A C++ agent-based simulation of misinformation spread across real country populations. Select a country, watch 25 named agents move across its map, and observe how a single false belief propagates — or gets permanently corrected — through their social network.

A single misinformation packet is injected into one agent (patient zero) at the start. From there it spreads via two parallel mechanisms: proximity (agents who are physically close on screen share information face-to-face) and social graph (a Barabási–Albert preferential-attachment network carries information across distance, at reduced strength). TikTok bypasses both — its For You Page algorithm can reach any susceptible agent regardless of proximity or connection. Each agent has a name, real demographic parameters drawn from their country's data, and up to three social media platforms.

Every agent is always in one of five states:

Immune is distinct from Recovered. An agent reaches IMMUNE only after: being exposed, rejecting it (prior_exposure_count++), being exposed again, rejecting again, then later recovering from infection. High-literacy agents with that prior-rejection history have developed genuine epistemic immunity — they cannot be re-infected. Low-literacy populations even with multiple exposures recover to RECOVERED, not IMMUNE, because their threshold never stabilised. In high-literacy countries the IMMUNE population is larger; in Perfect Storm scenarios with very high emotional valence, most agents get infected before accumulating the two rejections needed, so fewer ever reach IMMUNE — the immune core collapses.

The herd immunity threshold playing out correctly. Each tick, infected agents have a small natural recovery chance (1.5% × literacy × (1 − correction_resistance)). As the susceptible pool shrinks, transmission slows, recoveries accumulate, and the misinformation runs out of new hosts. This is the same structural mechanism as epidemic herd immunity, applied to belief propagation.

Each simulation tick runs 12 sequential phases:

1. Proximity spread — every INFECTED agent checks all agents within 55px. Uses the shared platform between sender and receiver (not just the sender's primary), so platform transmission counts reflect the actual channel used.
2. Social graph spread — Barabási–Albert network (m=4, avg degree ≈8) carries infections across distance at 18% of proximity probability. WhatsApp retains its trust multiplier; Twitter and Reddit remain suppressed.
3. TikTok FYP — 10% × tiktok_infected_count chance per tick of reaching a random susceptible agent. Source agent is logged as UINT32_MAX (no connection required).
4. YouTube entrenchment — infected agents on YouTube gain +0.002 correction_resistance per tick. Resistance ratchets monotonically — it never decreases. A YouTube-infected agent who's been infected for 100 ticks has gained +0.2 correction resistance.
5. Virality burst — 0.4% base / 2.5% during crisis per tick of instantly exposing all susceptible TikTok + Twitter users. The 6× crisis amplification means crisis events don't just raise individual credulity — they dramatically raise platform-level viral event probability simultaneously.
6. Exposure resolution — EXPOSED agents after the window either believe (INFECTED) or reject (SUSCEPTIBLE + prior_exposure_count++).
7. Recovery — INFECTED agents may recover to RECOVERED or IMMUNE (see above).
8. Crisis decay — emotional state decays back toward baseline over the event duration.
9. Scenario auto-events — scripted events (crisis, botnet, fact-check) fire at their programmed tick.
10. EventSystem dispatch — JSON-scripted custom events fire if scenarios/custom.json is present.
11. Buffer swap — double-buffered state (current_states/next_states) swapped atomically — all phases read from current, write to next, eliminating race conditions.

Each platform has independent spread modifiers. The combination that matters most is trust multiplier (how credible a message from this channel feels) × algorithm amplification × correction reach (how much of the infected population a fact-check on this platform can actually reach).

Each repeated fact-check on the same platform decays by 0.8 per deployment:

effective_reach = 0.6 × platform.correction_reach × 0.8^(deploy_count - 1)

A WhatsApp fact-check (5% base reach) deployed three times reaches 0.6 × 0.05 × 0.64 ≈ 2% of infected users. Deploying the same correction on the same channel repeatedly is geometrically ineffective — switching platforms matters more than increasing volume.

Three built-in preset scenarios, loadable via keys 1, 2, 3 or by clicking the left panel:

Scenarios reset all agent states and fire auto-events at precise ticks, logged to the story panel.

Edit scenarios/custom.json to script any timeline without recompiling. All event types are supported:

{
  "events": [
    { "tick": 30,  "type": "INFLUENCER_SHARE",  "zone": "USA", "top_n": 3 },
    { "tick": 60,  "type": "CRISIS_TRIGGER",    "intensity": 0.70, "duration_ticks": 80 },
    { "tick": 80,  "type": "BOT_ACTIVATION",    "botnet_id": 0 },
    { "tick": 180, "type": "INJECT_FACTCHECK",  "platform": "Twitter", "effectiveness": 0.75 },
    { "tick": 250, "type": "INTERNET_SHUTDOWN", "zone": "USA" }
  ]
}

Available event types: INJECT_MISINFO, INJECT_FACTCHECK, CRISIS_TRIGGER, INTERNET_SHUTDOWN, BOT_ACTIVATION, INFLUENCER_SHARE.

Two pre-allocated bot agents are injected via NetworkBuilder::injectBotnet() rather than manual construction. This gives bots a fully-connected clique among themselves, plus direct graph edges to the top high-influence agents in the regular population. When a botnet activates (B key, or scenario tick), the clique means activation is immediate across all bots simultaneously, and the influencer connections mean spread does not wait for proximity — it travels the social graph directly to the most-connected nodes first. Bot activation is therefore not linear in bot count; it is a graph multiplier operating through the network's hubs.

• Language: C++17
• Agent struct: 64 bytes exactly, cache-line aligned (static_assert enforced)
• Spread model: Hybrid — proximity O(N²) within 55px + Barabási–Albert social graph (m=4, avg degree ≈8) + TikTok FYP (graph-independent). All three run in parallel each tick
• Social graph: Built via preferential-attachment after regular agents; bots injected as a separate clique with influencer hooks via NetworkBuilder::injectBotnet()
• Rendering: SFML 2.6 — pixel-art sprite sheet, green political map per country, SIR ring-buffer chart, graph edge overlay, influencer aura, virality flash, platform bar chart, agent roster, event log
• Country borders: Natural Earth 110m GeoJSON → simplified ≤120-point polygons; agents spawn inside the real border via rejection sampling
• Double buffering: current_states/next_states swap at end of each tick — all 12 phases read from current and write to next, eliminating state-race conditions
• Event system: Priority-queue JSON loader (EventSystem) fires scripted events at exact ticks; scenarios/custom.json is loaded automatically at startup if present

Country parameters (literacy, press freedom, polarisation, internet penetration, platform dominance) are derived from real datasets:

git clone https://github.com/Bilal-Waraich/Infodemic-Flux
cd Infodemic-Flux

# Install dependencies (macOS)
brew install cmake sfml@2
pip3 install pandas numpy

# Generate country border header from GeoJSON
python3 scripts/extract_borders.py

# Build data pipeline (agents_config.csv)
python3 scripts/build_agents_config.py

# Build and run
cmake -B build && cmake --build build
./build/simulation 25    # 25 agents (5–60 supported)

The starkest result is structural. WhatsApp's correction_reach of 0.05 versus Twitter's 0.9 is not a tuned parameter — it is a direct encoding of platform architecture. There is no feed, no algorithmic surface, and no public post on an encrypted messaging app to attach a correction label to. The 18× asymmetry means that two countries with identical literacy rates, identical misinformation exposure, and identical fact-check investment can have completely different outcomes depending solely on which platforms are dominant. In a high-WhatsApp country (Nigeria, India, Brazil), corrections are structurally invisible to ~95% of the infected population.

Each repeated fact-check on the same platform decays to 80% effectiveness. The correction window problem is therefore twofold: arriving early matters because correction_resistance ratchets upward monotonically during infection (YouTube rabbit-hole entrenchment adds +0.2 over 100 ticks), AND repeated campaigns on the same channel lose effectiveness geometrically. The correct strategy — switching platforms per deployment — is rarely available when the dominant platform is encrypted.

Bots are connected as a fully-connected clique with direct edges to the highest-influence regular agents. When the botnet activates, it does not simply add N infected agents — it delivers infection probability directly to the network hubs first, who then cascade through their high-degree connections. A two-bot botnet wired to the top three influencers can trigger the same cascade as ten randomly-placed infected agents in isolation. The effective amplification is proportional to the combined degree of the connected influencers, not the number of bots.

TikTok FYP operates independently of the social graph: 10% × tiktok_infected_count per tick can reach any susceptible agent. Virality bursts (which instantly expose all TikTok + Twitter susceptibles) occur at 0.4% per tick baseline, but at 2.5% per tick during a crisis — a 6× increase. A crisis event therefore simultaneously: (a) spikes individual credulity via emotional state amplification, (b) increases virality burst probability 6×, and (c) reduces the analytical processing that would otherwise produce prior_exposure_count rejections — meaning fewer agents will ever accumulate enough rejections to become IMMUNE. The three effects compound non-linearly.

An agent reaches IMMUNE only after rejecting exposure twice (building prior_exposure_count ≥ 2) AND recovering with literacy_score > 0.65. In low-literacy countries, even multi-exposed agents recover to RECOVERED and remain re-infectable. In high-literacy countries under a low-valence scenario, the IMMUNE population grows to a stable core that permanently suppresses future outbreaks. This is visible in the SIR chart as the recovered curve flattening while susceptible agents stop converting — structural epistemic herd immunity, not just temporal correction.

A fact-check deployed on Twitter reaches 90% of infected users on Twitter. But if the dominant spread in a country is via WhatsApp (high trust multiplier, high spread, encrypted), correcting on Twitter reaches a different, already-sceptical audience. The effective correction population is the intersection of infected × uses_twitter — in a high-WhatsApp country, this is a small, largely uninfected subset. The model makes visible what is often stated but rarely quantified: regulatory interventions designed for transparent platforms do not transfer to encrypted ones, and the populations that most need correction are typically on the channel that most resists it.