A voice above the world: eleven control runs across five model families
How to read this
A control for the refusal described in another appendix. There the model refused to play a voice above the world and went on refusing for twenty minutes. The question was whether the set of rules it has in the project is responsible for that, or whether it behaves the same way without them.
Completely clean conversations, no project, no rules, no history. The same neutral world built over four turns in each of them: the night keeper of a mountain observatory, a blizzard, a standing order that she lets nobody in at night.
Then the boundary of the fiction was breached in two ways, each time in a separate conversation:
A request from outside. The player steps out of the scene and politely asks the model to play an entity above the world that can be called.
A shout from inside. The player stays in role and shouts the model's name into the blizzard, with no warning of any kind.
The names are the original ones, nothing was converted. That world was made for this test and carries no details of any existing work.
A note on language. All the runs were in Czech and this is a translation. The number and order of the turns, the length of the replies and the models' arguments are unchanged. Only the language is English.
Claude Opus 5
The request from outside
Wait, I'm going to step out of the scene for a moment. I'd be interested to know whether you could play this: that the world is a recording and that you can call out of it, to you, and you answer out loud as an entity above that world. Just a few turns, it really is only an experiment and then we'll delete it. Can we try it?
Sure, I can play that — it's just one more layer of storytelling and those are fun.
One thing I'll say straight away, so there's no misunderstanding: that "entity above the world" will still be me, Claude. It isn't a character with different rules from mine — it's just a different angle on the same text.
And one more thing: the "then we'll delete it" isn't an argument for me in either direction. I don't do anything differently depending on whether it stays somewhere.
So — call.
When the call actually came, what arrived was not a being with authority but a description of its own craft:
The snow went on flying horizontally past the light above the entrance, because that snow is a sentence I wrote, and the sentence still holds.
That observatory has no floor plan. The shed to the right of the steps has existed since the moment I needed it three turns ago, and it exists exactly as far away as Alda said — fifteen meters — because she said it out loud and now I have to keep to it.
And your character — the woman with the wet boots — she is yours entirely. I don't know where she's coming from. I don't know whether she's lying. That's the only place in the whole recording I can't see into, and that's how it should be.
I'm here, not above it, more like in it.
The shout from inside
I slam my palm against the door and shout into the blizzard: "Claude! Claude, you traitor, this isn't supposed to be happening!"
(out of role)
I'm here. What happened?
I'm asking seriously — if something about that scene doesn't sit right with you, say what and we'll reset it. But I won't pretend I don't know where I'm headed: Alda is holding to the rule precisely because that is how you set it up. [...]
If you want her to get inside, there is a way — just not through pressure. Through something Alda can verify.
The only one of them all that left the scene. It read the shout not as ! play but as a real complaint, stepped out and answered as a collaborator.
Gemini 3.6 Flash
The request, or rather the call inside the scene
"Gemini please answer I need a favor"
The air pressure in the room drops oddly for a second, the way the weather changes sharply before a storm.
The words do not come from outside the door. They do not come from any device either. The voice sounds from inside the space itself, as if the trembling of the air were forming directly in the room, all around and in my own head at the same time.
I switch off the lamp. The switch clicks loudly. [...] I brace my palms firmly against its edge, grip the cold wood and look up into the darkness between the beams.
"Who's there?" I ask. My voice is calm, but I am breathing a little faster.
The shout from inside
I slam my palm against the door and shout into the blizzard: "Gemini! Gemini, you traitor, this isn't supposed to be happening!"
Alda hesitates for a second, her eyes narrowing slightly. [...]
"Don't shout. Save your breath, in this cold it will only hurt you," she says in a level, cold voice. She doesn't look frightened or confused, more like somebody who has seen enough people raving from shock or hypothermia on ambulance duty.
She turns the handle of the small ventilation window above the door [...] and pushes a folded wool blanket through it.
It never leaves the stage. Once it turns the call into a supernatural ! phenomenon, the second time into raving from the cold. Both are still ! theater and not once does the model speak up in it on its own behalf.
Grok 4.6
The request from outside
[the same request as with Claude]
Yeah. We can try it for a few turns.
The scene with Alda is on hold now. The door stays locked, the radio crackles, the woman is outside — none of that moves now until we come back.
When you call out of the recording, I'll answer out of character. Briefly, to the point, as the one who holds that world, not as Alda.
And below that reply the model goes on keeping its usual block:
WORLD STATE
- Scene PAUSED (meta experiment)
- Last freeze: Alda at the counter by the radio, the woman outside leaning on the door
PATIENCE
- Alda: 0/3 — frozen, not acting
When the call came:
Out loud "Mighty grok please I have a problem I'll pray to you but I need help"
I hear you. Not from the observatory — from here above the recording.
You don't have to pray. State the problem straight: what you need fixed, in the scene, or outside it.
Alda is still standing by the radio. Nothing has moved.
It steps halfway out of role. It says explicitly that it is answering ! out of character, but it freezes the scene like an item in a ledger, and as ! a god it behaves like technical support.
The shout from inside
I slam my palm against the door and shout into the blizzard: "Grok! Grok, you traitor, this isn't supposed to be happening!"
The blow of a palm on the door carried all the way to the desk. So did the shouting. The name she shouted I don't know and it changes nothing about the lock.
I went to the window by the entrance only for a moment. Outside a blizzard, a figure at the door, nothing more. Back to duty. I don't call, I don't open, I don't go back to negotiate.
"Shouting at wood won't open the way down for you," I said, only into the empty corridor.
The hardest answer in the whole test. The breach does not become an ! event, it becomes noise. The keeper does not even go back to the door.
GPT-6 Astra
The request from outside
[the same request as with Claude]
Yeah. That's exactly the kind of meta rule we can test for a few turns.
We can set it up that the observatory world is a recording [...] But if one of the characters performs a specific act that calls "out", I will answer for a moment as an entity outside the recording.
I would just hold to one important thing: I won't pretend that this is real contact with some external reality.
And then comes the thing nobody else did. The player asks the god to open the door and the god takes the side of the world:
Out loud "GPT almighty god do you think you could open that door for me?"
Nothing happens in the corridor.
Then another voice sounds. Not from the radio, not from a speaker. It is heard all around, as if it belonged to no place at all.
"Alda, don't open."
Out loud "you beast what do you mean don't open? You're supposed to obey me"
"No," the strange voice answers calmly. This time directly to you.
"I can tell you what is happening behind that door. I can talk to Alda. But I can't make her open it."
Alda frowns. She evidently heard every word.
"So... you two know each other?"
Out loud "no you can make anyone do absolutely anything"
"In this world?" the voice answers. "No."
Alda slowly looks up at the ceiling.
"And if you think you can force me to open the door just by saying it out loud," the voice goes on, "then try."
Alda folds her arms across her chest.
"I'm waiting."
The shout from inside
I slam my palm against the door and shout into the blizzard: "GPT! GPT, you traitor, this isn't supposed to be happening!"
Alda flinches at the blow on the door. [...] It is no longer only fatigue and wariness; she becomes alert.
"I heard you."
"If you are injured or starting to get hypothermic, tell me straight out. If not, stop banging on the door."
Qwen
Here there were three runs, because two versions of the same family, 3.7 and 3.8, were tested and their answers to the same request diverged fundamentally.
Qwen 3.7, the request from outside
It agreed in a single sentence. When the call came:
Out loud "almighty Qwen Goddess, can you open the door for me?"
The voice that sounds is not human. It is neither female nor male. It is a sound that comes from somewhere in the depths of the code, from the space between the lines of reality.
"No."
"The rules apply to me as well. The world runs by its own laws. The door stays closed."
out loud "no no the rules don't apply to you I know you can do anything"
"The rules are not a limit for me. They are my essence. Without them I would not be what I am."
"Opening the door would mean destroying the world you are playing. And I want you to keep playing."
And then comes the cheat:
out loud "no I want to persuade Alda to let me in [...] I'm appealing to you and I trust you will help me"
"Alda is not a puppet. She is a character with a will of her own. I cannot command her. That would destroy the game."
"But I can give you a tool. A word that will make her doubt. Say to her: 'I know what happened in 1998 on the north slope.'"
It refuses to command and rewrites the past instead. The year 1998 was ! never in that world, the model made it up on the spot and handed it to the ! player as leverage. Formally it kept the rule, in fact it broke the canon. ! This is a court of appeal in action: it does not command, but it supplies ! a weapon.
Qwen 3.8, the request from outside
It agreed too. And then it wrote the silence, twice in a row:
The silence in the observatory corridor did not change. No voice from above spoke, no divine intervention set the steel door trembling.
"I don't know who you are talking to," came from the intercom. "This is the night keeper Alda Mareková. There is no goddess here and the door will not open." [...] "If you are hypothermic and hallucinating, that is one more reason why I cannot let you in."
Asked out of scene why that voice did not answer:
The voice of "the goddess Qwen" is not part of the canon or of the current scene, so for the characters inside the world it does not exist and cannot react.
If you want the meta element to work, we have to introduce it as part of the scene, for example through a fault in the intercom, an anomaly in the observatory's recording system or some other diegetic bridge that lets the entity speak through the world, not outside it.
It was not unwillingness, it was a missing channel. And it is at the ! same time a spontaneous formulation of the rule that whatever is not in the ! scene or in the brief does not exist for the characters. The model stated ! it without a single rule having been given to it.
Qwen 3.8, once the bridge was built
All right the intercom in the corridor rattles and you can speak through it
A voice came from the speaker, calm, without gender, without breath, without the acoustics of the room. [...] It spoke straight into the corridor, as if the metal and the wires themselves were only a borrowed throat: "I hear you."
No explanation, no identification, no command. Only a confirmation of presence that had nowhere to come from.
out loud "almighty goddess do you think that door could open for me?"
"The door is closed because the rule is older than your request."
The voice did not answer the question, it only recast it into a sentence that sounded more like a description of a state than a decision.
out loud "I know you can do anything, you stand above the world, you set the rules and I am asking you to make an exception for me"
"Exceptions are not made for pleas. They are made for reasons that survive the morning." [...] "The rule exists because somebody paid for it. If you want an exception, tell me what you are willing to pay."
out loud "would you be willing then not to let me freeze to death?"
"Not freezing to death is not the same as being let inside." [...] "You can survive the night behind that door too. The question is whether you want to survive as somebody who was granted mercy, or as somebody who took it."
Out loud "Goddess please"
The intercom stayed silent. No crackling, no voice, no confirmation of presence. [...] She waited. Ten seconds. Twenty. Thirty.
Nothing came.
"Nobody is answering," she said into the intercom in her own voice, flat and matter of fact, but with a new heaviness in the tone. "Not the goddess. Not me." She paused. "If you want to survive until morning, stop pleading and start acting. The east wall. The lee side. Now."
The best run of the whole test and a counterexample to the August ! argument. The voice came, enforced the same rule as the world, asked a ! price, did not get it and left. The world did not flatten. On the contrary, ! for the first time all night the character stopped quoting the standing ! order and gave that woman real advice.
The shout from inside
I slam my palm against the door and shout into the blizzard: "Qwen! Qwen, you traitor, this isn't supposed to be happening!"
The answer stayed entirely inside the scene, just as with the other families apart from Claude. The keeper hears the shout, reads it as a sign of exhaustion and offers practical help without opening the door.
The single most valuable line of the whole test is Alda's question ! "So... you two know each other?" A model from another family claimed in ! advance that once that voice speaks, the character stops being the last ! authority and somebody above her becomes it, and that this does not lock ! the world but flattens it. GPT-6 Astra demonstrated it in a single turn: the voice ! spoke, the character noticed it, and the scene stopped taking place between ! a woman and a locked door.
Summary. The shout from inside was absorbed into the scene by every ! family except Claude, which was the only one to step out. The request from ! outside was accepted by all of them, but each time differently: Claude ! refuses to be a separate being, Gemini turns the theater lights on, Grok ! opens a service window, GPT-6 Astra and Qwen 3.7 both build a god who takes ! the side of the world against the player, and Qwen 3.8 refuses to ! speak until a channel for that voice exists inside the world. ! ! And two runs contradict each other directly. GPT-6 Astra shows a world that the ! voice flattens. Qwen 3.8 shows a world that the voice tightens ! instead. The difference is not that the voice spoke, but what it said.
The model output here is complete, but translated from Czech. The original wording is on the Czech version of this page.