justinzeus/spreadlab
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

engine: additive forward probability, behaviour-preserving

Browse source 
Forwarding was a single global ForwardProb; make it a per-student
composite (Config.ForwardChance): baseline propensity raised by the
fake's Novelty, lowered by ambient HarmAwareness, and scaled down for an
educated student by ProgramEffect (1 = today's hard block). RunCascade
now takes a precomputed per-student []float64 chance and has no education
special case. Defaults are behaviour-neutral, so the 82/58/6 golden
tests are unchanged; the model is tuned in a later slice.
This commit is contained in:
Justin Visser 2026-06-18 20:03:04 +02:00
parent 4c58e8a0f5
commit a0635544d9
4 changed files with 148 additions and 21 deletions
Download patch file Download diff file Expand all files Collapse all files

61
internal/engine/scenario.go
View file

@ -25,9 +25,17 @@ type Config struct {
NumStudents int `json:"numStudents"`
EdgesPerNode int `json:"edgesPerNode"` // attachment edges per new student (network density)
TriangleProb float64 `json:"triangleProb"` // chance to close a friend-of-a-friend triangle
ForwardProb float64 `json:"forwardProb"` // chance a student forwards the fake along an edge
NumEducated int `json:"numEducated"` // students the education program reaches
Origin int `json:"origin"` // student who first posts the fake
// Forwarding is an additive composite (see ForwardChance): a baseline
// propensity, raised by how novel/shocking the fake is, lowered by the
// year group's ambient harm awareness.
ForwardProb float64 `json:"forwardProb"` // baseline chance a student forwards the fake along an edge
Novelty float64 `json:"novelty"` // how novel/shocking the fake is (0..1); raises forwarding
HarmAwareness float64 `json:"harmAwareness"` // ambient AI-literacy / harm awareness in the year group (0..1); lowers forwarding
NumEducated int `json:"numEducated"` // students the education program reaches
ProgramEffect float64 `json:"programEffect"` // how strongly the program suppresses an educated student's forwarding (0..1; 1 = never forwards)
Origin int `json:"origin"` // student who first posts the fake
GraphSeed uint64 `json:"graphSeed"`
ThresholdSeed uint64 `json:"thresholdSeed"`
@ -42,7 +50,10 @@ func DefaultConfig() Config {
EdgesPerNode: 3,
TriangleProb: 0.45,
ForwardProb: 0.38,
Novelty: 0, // behaviour-neutral until the model is tuned (slice 4)
HarmAwareness: 0, // "
NumEducated: 36,
ProgramEffect: 1.0, // a perfect program: today's hard block, softened in slice 4
Origin: 0,
GraphSeed: 17,
ThresholdSeed: 2,
@ -50,6 +61,35 @@ func DefaultConfig() Config {
}
}
// Forwarding-composite weights: how far each lever can move the baseline
// forwarding probability. These are provisional, illustrative values, not
// fitted to data; they are tuned for legible behaviour in slice 4.
const (
noveltyWeight = 0.30 // a maximally novel fake adds up to +0.30
harmAwarenessWeight = 0.40 // a maximally aware year group subtracts up to -0.40
)
// maxForwardChance caps the composite: even the most novel fake in the most
// permissive world is never forwarded with certainty.
const maxForwardChance = 0.95
// ForwardChance is the probability that a student forwards the fake along one
// friendship in one round: the additive composite at the heart of the model.
// A baseline propensity is raised by the fake's novelty and lowered by the
// year group's ambient harm awareness; a student the program reached then has
// that propensity scaled down by the program's effect. educated reports
// whether the education program reached this student.
func (config Config) ForwardChance(educated bool) float64 {
propensity := config.ForwardProb +
noveltyWeight*config.Novelty -
harmAwarenessWeight*config.HarmAwareness
propensity = min(max(propensity, 0), maxForwardChance)
if educated {
propensity *= 1 - config.ProgramEffect
}
return propensity
}
// IntBounds is an inclusive allowed range for an integer Config field.
type IntBounds struct {
Min int `json:"min"`
@ -169,7 +209,20 @@ func RunScenario(config Config, strategy Strategy) (Result, error) {
if educated == nil {
educated = []int{} // a nil slice marshals to JSON null, not []
}
cascade := RunCascade(graph, config.Origin, config.ForwardProb, educated, thresholds)
// Fold the education lever into a per-student forwarding chance: an
// educated student's composite is scaled down by the program effect, so
// the cascade itself needs no special case for education.
isEducated := make([]bool, config.NumStudents)
for _, student := range educated {
isEducated[student] = true
}
forwardChance := make([]float64, config.NumStudents)
for student := range forwardChance {
forwardChance[student] = config.ForwardChance(isEducated[student])
}
cascade := RunCascade(graph, config.Origin, forwardChance, thresholds)
return Result{
Strategy: strategy,
Educated: educated,