21csim

Monte Carlo counterfactual simulator for 21st century world history (2000–2100). Live at 21csim.com.

A command-line tool and web viewer that models the major inflection points of the 21st century as a directed acyclic graph with probabilistic branching, runs Monte Carlo simulations sampling from historically-informed probability distributions, and renders alternate histories as narrative output.

Each run produces a plausible alternate 21st century. Batch runs reveal which decisions had the most leverage over how the century turned out.

Credit to Alex Langston's NYX99 for the simulator inspiration.

Quick Start

pip install .
21csim run --seed 42                    # Watch seed 42 unfold
21csim run --seed 42 --mode cinema      # Cinematic mode (default)
21csim run --seed 42 --mode interactive # Step through with keyboard
21csim run --seed 42 --mode sprint      # One-line-per-decade summary
21csim run --seed 42 --format json      # JSON output
21csim run --from-reality               # Start from the real 2026 world state
21csim batch --iterations 1000          # Run 1000 simulations
21csim diff 42 7714                     # Compare two seeds
21csim sensitivity --node 2003_iraq     # Sensitivity analysis
21csim what-if --set 2000_election=gore_wins --set 2001_911=plot_disrupted

Web Viewer

cd web && python3 serve.py
# Open http://localhost:8080

Browse 497 pre-generated runs with cinematic playback, AI narration, world-state dashboard, and a generative Tone.js score. Every curated run has a shareable /century/<seed> page with its own social card.

Forge mode ("Forge Your Own" on the landing page) runs the full simulation engine in the browser (web/sim.js, a verified port of the Python engine — exact output parity on locked runs, verdict distribution within ~2pp of the Python 10k batch). Playback pauses at eleven high-leverage decision nodes — the 2000 election, 9/11, Iraq, the 2008 crisis, COVID response, the AMOC collapse, the AI existential moment, and more — and your choice re-simulates the rest of the century instantly. Finished centuries get a shareable ?forge= link that replays your world and lets the recipient re-decide.

Life thread — a quiet strand under the world events follows one person through each run (birth year and hometown vary per seed; milestones are chosen from the run's actual world state). Toggle with the "Life" button.

Reality Stream

The simulator carries a rolling snapshot of the actual world (data/reality_2026.yaml) so runs can start from today instead of 2000:

21csim update-reality --dry-run           # NewsAPI + LLM analyst propose updates
21csim apply-reality-updates updates.json # Apply a pre-computed analysis (no API calls)

Both paths validate and clamp every value against the engine's legal ranges before anything touches disk. A scheduled Claude Code routine runs the analysis weekly from web research and opens a reality-update/YYYY-MM-DD pull request for review (see .claude/skills/update-reality/).

Architecture

• 303 event nodes spanning 2000–2100 as YAML files
• 50 world-state dimensions (climate, politics, technology, security, demographics)
• Cascading dependencies — upstream outcomes shift downstream probability distributions
• Conditional gating — future nodes only fire if world-state conditions are met
• LLM narration — pre-generated Claude narrations for curated seeds, local Ollama for custom seeds
• 8 verdict categories — Golden Age, Progress, Muddling Through, Decline, Catastrophe, Extinction, Transcendence, Radically Different

Project Structure

src/csim/
  cli.py            — Typer CLI
  engine.py         — Simulation engine (DAG traversal + sampling)
  graph.py          — DAG construction, dependency resolution
  models.py         — Data models (SimEvent, SimOutcome, BatchResult)
  world_state.py    — World state + composite scoring + verdict classification
  renderer.py       — Rich terminal output (cinema, sprint, interactive modes)
  sound.py          — Runtime audio synthesis (numpy + afplay)
  analysis.py       — Batch statistics, sensitivity analysis, what-if
  exporter.py       — JSON export for web viewer
  reality_stream.py — News ingestion -> world state updates
  llm/              — LLM providers, narration, news analyst
  data/nodes/       — 303 YAML event definitions
  data/curated/     — 497 pre-generated seeds with narrations

web/
  index.html        — Cinematic viewer SPA
  explore.html      — Corpus analytics / Explorer
  sim.js            — In-browser simulation engine (port of engine.py)
  forge.js          — Interactive decision mode
  persona.js        — Life-thread generator
  serve.py          — Development server

scripts/
  export_sim_nodes.py       — Node data export for the browser engine
  validate_sim_parity.py    — Python/JS engine parity fixtures
  generate_century_pages.py — Per-seed share pages + sitemap

tests/              — pytest suite (100 tests, run in CI with ruff)

Development

pip install -e ".[dev]"
ruff check src tests && pytest          # what CI runs
python scripts/export_sim_nodes.py     # after editing node YAMLs
python scripts/generate_century_pages.py  # after regenerating the run library

Requirements

• Python 3.11+
• Dependencies: typer, numpy, networkx, rich, pyyaml

License

Split license:

• Code (engine, CLI, web viewer, scripts) — MIT. Use it for anything.
• Simulation data & creative assets (the node library, curated runs, narrations, audio, social cards) — CC BY-NC-SA 4.0. Share and adapt with attribution for noncommercial purposes; commercial use of the data requires permission.

The node library is the research product of this project — hundreds of hand-calibrated probability distributions and cascade structures. The split keeps the engine open while protecting that work from commercial reuse.