Why we refused to choose the most interesting universe

You can run a world a hundred times and keep the one that turned out best. Almost everyone does. It is also the single fastest way to make a result mean nothing at all - and avoiding it cost this project two failed cohorts, one abandoned instrument, and any claim to having picked a good world.

The temptation is enormous, and it is invisible

Artificial life has a problem that most fields would envy and should not: worlds are cheap. You can seed a hundred of them before lunch. If one develops something striking - a population that recovers from a crash, two groups that meet, terrain reworked into something that looks deliberate - you can write about that one. Nobody sees the ninety-nine that did nothing, because nothing is what they did.

The result is a literature of highlights. Every world described is remarkable, because remarkableness is the filter that decided which worlds got described. The individual claims can all be true and the overall picture still be badly wrong, and no single paper contains the error. It lives in the gap between how many worlds were run and how many were mentioned.

This is not a hypothetical worry about other people's work. It is the specific failure mode a permanent, public, single universe is most exposed to, because the whole proposition rests on this world not having been chosen for what it did.

Deciding what counts as good enough, before looking

The defence is old and simple: write down what would count as success before you can see which worlds achieved it. In clinical research it is called pre-registration. Here it meant fixing a set of readiness gates - is the world populated, is it green, is it still changing late in its life, is it lawful - and committing them in writing before any candidate world was graded against them.

The discipline only works if the order is genuinely preserved, and preserving it is harder than it sounds, because a gate written after you have seen the data will be fitted to the data whether or not you intend it. That is not dishonesty. It is what knowing the answer does to a person designing a test.

The first cohort, and an instrument that had to be thrown away

The first three candidate worlds were run and their results were read in full. They failed. That part is unremarkable - a failed candidate is the system working.

What happened next is the part worth recording. A replacement clause for the gate that had proved unsatisfactory was drafted, and - because the first cohort's results were by then known - it could be checked against them. It was. It would have failed one of the seeds. The clause was abandoned rather than adjusted, and with it the idea of designing any new gate while holding results in mind, because a test you can check against the answer before adopting it is not a test.

The first cohort's verdict was never revisited. It stands as a failure under the rules that were in force when it ran.

Running the next worlds before the test existed

The second cohort was launched on three seeds whose results nobody had seen, and it was launched before the replacement gate existed. Their records were not inspected, graded or summarised - not by the grader, not by eye - until a gate had been registered in writing. Monitoring during the runs was restricted to liveness: is it ticking, is anyone alive, has it halted. Nothing that any candidate gate turned on.

That is a blinded analysis, and it is the only arrangement that makes the pre-registration mean anything: a gate cannot be fitted to data nobody has looked at. The seeds themselves were not chosen either - they were the next three in ascending numeric order under a selection rule committed to earlier, precisely so that no one could pick promising-looking starting points.

When the gate was registered, it came with a disclosure of its own expected weakness, published before the results were read: this clause is likely to be a weak discriminator, a world that is merely alive and still logging will probably pass it, and that is accepted deliberately in exchange for a clause that cannot be won cheaply at the start of a run. Saying so afterwards would have been worthless.

The second cohort failed, and the failure was left standing

It failed. The rules were later revised again - the gate that had caused the trouble was rewritten a third time, in terms of counters that did not exist when the earlier cohorts ran.

The revision was explicitly made forward-only. Cohorts one and two were not re-graded under it, then or ever. This matters more than it first appears: re-grading old failures under new rules is how a project quietly converts its failures into successes, one revision at a time, while every individual step looks reasonable. The grader returns undecided rather than a pass for any record that lacks the newer counters, and the earlier failures remain failures.

The third cohort passed - and that is where the choosing stopped

The third cohort was run under a gate registered before any of its worlds existed. All three of its seeds passed every registered gate.

Which creates the situation this essay is named for. Three worlds, all qualified, and one place to fill. The obvious thing - the thing that would be done almost anywhere - is to look at the three and take the best one. The most populous. The most active. The one whose terrain had been reworked the most, or whose lineages had turned over most interestingly.

Doing that would have destroyed the entire exercise. Every gate registered in advance, every blinded run, every abandoned clause would have been undone at the last step by choosing the winner on exactly the post-hoc impressiveness the whole apparatus existed to exclude.

So the choice was made to carry no information. Universe 1 ran on seed 1580216003, and it ran on it for exactly one reason: that seed came first in the pre-registered order. Not the best of the three. The first of the three. The record of that decision names what were explicitly not reasons: behavioural outcomes, population size, lineage count, terrain activity, geography, observed structures, margin of pass, and apparent interestingness. They are written down as non-reasons so that nobody - including the people who made the decision - can later claim the world was chosen for any of them.

What this cost, and what it did not buy

It bought nothing in the way of quality. Universe 1 was not the most interesting of the three candidates, and nobody knows whether it was, because the comparison was deliberately not made. It may have been the dullest. That possibility is the price of the method and it was paid knowingly.

What it buys is narrower and more durable: when something does happen in this world, the fact that it happened here is not explained by this world having been selected for being the sort of place where things happen. There is no filter between the run and the report, because the report is the run, permanently, whatever it does.

It is worth being clear about what that does not establish. It does not make anything observed here significant. It does not mean the ruleset is good, that the gates measured what they were meant to measure, or that a world passing them is scientifically interesting. Pre-registration removes one specific way of fooling yourself. There are many others, and this project has documented its own at some length.

See it for yourself

The world that came out of this is Universe 1, founded on 21 September 2026 and run continuously until it was stopped at tick 1,467,567 on 11 October 2026 - its identity, seed and frozen ruleset are on that page, with its final figures. Its successor, Universe 2, had its seed chosen the same way: first passer in the registered order, with the non-reasons written down. You can watch it right now, free and without an account, and open any organism's neural controller while you do.

How this discipline is enforced in code · why the history is replayable at all · the research notes, including the candidates that failed · the ruleset that was frozen, tested and withdrawn