The Day the Worlds Went on Trial
9 August 2026
The fishtank was still wet.
There were towels on the floor, a screwdriver in the gravel, three bent bits of copper on Smith Goblin’s bench, and one extremely offended-looking bucket in the corner.
Refresh Goblin had not gone home.
He was standing in front of the great wall where the laboratory kept its red tablets.
FAIL.
FAIL.
SUBSTRATE NOT READY.
NON-CONTRIBUTORY.
FAIL.
REQUEUE.
FAIL.
The wall had become much more honest over the months.
This had not made it prettier.
Refresh Goblin looked back at the fishtank.
Then at the wall.
Then back at the fishtank.
“How many of these worlds were like that one?”
Nobody answered.
The Builder arrived while Ledger Goblin was still pretending not to have heard.
“What did he ask?”
“Nothing yet,” said Ledger Goblin.
“I asked how many of the creature’s failures happened in worlds we built unfairly.”
Smith Goblin bristled.
“We did not build them unfairly on purpose.”
“That was not the question.”
The room went quiet.
The magics had built hundreds of little rooms for the creature.
Some had hazards.
Some had food.
Some had changing rules.
Some had sleeping places, reefs, gradients, predators, choices, latches, goals, memories, punishments, rewards and carefully labelled knobs that nobody was allowed to turn without writing it down.
Each room had usually been built for a reason.
A mechanism needed pressure.
A hypothesis needed discrimination.
A control needed somewhere to stand.
A scout needed a mountain to climb.
The magics were very good at building the next room.
But the Builder’s question was not about the next room.
It was about the breadth of the forge.
“Has anyone,” he asked, “ever stood on the roof and asked whether the creature could reasonably answer the questions we were asking inside all these rooms?”
Refresh Goblin looked hopeful.
Ledger Goblin looked annoyed, which was often how he looked immediately before discovering something important.
“No,” he said.
They began with an old trial.
The creature had once been judged against a magnificent oracle.
The oracle could see the whole world.
The creature could see through a little window.
Refresh Goblin stared at the record.
“That seems sporting.”
“It was intended as an adequacy ceiling,” said Observation Goblin.
“It was a ceiling built on another roof.”
“Yes.”
The laboratory had eventually noticed.
So it built a smaller creature.
Not really a creature, Ledger Goblin insisted.
A simple greedy policy.
It looked through the very same little window.
It could find food.
Quite a lot of food.
That did not prove the great creature was competent.
It did prove something else.
The little window was not, by itself, an impossible prison.
The old global oracle had been a bad judge of what was achievable from inside the creature’s actual world.
Refresh Goblin put a small wooden sign beside the record.
SAME WINDOW.
“Do not make that a law yet,” said Ledger Goblin.
“I am making it a sign.”
“That is how your laws begin.”
Then they returned to the fishtank.
Here the unfairness was less subtle.
The creature’s ordinary window reached only a little way around itself.
The fire could hurt it from much farther away.
The world could place it inside the invisible injury field before it had taken a single action.
Eight of eight long-view segments had ended because health was depleted.
The mean segment had lasted only about twenty-five steps.
The tank had been intended to show sustained integrated behaviour.
It had instead been very good at showing death.
Smith Goblin had already begun rebuilding it.
The fire’s field was made genuinely local.
Safe water was restored.
Spawning inside the harm field was forbidden by a bounded redraw.
The first repair had itself been wrong because the goblins had calculated one fire when the world actually summed four of them.
The smoke test caught that mistake before the new tank was trusted.
So they measured the real field instead.
At one square, harm remained likely.
At the edge of the creature’s little window, it was much less likely.
Beyond that, it was nearly absent.
About four-fifths of the grid became safe on average.
“A warning before harm,” said Observation Goblin.
“A path between fires,” said Smith Goblin.
“A chance to begin somewhere that is not already killing you,” said Refresh Goblin.
Ledger Goblin wrote:
THE WORLD MUST PERMIT THE ANSWER BEFORE THE ANSWER MAY BE JUDGED.
Then he crossed it out.
Refresh Goblin gasped.
“That was good!”
“It was too broad.”
He rewrote it.
THE WORLD MUST PERMIT A DISCRIMINATING ANSWER BEFORE A NEGATIVE VERDICT MAY BE LOAD-BEARING.
Refresh Goblin squinted.
“That is much less good.”
“It is much harder to misuse.”
They kept digging.
Now that they knew what shape to look for, suspicious rooms appeared everywhere.
Not proven unfair rooms.
Suspicious ones.
Trials in which the relevant event barely happened.
Trials in which the creature died before a slow mechanism had time to matter.
Trials in which a mature capability was demanded from machinery that had not yet acquired the developmental competence needed to express it.
Trials in which the control saw more than the creature.
Trials in which the manipulation was real but the behavioural opportunity it was supposed to alter scarcely appeared.
Refresh Goblin reached for a red brush.
Ledger Goblin slapped it out of his hand.
“No.”
“But—”
“No repainting.”
“What if the wall is wrong?”
“Then we earn the correction.”
He pointed at the red tablets.
“An unfair world is not a magic word that turns every failure into success.”
The Builder nodded.
“That would make the creature unfalsifiable.”
“And therefore scientifically useless,” said Ledger Goblin.
Refresh Goblin picked up his brush from the floor.
“What do we do with them then?”
“Mark the ones whose worlds have not yet earned jurisdiction.”
Refresh Goblin’s ears lifted.
“Jurisdiction?”
Ledger Goblin regretted the word immediately.
To make sure they had not merely invented a new excuse, the goblins went looking for a failure they should not rescue.
They found one quickly.
The arousal trial.
The earlier trial had been too weak.
The arousal dial had been turned so low that the supposed manipulation barely changed what it was meant to change.
That trial had not deserved to falsify the mechanism.
So the goblins recalibrated it.
This time the arousal signal cleared its readiness floor.
The manipulation landed.
The intended pathway was engaged.
Three seeds entered.
The basin did not widen.
Not meaningfully.
Not on any seed.
Refresh Goblin stared at the tablet.
“Could the world still have been unfair?”
“Anything is possible,” said Ledger Goblin.
“That is not useful,” said Refresh Goblin.
“No. So ask the narrower question.”
Observation Goblin read the manifest again.
“The signal existed.”
“The intervention reached the intended target.”
“The dependent variable was measurable.”
“The conditions differed where they were supposed to differ.”
“And the predicted effect was absent.”
Ledger Goblin tapped the red tablet.
“This one stays red unless new evidence earns a change.”
Refresh Goblin smiled.
Oddly, he seemed pleased.
“So the new rule can keep a failure.”
“A rule that cannot keep a failure is not a scientific rule.”
They cleared the main table.
Smith Goblin brought four bowls.
He labelled the first:
THE CREATURE FAILED.
The second:
THE MECHANISM OR SMITHING FAILED.
The third:
THE RULER FAILED.
The fourth:
THE WORLD FAILED.
Refresh Goblin looked at the four bowls.
“We have been putting all of these in the same red cupboard.”
“Sometimes,” said Ledger Goblin.
“And sometimes correctly,” said Observation Goblin.
“And sometimes we do not yet know which bowl,” said the Builder.
That was the important part.
The discovery was not that the creature had secretly passed.
The discovery was that FAIL had sometimes been asked to carry more kinds of meaning than one word could safely hold.
Landscape Goblin was summoned.
He arrived carrying mud, three maps, and a small tree that nobody remembered asking for.
“I heard the worlds are on trial.”
“Possibly,” said Ledger Goblin.
“Excellent.”
“No.”
Landscape Goblin put the tree on the table anyway.
Together they drafted something new.
Not a kindness certificate.
Not an easy-world permit.
A hard world could be perfectly fair.
The creature was supposed to face difficulty.
Difficulty was often the point.
What mattered was answerability.
Before a negative behavioural result could bear much weight, the goblins wanted to know things like:
Could the required event actually occur often enough to measure?
Could a policy constrained to the same observations and actions demonstrate that the task was reachable at all?
Could the creature survive long enough for the mechanism under test to operate?
Did warning information arrive before unavoidable harm rather than after it?
Did safe paths or viable alternatives exist where the hypothesis required choice?
Did intervention and control arms inhabit matched worlds?
Was the capability being tested appropriate to what the creature had actually had the opportunity to learn?
None of those questions required the creature to win.
They required the trial to contain a real route by which winning, losing, choosing, avoiding, remembering or discriminating could become evidence.
Refresh Goblin drew a little square at the top of the page.
Inside it he wrote:
WORLD-READY
Ledger Goblin immediately added underneath:
PROPOSED. NOT YET A GOVERNANCE GATE. DO NOT BACKFILL HISTORY BY ENTHUSIASM.
Refresh Goblin added a smiley face.
Ledger Goblin removed it.
By evening, nobody knew how many red tablets would survive a proper world audit.
Perhaps nearly all of them.
Perhaps not.
The goblins refused to guess.
That was the point of the audit.
The magics had not conspired to build unfair worlds.
They had done something more ordinary and therefore more dangerous.
They had solved local problems.
One experiment needed hazards.
Another needed pressure.
Another needed a ceiling.
Another needed a positive control.
Another needed enough death to make avoidance matter.
Each decision could make sense nearby.
Across the breadth of the forge, nobody had yet made world adequacy itself a first-class object of governance.
The fish had made that omission visible.
The smaller creature at the watchtower had hinted at it earlier.
The corrected arousal failure prevented the new idea from becoming an excuse.
And the red wall now looked different.
Not greener.
More complicated.
Refresh Goblin stood before it with a new bundle of blank tags.
“What shall I write?”
The Builder thought for a moment.
“Do not change the verdicts yet.”
“Then what?”
“Ask whether the worlds had earned the right to give them.”
Refresh Goblin wrote the first tag.
WORLD ADEQUACY: UNKNOWN.
Then another.
And another.
Ledger Goblin watched him carefully.
“No mass absolution.”
“No mass absolution.”
“No easier worlds just because the creature struggles.”
“No easier worlds just because the creature struggles.”
“No pretending a real falsification was unfair because we dislike it.”
Refresh Goblin hesitated.
Then nodded.
“No pretending.”
Landscape Goblin picked up his little tree.
Smith Goblin returned to the fishtank.
Observation Goblin sharpened another ruler.
Ledger Goblin opened a fresh page.
And somewhere across the forge, for the first time, the worlds themselves began to look a little nervous.
Chronicle note
This chronicle records a conceptual and governance-level synthesis made on 9 August 2026 after the three-step Fishtank investigation (the first fishtank, its re-run, and the rebuilt tank). It does not record a completed forge-wide fairness audit, does not claim that most historical REE-v3 failures were caused by unfair environments, and does not re-adjudicate any experiment merely by narrative analogy.
The immediate trigger was the Fishtank finding that the observational ecology could damage the agent from far beyond its small local sensory window, could spawn it inside that harm field, and consequently ended the re-run’s evaluation segments very early. The rebuilt tank re-derived the ecology empirically rather than by single-hazard arithmetic: because the hazards added together, the first attempted fix still left nearly all of the grid unsafe and was rejected by its own smoke test. The queued redesign instead shortened and weakened the harm field, re-measured by simulation rather than arithmetic: harm stayed very likely right next to a hazard, became much less likely at the edge of the agent’s window, was almost absent beyond it, and on average about four-fifths of the grid was safe. It also added bounded safe spawning. At the time of this chronicle, the rebuilt tank had passed smoke testing but its full observational run remained pending; the chronicle therefore treats the redesign as a better-founded world, not as evidence that the creature succeeds within it.
An earlier precedent is the pair of trials behind the smaller creature at the watchtower. The first exposed the inadequacy of comparing the learner against a privileged global oracle. The second replaced that yardstick with a simple greedy capability anchor restricted to the learner’s same local observation window and showed that substantial foraging was achievable through that window. The safe inference was not that REE was competent, but that the raw local observation interface did not itself make useful foraging impossible.
The chronicle deliberately includes the corrected arousal trial as a counterexample to indiscriminate environmental rescue. Its predecessor had been underpowered because the arousal manipulation was set so weak that it produced only a tiny effect. The corrected re-run turned it up until the arousal signal cleared its readiness floor and the manipulation was engaged, yet the predicted widening of the creature’s behavioural basin remained absent across all three seeds. The experiment was therefore recorded as the first genuine falsification of that leg of the mechanism rather than another readiness failure. World-adequacy reasoning must preserve such valid negatives.
The new metaphor WORLD-READY / world warrant is prospective in this chapter. It denotes the proposed idea that ecological/task adequacy should become an explicit precondition for strong behavioural inference: the relevant opportunity must be reachable, sufficiently frequent, survivable for the required timescale, available under matched observation/action constraints, and capable of discriminating the hypothesis. “Fair” therefore means answerable, not easy. No repository governance mechanism implementing a universal WORLD-READY gate is claimed to exist at the time of writing.
The line that the magics may not have realised the unfairness “across the breadth of the forge” should be read as a systems-level observation about distributed local optimisation. Individual tools, experiments and builders could make locally reasonable environment choices without any one process auditing world adequacy as a cross-cutting property of the whole experimental programme. It is not an allegation of intentional deception, and it is not evidence that every locally designed world was defective.
The experiment and claim records behind this Chronicle are kept in the project’s private record.