Forecast.

First, does the engine say what the paper measured?

In the engine a child's setup can fail. The chance it works starts at 45% and rises with every generation of inherited notes, levelling near 75%. That rule is written from the paper's numbers — but whether a whole swarm, with rent, crowding and defenders, then measures like the paper is a separate question. It does:

Measured over 6,362 setups in 24 undefended weeks. Generation 1 is the founders' children. Lines past generation 6 are too thin to show: few get that deep in seven days.

Three steps past the paper

Every step below is our extrapolation, not a finding of the paper. In Outbreak they switch on in endless mode (day 8 is 2027); here they switch on by year.

Typical R per day, by defence and year

Each panel is one defence package, stacked on the one before it: quarantine; then an alert over every quarantine (the paper's 92/74); then honeypots and signatures; then a detection drone that strikes crawlers in transit. Each point is the geometric mean of daily R over 24 dishes and seven days. New infections arrive at the rim every morning, as they would from outside.

What the model says

• No single layer gets R under one, in any year. Quarantine alone: R 1.28 in 2026. Adding the paper's notification lift moves it to 1.25: real, and small. Honeypots with signatures: 1.16.
• Only the full stack holds — and it's detection that tips it. With the drone, typical R is 0.85 in 2026 and the week ends contained in 92% of dishes.
• Capability erodes the margin. By 2038 the full stack's R is 1.08 and it holds in 42% of dishes. From 2036 on, its typical R never comes back under one in this model.
• Undefended, the dish fills. With no defence the swarm hits the lid in 38% of weeks already in 2026.

The shape of the answer is the paper's own: no tool wins on its own; it's the product of layers that keeps R under one, and every year of capability takes a little of that margin away. What the model adds is a way to see how much, and to argue about it with numbers anyone can re-run.

Limits — read these before quoting a year

• It's a game engine, not a network simulator. Machines are cells, compute is rent, time is ticks.
• One capability knob (15% a year) stands in for a whole trend line. Change it and the years move.
• Defenders here don't improve with time except through signatures. Real defenders would.
• The scripted defender is simple; a better policy could hold longer. The swarm's notes are simple too.
• Twenty-four dishes per point: differences under about 0.05 in R are noise.
• The years are labels on a model, not dates in the world. We make no claim about when, or whether, any of this happens.

Re-run it

npx tsx scripts/forecast.ts            # 65 cells × 24 dishes, deterministic
cat app/lib/forecast-data.json          # what this page draws

Engine outbreak/2 · study forecast/1 · same seeds, same numbers, on any machine. The study and the game share one engine (sim/outbreak.ts); its tests pin the rules.