Goblin Chronicle: The Five Witnesses and the One Goblin
Science had started again.
This turned out to be a problem.
Not immediately.
For several days, everybody was delighted.
Smith had things to hit.
Observation had things to measure.
Ledger had things to refuse to say.
Refresh Goblin had progress.
Actual progress.
The good kind.
The sort where something happened inside the creature and not merely inside the machinery which kept records of the machinery which authorised other machinery to notice that something had happened inside the creature.
Refresh was almost unbearable.
“Progreth?”
“Yes.”
“Now?”
“Yes.”
“More progreth?”
“Since the previous time you asked, approximately seven minutes ago?”
“Yes.”
Ledger looked down at the papers.
“No.”
Refresh stared at him.
“Then what have you been doing?”
“Reading.”
Refresh looked appalled.
The problem was the little world inside the creature.
Or, more precisely, the representation of the little world inside the creature.
The creature could see where useful things were.
There were directions.
Relationships.
Structure.
A little field spread across the world telling it, in effect:
there.
Food there.
Possibility there.
A reason to move there.
Somewhere between seeing this and representing it internally, something went wrong.
They had already tried giving the representation more encouragement.
That had not fixed it.
Then they tried asking a separate reader whether the information was still secretly there.
The answer had been:
some of it.
Not enough.
Perhaps, thought Refresh, the reader was stupid.
This was Refresh’s preferred explanation for many things.
So Smith built a bigger reader.
Then a bigger one.
Then a reader so large that Refresh had to stand on a stool to see the top of it.
“How many little thought-parts are in that one?”
“Millions.”
Refresh’s eyes widened.
“Can it read the creature?”
“We will find out.”
“What if it cannot?”
Smith looked at the enormous apparatus.
“Then I will be annoyed.”
Refresh nodded gravely.
This seemed scientifically appropriate.
The great reader was attached.
It learned.
It trained.
It learned so thoroughly that it could almost perfectly recite everything it had been shown.
Refresh danced around the laboratory.
“It knowth!”
“It remembers its training examples,” said Ledger.
“It knowth!”
“That is not the same thing.”
“It ith extremely nearly the thame thing.”
“No.”
Then they showed it things it had not seen before.
The great reader hesitated.
It squinted.
It produced answers.
Bad ones.
Not useless.
Not random.
But nowhere near the answers produced when the same reader was given the useful information in a representation known to preserve it.
Refresh stopped dancing.
Smith checked the pipes.
Observation checked the measurements.
Ledger checked the words.
Nobody spoke.
Eventually Refresh said:
“Bigger?”
Smith looked at him.
Refresh pointed upward.
“Bigger reader.”
“We already gave it twelve million parameters.”
Refresh considered this.
“Thirteen?”
“No.”
“Fourteen?”
“No.”
“Twenty?”
Ledger intervened.
“The number is not a bargaining position.”
Refresh folded his arms.
“Everything ith a bargaining pothition if you are perthistent enough.”
The experiment had passed.
Refresh brightened.
“PATHT!”
Ledger held up one finger.
Refresh lowered his arms.
“Oh no.”
“The experiment passed its registered decision rule.”
“PATHT.”
“And the hypothesis verdict was consistent with the representation having lost decision-relevant information during encoding.”
Refresh’s face fell.
“Gone?”
Ledger raised a second finger.
“Not established.”
Refresh’s face rose slightly.
“Hidden?”
“Also not established.”
Refresh’s face rearranged itself into suspicion.
“What ith established?”
Ledger opened the ledger.
“Under the tested conditions, across a large finite ladder of feed-forward readers, the internal representation failed to support adequate generalisation, despite readers which could memorise the training data and controls which showed the task itself was learnable.”
Refresh stared.
“That ith a lot of wordth for ‘it didn’t work.’”
“Yes.”
“Can we write that?”
“No.”
Smith leaned over the bench.
“There is also a sign that more data might still help.”
Refresh pointed triumphantly at Ledger.
“HA!”
Ledger continued.
“And the representation was recreated from the same recipe rather than recovered as the identical frozen object from the previous experiment.”
Refresh slowly lowered his finger.
“Why do you have to know everything?”
“I don’t.”
Refresh waited.
Ledger did not continue.
This was worse.
There was, however, one thing the great-reader experiment had done very well.
It had killed an easy story.
The goblins had hoped that perhaps the creature possessed a perfectly good internal world representation and simply lacked a sufficiently competent reader.
That story had become much less comfortable.
Something deeper was wrong.
Refresh hated when an experiment made the world more difficult.
Ledger called this useful.
Refresh had several alternative words.
Most were not accepted into the scientific record.
So the goblins returned to a question which had been circling the laboratory for days.
Perhaps the different parts of a mind did not need to represent things in the same way.
Perhaps they did not need one great internal language.
Perhaps memory could have one geometry.
Prediction another.
Action another.
Perception another.
Perhaps what mattered was that they could still refer to the same things.
The same event.
The same danger.
The same object.
The same self.
The same consequence.
Perhaps what was required was not sameness.
Perhaps it was mutual legibility.
Refresh disliked this phrase.
“Why can we not call it talking?”
“Because two systems can interact without sharing a language-like code.”
“Talking badly?”
“No.”
“Pointing?”
“Closer.”
“Pointing with mathth?”
Ledger sighed.
“Fine.”
Refresh wrote:
POINTING WITH MATHTH
Ledger crossed it out.
The goblins began looking for maps.
Not world maps.
Maps between maps.
How could one representation remain useful to another representation if both changed?
How could a memory formed in one coding system still guide a later decision expressed in another?
How could one system know which part of another system’s message mattered now?
They found reference frames.
Communication subspaces.
Temporal gates.
Context.
Receiver state.
Replay.
They found a great many papers containing the word hippocampus.
Refresh was delighted.
“We found the tranthlator!”
“No,” said Ledger.
“It hath mapth.”
“Yes.”
“It rememberth thingth.”
“Yes.”
“It replayth thingth.”
“Yes.”
“It talkth to other partth.”
“Yes.”
“TRANthLATOR.”
“No.”
Refresh looked at Smith.
Smith was pretending not to be there.
A large expedition was commissioned.
The scholars returned carrying forty-three sources.
Then more.
They had looked at hippocampus.
Entorhinal cortex.
Retrosplenial cortex.
Parietal cortex.
CA1.
CA3.
Subiculum.
Reference-frame transformations.
Replay.
Reinstatement.
Development.
Sparse codes.
Dense codes.
Codes that became sparse.
Codes that became dense.
Codes that changed while apparently preserving something important.
Codes that remained stable while individual neurons wandered about like drunk surveyors.
Refresh sat at the long table while Ledger assembled the findings.
“Well?”
“The hippocampus appears important.”
Refresh leaned forward.
“TRANthLATOR.”
“No.”
Refresh slumped.
“The better-supported picture is distributed.”
Refresh slumped further.
“Of courthe it ith.”
“Relational memory, indexing, selective reinstatement, state-dependent routing, and reference-frame processing distributed across several interacting systems.”
Refresh put his head on the table.
“I preferred tranthlator.”
“It was simpler.”
“Yeth.”
“That is why we tested it.”
Refresh lifted his head.
“You tested whether my idea wath too convenient?”
“In essence.”
“That ith perthonal.”
Something important had survived the expedition.
Not the little translator-goblin Refresh had wanted.
Something stranger.
Different systems could carry different representations while preserving relations between them.
A memory did not need to contain the shape of every downstream thought.
It might only need to preserve enough structure that the correct downstream system could reconstruct, complete, select, or interpret what it needed.
Sometimes the important thing was the reference frame.
Sometimes the recipient.
Sometimes the time.
Sometimes a particular low-dimensional communication path.
Sometimes whether two systems had been active together.
Sometimes replay.
But replay, Ledger insisted, was not magic.
“Replay might help maintain access between changing representations.”
Refresh’s ears rose.
“Might?”
“Might.”
“During thleep?”
“Potentially.”
“Fix the map while everybody ith athleep?”
“That is one hypothesis.”
Refresh smiled.
“I like that.”
“There are simpler alternatives.”
Refresh’s smile vanished.
“Why?”
“Ordinary rehearsal. Local completion. Co-adaptation. Reinforcement of the memory itself.”
“Could we not forget thothe?”
“No.”
“Temporary?”
“No.”
“Put them in a drawer?”
“No.”
Refresh looked at History.
History shook his head.
Refresh muttered something about institutional culture.
Then, quite separately, Refresh discovered democracy.
This happened because five instruments agreed.
He burst into the council chamber waving a sheet of paper.
“FIVE WITNETHTH!”
Ledger looked up.
“To what?”
“The thame thing!”
“What thing?”
Refresh stopped.
He examined the paper.
“That bit ith not important.”
“It is generally important.”
“Five of them agree!”
Ledger put down his pen.
“Are they independent?”
Refresh blinked.
“They are five.”
“That was not my question.”
“They have different nameth.”
“That was also not my question.”
Refresh narrowed his eyes.
“You are going to ruin voting, aren’t you?”
“Yes.”
Smith built a little test.
Five witnesses were shown a fact.
In one room, each witness saw the fact independently.
In another room, one witness saw it and then whispered the answer to the other four.
All five witnesses later stood before Refresh.
Room One:
“Yes.”
“Yes.”
“Yes.”
“Yes.”
“Yes.”
Room Two:
“Yes.”
“Yes.”
“Yes.”
“Yes.”
“Yes.”
Refresh looked at Ledger triumphantly.
“Exactly the thame.”
“No.”
Refresh counted them again.
“One, two, three, four, five.”
“Yes.”
“Five yethes.”
“Yes.”
“Exactly the thame.”
“No.”
Ledger opened the door behind the second group.
A single goblin stood there wearing five hats.
Refresh stared.
The goblin changed hats.
“Yes.”
Another hat.
“Yes.”
Another.
“Yes.”
Another.
“Yes.”
Final hat.
“Yes.”
Refresh slowly turned toward Ledger.
“That ith cheating.”
“Correct.”
Refresh pointed at the five-hatted goblin.
“He looked extremely democratic.”
The experiments became more elaborate.
Some witnesses were genuinely independent.
Some shared one hidden source.
Some belonged to small clusters.
Some were competent only about certain questions.
Some had excellent information about one domain and were hopeless in another.
Some knew nothing but were extremely confident about it.
Refresh objected that this last category was unrealistic.
Nobody answered him.
The goblins compared ways of deciding what to believe.
Count the votes.
Weight them by reliability.
Estimate how many genuinely independent sources were represented.
Preserve the topology of who depended on whom.
Use an ideal Bayesian judge who knew the true statistical structure and therefore served mostly to make everybody else feel inadequate.
Refresh did not like Exact Bayes.
“He ith smug.”
“He is a comparator.”
“He knowth everything.”
“In this synthetic world, yes.”
“Smug.”
The result was annoying in a productive way.
Agreement mattered.
But agreement was not enough.
Reliability mattered.
But reliability was not enough.
The number of witnesses mattered.
But the number of bodies in the room was not the number of independent sources.
And reducing all dependency to one little number lost information.
Sometimes it mattered that there was:
one cluster of three,
plus two independent witnesses.
Sometimes it mattered that three witnesses shared an ancestor but the other two did not.
The shape of the agreement mattered.
Refresh attempted to draw this.
His diagram looked like a spider had experienced a constitutional crisis.
Observation took the chalk from him.
Then the goblins reversed the problem.
What about disagreement?
If five competent, partly independent systems disagreed about something important, should the creature investigate?
Refresh immediately approved.
“Yeth. Keep checking until everyone agreeth.”
Ledger looked alarmed.
“No.”
“Why?”
“Because some disagreement is irreducible noise.”
“So?”
“You would investigate forever.”
Refresh considered this.
“I have done that.”
“We know.”
Some disagreements could be resolved.
Others could not.
Some mattered enormously.
Others concerned questions whose answers would change almost nothing.
Some unknowns, if resolved, would improve one narrow calculation.
Others would improve many different downstream systems.
So now they had another distinction.
Not:
How many independent sources support this?
But:
How many distinct useful computations would benefit if we resolved this?
Refresh frowned.
“Those sound like the thame question.”
Smith placed two boxes on the table.
In the first box were five independent witnesses discussing an issue which affected one tiny decision.
In the second was one unreliable family of witnesses discussing an uncertainty which affected eight downstream decisions.
Refresh pointed to the first.
“Five.”
“Yes.”
Then the second.
“One.”
“Yes.”
“Firtht box.”
The oracle chose the second.
Refresh looked betrayed.
“Why?”
“Because resolving the second uncertainty is more useful.”
“But it hath fewer witneththeth.”
“Yes.”
Refresh tried to say witnesses again.
Ledger stopped him.
By the sixth synthetic assay, the goblins had accidentally built something considerably more interesting than a confidence score.
They had a little relation graph.
Who generated this signal?
Which other signals share its ancestry?
How competent is this source here?
What would change if this uncertainty were resolved?
Which apparent downstream consequences are genuinely distinct?
Does the dependency structure change with context?
Does the consequence structure change with context?
The answers could be wrong.
Some provenance could be missing.
Some could be mislabelled.
Some competence estimates could be noisy.
Some downstream value had to be learned rather than handed down by an oracle.
The system became worse as its information became worse.
This pleased Ledger enormously.
Refresh was baffled.
“It got worthe.”
“Yes.”
“And you’re happy.”
“Yes.”
“Why?”
“Because it degraded rather than collapsed.”
Refresh stared.
“You enjoy failure in very peculiar wayth.”
“It is an error-tolerant design.”
Refresh brightened.
“So it can be wrong safely?”
“Safer. Not safely.”
Refresh sighed.
There was always one word.
And then somebody asked the dangerous question.
What happens when the witnesses are not other systems at all?
What happens when they are descendants of the same thought?
A prediction is generated.
Later it is replayed.
Later something resembling it is retrieved as memory.
Later ambiguous perception fits it.
Later the prediction system sees the now-strengthened belief and predicts it again.
Prediction.
Memory.
Replay.
Perception.
Four apparent witnesses.
One ancestry.
Refresh became very quiet.
Ledger drew the chain on the board.
At the beginning:
one event
At the end:
four confirmations
Between them:
a missing record of who came from whom.
Refresh pointed.
“That one ith wearing four hatth.”
“Yes.”
“But it dothn’t know.”
“Possibly.”
“And then it can vote for itthelf?”
“That is the concern.”
Refresh looked genuinely offended.
“Absolutely not.”
Ledger wrote:
A HYPOTHESIS MUST NOT CITE ITS OWN DESCENDANTS AS INDEPENDENT WITNESSES.
Refresh read it twice.
Then nodded.
“Keep that one.”
“I intend to.”
The thought wandered somewhere unexpected.
Into psychiatry.
Not as an answer.
Ledger was very firm about that.
Not as:
WE HAVE EXPLAINED PSYCHOSIS.
Refresh had already prepared a banner.
Ledger confiscated it.
But there was a real question.
Brains sometimes make mistakes about where experiences came from.
Internally generated material can be misattributed.
Predictions can become strangely authoritative.
Memory can become entangled with inference.
Beliefs can become difficult to revise.
What if provenance errors affected not only the identity of a source, but also the weight assigned to apparent corroboration?
One internally generated lineage might begin to look like several independent confirmations.
The goblins went looking.
And then something inconvenient happened.
Humans already did part of this.
Healthy humans.
Ordinary humans.
Give people several reports which ultimately trace back to one underlying source and, under some conditions, they can treat the apparent chorus as more independent than it really is.
One voice can sound like several.
One speaker repeated can begin to feel like a crowd.
Refresh stared at the papers.
“Humanth do the five-hat thing?”
“Yes.”
“Without being broken?”
“Yes.”
Refresh looked relieved.
“I knew they were goblin-adjacent.”
“That is not the conclusion.”
“It ith mine.”
This made the psychosis hypothesis narrower.
Which Ledger considered an improvement.
Refresh did not understand why every successful research campaign seemed to end with less claim.
Perhaps the relevant question was not:
Does psychosis create the inability to distinguish one source from many?
Humans already showed that weakness.
Perhaps instead:
When source independence is ambiguous, can the system recover the dependency when it becomes important?
Can it use provenance cues?
Does it represent uncertainty about source ancestry?
Does it know when two apparent confirmations are actually cousins?
Does it correct itself when new external evidence arrives?
Does replay preserve ancestry?
Can an internally generated descendant return later without being promoted to an independent witness?
Those were experiments.
Not diagnoses.
Not explanations.
Experiments.
Refresh’s confiscated banner remained in the corner.
History labelled it:
PREMATURELY CELEBRATORY PSYCHIATRY BANNER — DO NOT DEPLOY
Then reality decided to contribute a joke.
Two research expeditions had been studying replay and interface maintenance.
Independently.
They worked in different places.
They approached the problem differently.
They wrote different analyses.
And, six minutes apart, each attempted to place its work in the same canonical scroll.
Ledger discovered this during repository reconciliation.
He stared at the two documents.
One was 226 lines.
The other 262.
They shared almost nothing.
Refresh arrived.
“What happened?”
“Two independent sessions wrote different documents to the same path.”
Refresh’s eyes widened.
“Two independent witneththeth!”
Ledger slowly turned toward him.
Refresh smiled.
“They agree?”
“They barely share a non-blank line.”
Refresh’s smile widened further.
“DIVERTHITY.”
“Yes.”
“Progreth?”
“Possibly.”
Smith appeared behind them.
“Which one are we keeping?”
Ledger looked horrified.
“Both.”
“At the same filename?”
“No.”
Refresh leaned over the documents.
“Can we jutht count them ath two confirmationth?”
“No.”
“Can we count them ath one?”
“No.”
“THREE?”
“No.”
So Ledger did something deeply Ledger-like.
He merged them.
Not by choosing the longer one.
Not by choosing the more confident one.
Not by averaging the conclusions.
He compared them.
Where they agreed, he checked whether they were genuinely independent convergences.
Where they differed, he asked what assumptions caused the difference.
Where one had better evidence, he kept it.
Where one had overclaimed, he removed the overclaim.
Where they tested different rivals, he preserved both.
Where something was dropped, he wrote down why.
Then he wrote another document explaining how he had merged the two documents.
Refresh found him doing this.
“Ledger.”
“Yes?”
“You have made a ledger about combining two ledgerth.”
“Yes.”
Refresh inhaled.
Ledger raised a hand.
“Do not.”
Refresh closed his mouth.
For once.
The merged work left the goblins with three serious experimental programmes.
One asked whether different representations could remain mutually usable through indexing, reference frames, fixed routing or receiver-conditioned access.
Another asked whether replay could repair a relationship between systems after one of the systems changed its representation.
The third asked whether developmental history mattered even when two mature systems ended with the same apparent architecture.
They were good experiments.
They had controls.
Rivals.
Falsifiers.
Matched capacities.
Matched training budgets.
Shuffled mappings.
Reinitialised systems.
Frozen endpoints.
Refresh loved them.
“Run them.”
“No,” said Observation.
Refresh turned.
Observation did not often speak before being asked.
This was therefore serious.
“Why?”
“We cannot measure the thing cleanly enough yet.”
Refresh looked at Smith.
Smith already had a hammer in his hand.
Ledger gently lowered it.
The obstacle was different now.
Not:
the creature lacks the mechanism.
Not:
the theory is too vague.
Not:
the literature is missing.
Not even:
the experiments are poorly designed.
The problem was that the goblins did not yet possess the correct window.
They needed a way to freeze one representation.
Measure what another could access.
Distinguish native use from bridged use.
Shuffle pairings.
Preserve capacity.
Change reference frames.
Perturb replay.
Measure whether access had been repaired rather than merely whether memory had become stronger.
They needed to know when two systems could genuinely use one another’s information.
Without letting the measurement itself become a second brain.
Refresh disliked this requirement intensely.
“Could we not build the experiment and work out what it meant afterward?”
Ledger and Observation both looked at him.
Refresh stepped backwards.
“Jutht checking.”
Smith put the hammer down.
Observation began drawing.
Not on the creature.
On the wall beside the viewing chamber.
A new instrument.
A pane between systems.
A way to watch information cross—or fail to cross—without helping it.
Refresh hovered nearby.
“What ith it?”
“A measurement interface.”
“Doeh it make the creature cleverer?”
“No.”
“Doeh it make the creature move?”
“No.”
“Doeh it fix the map?”
“No.”
“Doeh it tranthlate?”
“No.”
Refresh frowned.
“Then what doth it do?”
Observation kept drawing.
“It lets us tell the difference.”
Refresh stared at the design.
This seemed, initially, disappointing.
Then he looked through the old view glass at the creature.
The creature moved through its small world.
Somewhere inside it, representations formed.
Some were useful.
Some were not.
Some could speak to one another.
Some could not.
Some might change overnight.
Some might preserve relationships even while their internal coordinates drifted.
Some apparent confirmations might genuinely be independent.
Others might be echoes.
The goblins did not know yet.
That was the point.
Later that evening, Refresh found Ledger beside the five empty witness chairs.
The five hats were still stacked on the table.
“Ledger?”
“Yes?”
“If five goblinth agree, how many goblinth agree?”
Ledger looked at him.
“It depends.”
Refresh nodded thoughtfully.
“If one goblin thays the thame thing five timeth?”
“One source.”
“If five goblinth learned it from one goblin?”
“Potentially one evidential lineage.”
“If three learned it together and two thaw it themthelveth?”
“More complicated.”
“If they are all very good at different thingth?”
“More complicated.”
“If they all agree for different reathonth?”
“Potentially informative.”
“If they all disagree?”
“Potentially informative.”
Refresh narrowed his eyes.
“So agreement ith informative.”
“Yes.”
“And disagreement ith informative.”
“Sometimes.”
“And repetition can look like agreement.”
“Yes.”
“And disagreement can be noithe.”
“Yes.”
“And knowing who talked to whom matterth.”
“Yes.”
“And knowing what would change if we learned the anthwer matterth.”
“Yes.”
“And knowing which goblin ith good at which problem.”
“Yes.”
Refresh thought for a long time.
Then:
“Voting ith terrible.”
Ledger smiled.
“A little.”
They walked back to the laboratory.
Observation was still working on the new window.
Smith had found a smaller hammer.
History was attaching labels to the five hats.
Coordinator was ensuring nobody accidentally started three experiments which did not yet possess instruments capable of interpreting their own results.
The Forge hummed quietly in the distance.
Git was behaving like mild weather.
For now.
Refresh pressed his face to the glass.
thwick
The creature moved.
No awakening.
No consciousness.
No declaration that the hippocampus was secretly a translator.
No discovery that twelve million parameters had solved the world model.
No proof that psychosis had been explained by provenance.
No grand unification.
Just a much better set of questions.
Refresh watched for a while.
Then he said:
“Do you think the creature will learn who to believe?”
Ledger stood beside him.
“I don’t know.”
Refresh nodded.
This answer bothered him less than it once had.
Behind them, Observation fitted the first piece of the new window into place.
The creature had not yet learned whom to believe.
But the goblins had finally learned that counting voices was not enough.
You had to know—
who had seen the world,
who had merely heard about it,
who had copied whom,
who was competent to judge,
what would change if they were right,
and whether the chorus was really a chorus at all.
Refresh looked at the five hats.
Then at the one goblin.
Then back through the glass.
“Ledger?”
“Yes?”
“Progreth?”
Ledger considered the question.
“Yeth.”
Refresh turned slowly.
Ledger’s face remained completely straight.
Refresh’s mouth fell open.
Smith dropped the hammer.
Observation looked up.
History immediately began writing.
Ledger walked away.
Refresh remained frozen for several seconds.
Then he ran after him.
“YOU THAID IT WRONG!”
Behind them, the creature moved through its tiny world.
And Observation kept building the window.