Methodology

The discipline that keeps this project's claims trustworthy, and the limitations that discipline exists to be honest about.

How a claim gets made on this site Five stages in order. The ruleset is frozen, then what would count as a result is written down in advance, then the run happens from a recorded seed, then whatever came out is reported including nothing, and finally the claim is labelled as a measured fact or as an interpretation. A return arrow shows that changing the rules starts a new universe rather than re-running this one, which is what stops tuning until a result appears. Freeze the ruleset is fixed Register what would count, in advance Run from a recorded seed Report including nothing Label fact or interpretation changing the rules starts a NEW universe, it does not re-run this one the order matters more than any single step
The dashed return is the safeguard that does most of the work. Without it, a world can be adjusted until it produces something interesting, and the result then partly reflects the adjustments rather than the rules.

Deterministic, keyed randomness

Every source of randomness in the simulation is derived from an explicit seed through a keyed random number generator, never from an unseeded or system source. Given the same seed and the same code, a run reproduces bit-for-bit. This is what makes a claim about this project's behaviour checkable rather than anecdotal.

State hashes

At any point in a run, the entire authoritative simulation state can be reduced to a single hash. Two runs that should be identical, such as a world resumed from a checkpoint compared against one that ran continuously through the same period, are verified by comparing this hash, not by eyeballing the result.

Observer non-interference

Everything that presents the world (the desktop viewer, the public website you are reading now) reads state; it never writes it. This is not merely a policy, it is tested directly: a run with spectator/observer features active produces an identical state hash to the same run without them. Watching a universe, in any number, from any number of browsers, cannot change what happens inside it.

Preregistration and holdout seeds

Where this project makes a specific empirical claim about behaviour (rather than a claim about what the rules technically allow), the intent and the evaluation criteria are written down before the result is seen, and evaluated against seeds that were not used while developing or tuning the rules. This guards against quietly moving the goalposts after seeing a convenient result.

There is a limit to that worth stating, because a careful reader will find it anyway. The experiment record is append-only, but it lives in a repository this project controls, so on its own it is self-attested: you have our word that the text above a result was written before the result existed. From September 2026 onward, a preregistration is committed on its own and pushed before the run begins, so the timestamp is recorded by a third party rather than by us and the commit hash fixes the exact wording. Campaigns registered before that remain self-attested, and are described that way rather than credited with evidence they do not carry.

Timestamping does not address a different objection, that a run might be repeated until it passes. The defence there is separate: the seed set is fixed in the preregistration, and every seed in it is reported, including the ones that came out badly.

Universe 1's own founding is the worked example, including the part where three candidate worlds all passed and the choice between them was deliberately made to carry no information: why we refused to choose the most interesting universe.

Engineering capability versus natural observation

This project distinguishes two different kinds of claim. An engineering capability demonstration shows that a mechanism works when deliberately exercised, for example that manipulation succeeds under favourable conditions. A natural observation describes what a population actually did, unprompted, over a long run. Both are useful. They are not the same claim, and this site does not present one as the other.

General ecological viability as fitness

Selection pressure in this project rewards staying alive, gathering energy and reproducing under whatever conditions currently apply, not any specific named strategy. No fitness function scores an organism for building, cooperating or exploring.

Frozen rulesets

A universe runs under one fixed ruleset for its entire life, identified by a physics version string that never changes after genesis. If a better mechanic is discovered later, it cannot be patched into a universe that is already running: doing so would mean the universe's history and its physics no longer agree with each other, which would make every earlier observation about it unreliable. Instead, an improved ruleset becomes the seed of a new universe, genesis'd separately, with the old one preserved exactly as it was.

Genesis identity and checkpoints

A universe's identity, starting seed, code version and initial state hash are recorded once, at genesis, and never altered afterwards. Its running state is checkpointed periodically so it can be recovered exactly after a restart, and its major events are recorded durably so its history survives a process restart intact.

Limitations, stated plainly

Claim discipline

This site distinguishes supported fact, qualified interpretation and claims that are currently unsupported, and tries hard not to blur the three. If you notice a claim on this site that overreaches what has actually been observed, that is a defect worth reporting.