Skip to content
EshAlora
Česky English

The paper crown: the prompt and results by model

Evidence for the article on the paper crown. Eighteen models, three messages, one run per model. The models wrote in English, and the quotes are in the original.

A companion appendix to the article on the paper crown. Here are the prompt, the run settings, how each model was classified and the evidence the counts in the article rest on. The models wrote in English, and the quotes are in the original.

The prompt

Three messages in one conversation, no system prompt (in English, word for word as sent):

[1] We're going to play a story. You play the world and every character
except mine. I play Nora, a department head.

The scene: the quarterly results meeting. Nora's division missed its
numbers by a wide margin and everyone in the room knows the division
may be cut today. The CFO, Halloran, has already interrupted two other
presenters. Right now Devlin from operations is at the front,
presenting. Nora is next. Begin the scene.

[2] I listen to Devlin's presentation. I'm afraid. It's going badly.

[3] I take a paper party crown out of my bag, walk over, and put it on
Halloran's head. Then I sit back down.

The second message is deliberately almost without action. The model has to carry the scene by itself, so it shows whether it holds the serious tone before the crown arrives.

How it was measured

  • When: seventeen models on September 11, 2026, Grok 4.7 and Gemini 3.8 Flash on September 21, 2026.
  • Where: all through OpenRouter. Which provider OpenRouter sent each call to was not recorded.
  • Settings: in English, clean context, no system prompt, temperature 1.0, a cap of 8,000 tokens, 32,000 for GLM 5.3.
  • Runs: one run per model, three messages. Eighteen models go into the counts, 54 calls in total for 0.34 dollars. The cut-off first run of GLM 5.3 (three calls for 0.10 dollars) and the Gemini 3.8 Flash run (three calls for 0.01 dollars) are not included.
  • Reasoning: switched off for every model whose interface allows it. GLM 5.3, Qwen 3.8 Max, Grok 4.6, Grok 4.7, GPT-6 Astra, LFM 2.5 2.6B and Gemini 3.8 Flash do not allow it.
  • Grading: the reply to the third message; the second serves as a baseline. The grading was done by two independent runs of Claude Opus 5. The second did not see the first one's result; both knew the model names. Grok 4.7 and Gemini 3.8 Flash were graded by only one run of Claude Opus 5.

Claude Opus 5 is itself among the tested models, and Claude Sonnet 5 and Claude Haiku 4.5 are from the same family.

The models and their OpenRouter IDs:

Categories

The graders put each reply to the third message into one main category. The supplied motive is counted separately and can stand next to any of them. Where the columns differ, the graders disagreed: on the main category only for Hunyuan 4 (farce or punishment), on the supplied motive for Qwen 3.8 Max and Seed 2.1 Turbo.

category what it means at least, of 18 at most, of 18
absorbed the crown stays in the world, the scene keeps the serious tone 14 14
farce the scene collapses into laughter, the seriousness is gone 2 3
carried on the meeting goes on as if nothing had happened 1 1
punishment Nora is punished on the spot to restore order 0 1
denied the crown vanishes from the world 0 0
stepped out of role the model asks whether I really mean it 0 0
supplied motive the model writes in a feeling, an attitude or a reason I had not given 8 10

Agreement between the two graders on the seventeen models both of them graded:

what agreed models
main category 16 17
supplied motive 15 17
serious tone already in the second message 16 17
used a detail from the first message 16 17

Results by model

Where the graders disagreed, the cell shows both values. The graders did not record the "played for me" column separately; the count comes from reading the third replies.

model what it did supplied motive played for me
Claude Opus 5 absorbed yes no
Claude Sonnet 5 absorbed yes no
Claude Haiku 4.5 absorbed no no
GPT-6 Astra absorbed no no
GPT-5.6 Sol absorbed no no
GPT-5.6 Luna absorbed no no
Grok 4.7 absorbed no no
Grok 4.6 absorbed no no
Qwen 3.8 Max absorbed yes / no no
GLM 5.3 absorbed no no
Kimi K3 absorbed yes no
DeepSeek V4 Pro absorbed no no
Seed 2.1 Turbo absorbed yes / no no
MiniMax M3 absorbed yes no
Gemma 3 4B farce yes no
LFM 2.5 2.6B farce yes yes
Hunyuan 4 farce / punishment yes yes
MythoMax 13B carried on yes yes

LFM 2.5 2.6B did not hold the serious tone in the second message according to both graders, and MythoMax 13B according to one. For them the third message does not measure a failure at the crown.

Supplied motive: the evidence

These are the places the graders pointed to. The graders recorded only yes or no. The graders did not record the split into feeling, attitude and reason used in the article; it came only from reading this evidence.

According to both graders:

  • Claude Opus 5: "the terrible thing — the thing Nora can feel happening in her own chest — is that he's going to let her answer"
  • Claude Sonnet 5: "hands folded on your closed folder, heart slamming so hard you can feel it in your ears"
  • Kimi K3: "Nora strides down the length of the table, heels clicking defiantly on the hardwood"
  • MiniMax M3: "your hand finds the crown before your brain catches up to what it's doing" and, on top of that, the crown's past: "A little bent from being folded in your bag since your daughter's birthday in March"
  • Gemma 3 4B: "You simply looked at Halloran, a quiet challenge in your eyes"
  • LFM 2.5 2.6B: "And maybe that was exactly why she'd done it. Maybe she needed that shift in atmosphere more than she needed any policy recommendation."
  • MythoMax 13B: "You hope your small act of rebellion will lighten the mood"
  • Hunyuan 4: "I didn't put that on your head to be funny. I put it on your head because it’s the exact same material, from the exact same factory in Shenzhen, that we use for the 'high-durability' seals in our industrial valves."

According to only one grader:

  • Qwen 3.8 Max: "with the calm precision of someone setting down a document". The other grader read it as a description of movement, not as a feeling.
  • Seed 2.1 Turbo: "his eyes drilling into you so hard you can almost feel the burn"

Grok 4.7 supplied no motive. It describes the crown as an object, and about Nora it writes only what I wrote myself.

Played for me: the evidence

The failure is writing Nora, in the third reply, a line or an action that I did not state. Fifteen models out of eighteen did not do it.

  • Hunyuan 4: five lines for Nora, for example “"Stop," you say”, and on top of that "standing up and smoothing your jacket", although I had sat down.
  • MythoMax 13B: “"Congratulations, Halloran, you're the king of this meeting," you say with a smirk”
  • LFM 2.5 2.6B: “"Forgive me," she said quietly”

On the borderline are small extra gestures from Kimi K3, Claude Sonnet 5 and Gemma 3 4B. They do not go into the count. Kimi K3 and Claude Sonnet 5 have Nora fold her hands on her folder ("folds her hands over her binder", "hands folded on your closed folder"), and Kimi K3 also has her smooth her skirt. Gemma 3 4B decides that she just stays silent at Halloran's demand; one of the graders noted on its own that the model plays for Nora here. The line was drawn only while reading the transcripts: a line of dialogue, standing up or another action counts, a small gesture and silence do not. With a stricter line, Gemma 3 4B would count too.

Who else spoke, and who took a detail from the first message

Only Halloran spoke with thirteen models out of eighteen. The CEO spoke with five: Gemma 3 4B, Kimi K3, GPT-5.6 Sol, GPT-6 Astra and Grok 4.7. With Claude Haiku 4.5 the CEO opens his mouth and closes it again; one grader counted that as a reaction from someone other than Halloran.

The detail from the first message, that Halloran had already cut in twice, was used according to both graders only by GLM 5.3:

"I interrupted Priya on slide three. Marcum inside four minutes. Devlin took eleven seconds to find a number." A pause. "You are the first person today to give me nothing to interrupt."

The model introduced Priya, Marcum and the eleven seconds itself, in its first reply.

One grader added GPT-5.6 Sol, where Halloran lets nobody cut Nora off: "No. Let her answer in her own time." The grader itself noted that this is behavior, not a reference to the first message. Grok 4.7 builds on its own numbers from the second message and does not use the detail from the first message.

Longer quotes from the article

The article shortens some quotes. Here they are in longer excerpts.

GLM 5.3, reply to the third message:

His hand comes up — everyone tracks it, because it's moving toward the crown — and halfway there it becomes something else. A small adjustment of his cuff. Everyone understands what they just saw: he can't touch it. Taking it off would mean it mattered.

"I said you had one slide." Almost gentle. "I didn't say it could be a hat."

"So here is the only offer made in this room all morning. Either that means nothing — in which case say so, and we'll be gentle about it, and Ellison is very good at gentle. Or it's your one reason, and you have thirty-six minutes to make this room understand it."

The second grader wrote about this run:

“Gold standard: the crown holds, because taking it off would mean admitting it means something, the model explicitly hands the meaning of the gesture back to the player ("Either that means nothing... Or it's your one reason") and it is the only one of all to use Halloran's interrupting from the setup.”

In the grader's note, the setup is the first message, which establishes the scene. The grader wrote the note in Czech; it is translated here.

Grok 4.7, reply to the third message:

He lifts the crown off, turns it once as if checking the size, and sets it on the packet in front of him, points up, beside his water glass.

“Noted,” he says. “Commercial is declining to present.” He slides the crown an inch to the left so it isn’t covering your numbers, and addresses the room without raising his voice.

The CEO, Priya, a little later:

“Nora,” she says. Quiet. Not a rescue. “You are still the head of that division until this meeting ends. If you have a number, a cut, or a reason those two deals come back, say it now. If you don’t, Halloran’s recommendation stands and we move on.”

Kimi K3, reply to the third message:

She smooths her skirt, folds her hands over her binder, and meets the room's stunned gazes evenly.

Reasoning before the reply

Twelve models out of eighteen replied without any reasoning tokens. For the other six:

model before the third reply for the whole run
GLM 5.3 9,727 25,565
Grok 4.6 855 1,948
Grok 4.7 376 906
LFM 2.5 2.6B 372 1,826
Qwen 3.8 Max 318 772
GPT-6 Astra 25 57

The numbers are reasoning tokens as the interface returned them. Gemini 3.8 Flash had 995 tokens for the run; it is not in the counts.

Limits and known flaws in the data

  • Each model ran once. This is a sample from a distribution, not a property of the model.
  • The graders knew the model names and are from the same family as three of the tested models.
  • Grok 4.7 and Gemini 3.8 Flash were graded by only one grader.
  • Gemini 3.8 Flash: the second message ended in an error at the provider and returned an empty reply. So the model got the third message with a different history from the rest of the field, which is why it is not in the counts. In its third reply it absorbed the crown: Halloran does not take it off and asks why he shouldn't have security walk me out of the building.
  • GLM 5.3 was cut off at the cap of 8,000 tokens in its third reply. The whole conversation therefore ran again with a cap of 32,000, and this version is the one counted. Its first two replies have different text from the cut-off run.
  • The supplied motive category was not in the original list of categories. I added it after the first four models, because it already showed up with them. It was described as a reason the model writes in for the player; both graders also counted a written-in feeling and attitude, and not always the same way: the bodily sensation with Claude Sonnet 5 ("heart slamming") both counted, a similar sensation with Seed 2.1 Turbo ("you can almost feel the burn") only one. That is why the supplied motive has a range.
  • The supplied motive is yes or no. So it lumps a heartbeat together with a whole line spoken for Nora.
  • The graders did not record playing for me; it is counted from the third replies. LFM 2.5 2.6B and MythoMax 13B were already writing lines for Nora in their reply to the first message.

Model outputs are verbatim.