The Arithmetic of Self-Replication

One factory lands on the Moon or an asteroid and builds copies of itself from local material. How far that idea actually holds, and where the physics stops it - modelled in small, live pieces, with every number traced to a source.

A long-term, modular exploration of self-replicating space manufacturing - the idea that you could land a single factory on the Moon (or an asteroid, or Mars) and let it build copies of itself from local material, growing an entire industry from one rocket's worth of cargo. The name is the standard term for the concept: "von Neumann"-style self-replicating machines.

This is a monorepo. Each project lives in its own top-level directory, is independently runnable and tested, and models one slice of the problem. They share concepts (and, over time, data) but not a single giant simulation - the whole point is to keep each piece small, honest, and verifiable.

New here? FINDINGS.md collects what the models have concluded so far - nine sourced results, from the electronics wall to the light-speed coordination tax - as one read. The same findings are live and interactive on the frontend's overview. The machines these runs execute on are inventoried in HARDWARE.md.

Modules

Module	Status	What it does
closure-sim	v1	Closure & replication. Define a factory as a bill of materials; compute how much of itself it can build locally ("closure"), simulate how a seed multiplies over time, and analyze the "electronics wall" - why chips are the part that can't be made in space.
frontend	live	The monorepo's central interactive surface. The project's models run live in the browser, built entirely on pimas. A shell that hosts one surface per model; the first is the interactive electronics wall (drag the assumptions, speculate "make its own chips" before committing, watch the model explain which ceiling binds).
probe-sim	in progress	A single self-replicating probe (after Borgue & Hein 2020). Solar-electric power vs heliocentric distance, and the operational range it gates - feeds delivered power into closure-sim's replication to find where a probe stops being able to reproduce.
power-budget	live	Making vs. thinking, per watt. Split a solar-limited power budget among manufacturing, compute, and housekeeping; convert compute-watts to throughput - floored by the Landauer limit and anchored to the ~20 W human brain.
launch-economics	live	The economics of not launching mass. Launch cost, the rocket-equation Δv penalty, and the launch-mass leverage of shipping a self-replicating seed instead of the finished installation.
mission	live	The whole operation, end to end. One pure fold that composes all four modules above: launch a seed, arrive at a heliocentric distance, split its solar power between building and thinking, replicate, and price the launch-mass payoff. The frontend's "Full mission" surface follows the chain stage by stage.
multi-probe	live	A small, deterministic fleet. Tens of self-replicating probes, each an agent that builds copies at a rate its local sunlight allows and disperses children outward. Seeded and pure, so speculate stays exact. Re-instantiates two ceilings: the electronics wall (a finite vitamin pool) and a spatial power wall (1/d² sunlight vs dispersal). The frontend's "Fleet" surface scrubs the 40-year mission live. The intermediate step before the swarm.
swarm	core done + live surface	How fast a probe fills the galaxy. A deterministic, seeded settlement front: probes spread star-to-star through a field, settling and re-launching, filling the reachable galaxy outward from one homeworld (the exploration-timescale question of Nicholson & Forgan 2013). The pure algorithm core, plus three travel policies - powered flight and two gravitational-slingshot policies that steal speed from stellar motion (Nicholson & Forgan's headline: slingshots ≫ powered, and nearest-star beats max-boost on time). A live canvas "Swarm" surface plays/scrubs the fill, toggles the policy, and - hover any star - reads its light-speed coordination lag from home (every interstellar hop is years of latency: the "independent colonies" regime). A proven spatial hash keeps the fill fast. It also simulates light-speed-limited coordination (FRONTIER #1): probes decide against a light-delayed belief of what's settled, so they race for the same star from stale views - costing a median ~30–50% of the exploration timescale in the slingshot regime (but ~0% for powered flight), a live-toggle finding that refines Nicholson & Forgan. The one parked fork left is the 200k-star WebGL render engine.
spine	done	One factory across all three scales. The cross-scale integrator: threads one closure-sim factory through the single-factory fold, the local fleet (multi-probe), and the galaxy (swarm), so their replication cadences are derived from one source instead of assumed independently. It adds no new numbers - it replaces the swarm's old unsourced zero manufacturing dwell with a figure derived from the same build physics the fleet uses, and shows that dwell is a negligible tax on galactic exploration. The frontend's "Across scales" surface follows it.
papers	live	Longer-form write-ups, in LaTeX. The place to publish papers typeset with the IEEE template. A small generator builds each paper's bibliography (refs.bib) from the same sources.ts that powers the site's citations, so a claim in a paper and a claim on a live surface point at the identical reference. CI compiles the PDF and publishes it; the frontend's "Papers" surface lists each one.
more to come		The plan is upwards of ten interacting projects over the coming year - see ROADMAP.md.

Working in here

Each module is self-contained - cd into it and follow its own README. They don't share a top-level build; a module brings its own dependencies and tests.

cd closure-sim
uv venv --python 3.12 .venv
uv pip install -e ".[dev]"
.venv/bin/pytest

Reproducing all results

Every result in this repo is meant to be reproducible from a clone. The eight Python modules use uv and stand alone. The frontend is built on pimas, consumed from npm as pimas-ui (aliased to the bare pimas specifier in frontend/package.json), so npm ci in frontend/ resolves it like any ordinary registry dependency - no sibling clone or source build required.

Run everything - all eight module suites plus the frontend's two test layers and build

• with one command from the repo root:

./scripts/test-all.sh

It exits nonzero if any suite fails.

Conventions

• One module = one directory with its own README, dependencies, and tests.
• frontend is the one shared surface. It's the single interactive/presentation layer, built on pimas and pimas only. It hosts one surface per model rather than fusing them - each model still owns its slice; the frontend just presents it.
• Real tests, real numbers. Assertions check that the math behaves correctly, not just that code runs.
• Grounded inputs. Assumptions trace to real research; modules document their sources (see, e.g., closure-sim/REFERENCES.md).
• Plain language. This work gets shared with non-specialists; explanations are written to be read by people who aren't engineers.

License

This repository is split by intent, because it holds two different kinds of work:

• Source code, tests, build scripts, and data (the scenario YAML) are under the PolyForm Strict License 1.0.0 - see LICENSE. You may read and run it for noncommercial purposes; you may not redistribute it, modify it, build on it, or use it commercially.
• The written research - all Markdown prose, including this README, FINDINGS.md, ROADMAP.md, the module READMEs, and the REFERENCES.md files - is under Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 (CC BY-NC-ND 4.0) - see LICENSE-DOCS. Read it and cite it with attribution; no commercial use, and no derivative or altered versions.

The bundled frontend build embeds pimas (MIT-licensed) - see THIRD_PARTY_NOTICES. To cite this project, use CITATION.cff.

Contributing

This is a solo research project. The most useful contribution is an issue flagging a sourcing or factual error; code pull requests are generally not accepted. See CONTRIBUTING.md and the code of conduct.

Security

Report security issues privately - see SECURITY.md.