. Not policy, security or investment advice. Everything runs locally in your browser; no data leaves the page.