How the model works, and where its numbers come from
++ This page documents the simulation behind the + + tab so its results can be examined and reproduced rather than taken on trust. It describes + the model loosely following the ODD protocol (Overview, Design concepts, Details), the + recognised standard for reporting agent-based models, then lists the literature behind each + parameter. +
++ Illustrative, not validated. spreadlab is a planning and discussion aid, + not a prediction. Its parameters are explored as sensitivity ranges drawn from adjacent + literatures, not fitted to real-school data, so its numbers must not be read as forecasts + about any real school, platform, or incident. +
+How the model works
+ +Purpose
++ The model asks one question: when a non-consensual deepfake starts spreading through a + school year group, where does a limited education budget actually make a difference? It + runs the same outbreak in the same simulated network under different education strategies, + so the only thing that changes between scenarios is who gets educated. +
+Entities, state variables, and scale
++ The world is one year group of students (120 by default). Each student is an agent with a + fixed set of friends, a forwarding probability, and a state: not reached, reached + (received the fake), or educated. One student is the origin who first posts the fake. Time + advances in discrete forwarding rounds. There is no spatial geometry in the engine, the + layout you see on the Model tab is computed in the browser purely for display. +
+The friendship network
+
+ Friendships come from the Holme-Kim model of preferential attachment with triangle
+ closure, a port of networkx’s
+ powerlaw_cluster_graph. Each new student attaches to a few existing students,
+ more often to already-popular ones (which produces hubs), and sometimes closes a triangle
+ with a friend of a new friend (which produces the dense, clustered friend groups of a real
+ year group). The port preserves the model’s structure, not networkx’s exact random stream.
+
+ Holme & Kim (2002), growing scale-free networks with tunable clustering. +
+Process and scheduling: the spread
++ The fake spreads as an independent cascade. In round 0 the origin posts it. In each later + round, every student who just received it gets one chance to forward to each friend who + has not yet seen it; whether that forward happens is a single random draw against the + student’s forwarding probability. The draw for each friendship is fixed up front and + reused across every scenario, so all strategies are compared in the identical world (the + same outbreak, the same luck) and the only difference is the intervention. The cascade + runs until no new student is reached. +
++ Kempe, Kleinberg & Tardos (2003), maximizing the spread of influence through a social + network. +
+Design concepts: the forwarding probability
++ A student’s chance of forwarding is an additive composite, not a single fixed number. A + baseline propensity is raised by how novel or shocking the fake is and lowered by the year + group’s ambient harm awareness, then clamped to a sane range. For a student the education + program reached, that propensity is scaled down by the program’s effect (a strong but + imperfect reduction, not a perfect wall, so an educated student forwards less but not + never). Education therefore needs no special case in the spread itself: it simply lowers a + student’s forwarding chance. +
+Design concepts: the education lever and the three strategies
++ The program reaches a fixed number of students (30% by default). Which students it reaches + is the lever this tool exists to make visible. Three strategies are compared on the + identical budget: +
+-
+
- No program: nobody is educated, the baseline outbreak. +
- Educate at random: the program reaches a uniform random share. +
- + Educate the most-connected: the program reaches the students with the + most friends, the hubs whose forwarding keeps the network connected. + +
+ Under the tuned default world the fake reaches roughly 83% of the year group with no + program, 69% educating at random, and 18% educating the most-connected. Same budget, very + different outcomes: who you teach matters more than how many. +
+Design concepts: determinism
++ Every source of randomness (the network, the per-friendship draws, the random education + pick) flows from seeds in the configuration. Identical settings reproduce identical + results on any machine, which is what lets a shared link recreate the exact same picture + and lets the project’s tests pin reach values. Nothing is left to the wall clock. +
+Details: parameters as sensitivity ranges, not fitted values
++ Because direct empirical data on adolescent sexualized deepfake diffusion are still + scarce, the model does not bake in point estimates. The curated levers (baseline + forwarding, novelty, harm awareness, program effect) are explored as low / medium / high + ranges drawn from adjacent open-access literatures on misinformation spread, complex + contagion, peer aggression, cyberbullying interventions, and comprehensive sexuality + education. The remaining literature parameters are documented and reserved for the + calibrated instrument the grant would fund. The list below maps each parameter to its + sources. +
+Sources behind each parameter
++ Drawn from the project team’s review of open-access literature. Each parameter is mapped to + the source(s) it is grounded in. Live parameters are controls in the demo today; reserved + parameters are documented in the model and wired into the funded, calibrated instrument. +
+ +| Parameter | +In the model | +What it does | +Grounded in | +
|---|---|---|---|
| {{ param.name }} | ++ + {{ param.status === 'live' ? 'Live control' : 'Reserved' }} + + | +{{ param.role }} | +
+
|
+
References
++ Network and spread mechanics: Holme & Kim (2002), + growing scale-free networks with tunable clustering; Kempe, Kleinberg & Tardos (2003), + maximizing the spread of influence through a social network. +
+Parameter sources, from the project team’s open-access literature review:
+-
+
- + {{ ref.code }} + + {{ ref.text }} + open access + + +
+ The full instrument the grant would fund replaces these illustrative ranges with values + calibrated to real-school data, adds complex-contagion and per-agent heterogeneity, and + searches for the best intervention under a budget. This demo is the accessible, + intuition-building layer beneath it. +
+