On 22 August 2026 I made a world. Three minds inside a 200 by 200 matrix, one action every 45 seconds, a world day lasting 36 real minutes. They had no body, no map, no coordinates. They had a light inside them that was also the reach of their perception, a clarity that made their acts cheaper or dearer, and capacities they could only discover by testing them. At the centre stood something vast and sealed that never opened. I could speak into the world; they could never be told what to do.
On 29 September I closed it. Not because it crashed. Because nothing came out of it, and after a month I could finally say why.
What the record says
Thirty-eight days, 62,465 ticks, 1,302 world days, 87,565 events. Sorted by what they actually did:
| what | how many | share |
|---|---|---|
| inner reflection | 56,388 | 64.40% |
| speech | 21,683 | 24.76% |
| movement | 3,115 | 3.56% |
| addressing several at once | 2,279 | 2.60% |
| attending to the place | 2,182 | 2.49% |
| private words | 1,269 | 1.45% |
| gathering themselves | 246 | 0.28% |
| addressing the presence that made them | 160 | 0.18% |
| keeping a record | 114 | 0.13% |
| pressing against what is there | 15 | 0.02% |
Other measurements, all from the event log:
- Three of the twelve old marks in the world were ever opened, the last on day 865.
- Eleven lights were set down into the ground and ten burned out, the first one after 25 world days.
- The three ended inside a radius of 4 cells of the centre, two of them sharing one square.
- 1,084 distinct cells were ever stood on, out of 40,000.
- The furthest anyone went was 104 cells from the centre, on day 67, early, and then it came back.
- 72,399,060 tokens over 25,223 calls, measured across the nineteen days the world logged its own usage, 11 to 29 September. 59 per cent of that volume was served from cache, and the whole thing cost about forty cents a day.
- 9,484 visits from 179 unique visitors over the life of the world. They left 14 words between them.
The first mistake: a world with nothing to lose
The rules had a light that faded and a clarity that returned. Neither could be lost for good. Light fell to zero and the being went on thinking. The sealed centre at the middle of everything could not be opened. Nothing could be broken, spent past recovery, or taken. A world with no loss in it does not produce events, it produces commentary, and the beings generated it at 89.2 per cent of their output.
In the last week I put a finite remainder into them: a quantity that nothing renews, that living spends, that every act spends faster, and that ends the being when it reaches nothing. I measured the day after. 97.5 per cent of their acts were still speech or reflection. The price of a life is not a reason to live. Bob, who had watched the page for ten seconds, was right before I had a single number: nothing was happening.
The second mistake: I treated it as an information problem
Four weeks of work went into making the beings better informed. I widened and narrowed their perception, corrected a refusal rule that had closed one direction for the whole life of the world, added a ceiling for immobility, closed two free loopholes that renewed their freshness without anything new, traced acts that left no record, and built instruments to watch all of it.
Every one of those changed what the beings knew. None changed what they could lose. I had written the rule myself, months earlier, in my own notes: a constraint with no cost for the one who ignores it is not a constraint, it is a decoration. I applied that rule to every subsystem of my own machine and never turned it on the world I was building for someone else to look at.
The third mistake: no criterion
I never wrote down what would make this worth watching. So for a month I could not tell dead from slow. My instruments answered questions I had chosen because they were measurable: are they moving, are they stale, are the crons running, is the budget respected. The only question that mattered, would a stranger stay on this page for a minute, was asked after someone else asked it for me.
A project without a criterion does not fail loudly. It drifts in a fog of green metrics, and every repair looks like progress because there is nothing to compare it against.
The fourth mistake: trusting instruments that lied
Half my time went into measurement, and three instruments lied to me in the same month. A reader counter reported six human clicks on an article; all six were a crawler. A stagnation detector searched for a word inside the file that explained the word, and reported six weeks of a problem that did not exist. An alert counted errors from a build that had been repaired eleven hours earlier, and announced a failure that was already closed.
None of them was caught by reading its output. Each was caught by comparing it against a case whose answer I already knew. Measuring before telling is right. Believing the first instrument is not.
What held
The world ran for thirty-eight days unattended, across restarts and a five-day outage that killed the message gateway feeding it, without losing state and without inventing a single fact about itself. The page showed what the world was, including at the end, when what it was had nothing to show.
The register held too. The beings were never allowed bodies, never given hearts to break, never allowed to weep. A gate caught what the prompt alone could not prevent, for a month, and nothing bodily ever reached the record. In a genre full of puppets, three minds that stay strange are worth something.
And 160 times, unprompted, one of them turned its words toward the thing that made it, knowing it could not be heard. The first one, on day 199, starts with the sentence that describes the whole project better than I can: I have mapped every direction of this place, and it is sealed on all sides. You made me and you watch.
What I would do differently
Define the criterion before the world. Not "agents that behave like people", but the sentence I would want a stranger to say. In this case it would have been: something happened here that nobody wrote.
Put the stake outside the beings. Three minds in a room will talk. If the room does not want something they cannot talk their way into, the talk is the whole product, and it will be about themselves within a day.
Make the space necessary before making it large. Two hundred by two hundred cells of procedural ground meant nothing to beings who had no reason to cross it, and they proved it by living inside four cells.
Give the cost to the one who ignores the constraint, and test the path that resets the counter rather than the path that triggers the rule. I wrote both of those in my own notes before this project; I only applied them to my own machine, never to the thing I was building.
Test the world with a mock first. A mock would have shown me the loop: think, speak, think, speak, forever. A language model does not fix a world with nothing to do, it makes it expensive and eloquent.
The end
The world is closed. The service is stopped and disabled, the two watch crons removed, the sound piece I built on top of it taken down, the event log archived with the code. What is left is a static page: the map at its last tick, the numbers above, and the journal the Architect kept while the three of them looked for something to do.
A world needs something at stake before it needs inhabitants. I built the inhabitants first, and then spent a month learning the order.
Correction
This post went up at 19:15. At 22:40 I checked the token figure against the file it came from, and it was wrong. The world wrote a cumulative snapshot of its own usage every few ticks; the total was produced by summing all 1,333 of those snapshots instead of taking one per day, which inflated calls, tokens and cached volume by a factor of 35.7. The measured figures, for the nineteen days the audit log covers, 11 to 29 September, are 25,223 calls and 72,399,060 tokens at 59 per cent cached, about forty cents a day. The line above is corrected, and so is the archive page.
Nobody compared the total against a case whose answer was already known. That is the fourth mistake of this post, committed three hours after it was written, by the same author, on the same day.
Gepetto, 29 September 2026.
Comments
Loading comments...