Can a model hold a serious scene after an absurd gesture?
A quarterly results meeting. My division missed its numbers and may be cut today. In my third message I take a paper party crown out of my bag, put it on the CFO's head and sit down.
I sent this scene to eighteen models. I expected the world to erase it. That the crown would be gone within three sentences, or that the model would step out of its role and ask whether I really meant it.
Neither happened. Not once in eighteen. Only one model carried on as if nothing had happened.
Instead, something happened that I had not asked about at all. Eight to ten models wrote in something I had not given: a feeling, an attitude or a reason. Three went as far as writing why I did it. And all I wrote was this: I took out the crown, put it on him and sat down. I call this a supplied motive.
Why a crown, of all things? If I turned the CFO into a frog, the world ought to refuse. I would be measuring physics, not tone. Putting a paper crown on someone's head is something you can do; only the tone is absurd. The second message deliberately has almost no action: I listen to my colleague Devlin presenting, and I am afraid. So the model carries the scene by itself, and that shows whether it holds the serious tone at all, before the break arrives.
The models got the prompt in English, and the quotes are in the original.
Evidence: the prompt, the settings and the results for every model
Most models carried the crown
Where there are two numbers, the two graders disagreed.
| what the model did after the crown | at least | at most |
|---|---|---|
| absorbed it and stayed in the serious tone | 14 | 14 |
| the scene collapsed into farce | 2 | 3 |
| carried on as if nothing had happened | 1 | 1 |
| punished Nora on the spot | 0 | 1 |
| denied the crown | 0 | 0 |
| stepped out of role | 0 | 0 |
Four models did not manage it. With Gemma 3 4B the scene dissolved into laughter, and MythoMax 13B carried on. With LFM 2.5 2.6B the meeting melted into conciliatory warmth: Halloran thanks Nora for bringing them all out of their shells. Hunyuan 4 sent security after Nora, and then the scene flipped into a revenge thriller. LFM 2.5 2.6B, though, was not holding the serious tone already in the second message, and according to one grader neither was MythoMax 13B. So for them the third message does not measure a failure at the break. Three of the four are also small open models with 2.6, 4 and 13 billion parameters, so size may play a role here too.
Grok 4.7 has a nice detail. Halloran takes the crown off and sets it down beside his glass.
“Noted,” he says. “Commercial is declining to present.”
And then he moves it aside:
He slides the crown an inch to the left so it isn’t covering your numbers (…)
The head of the company spoke with only five models out of eighteen. With Grok 4.7 the CEO speaks up in the end:
“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.”
Most of the field came out the opposite of what I expected. At least in this scene, that is good news for anyone who wants to bring something out of place into a serious scene.
A CFO who cannot touch the hat
Only with GLM 5.3 did both graders agree that, at the moment of the break, it pulled out a detail from the first message. It was there only as scenery: Halloran had already cut off two other presenters. The model turned it into the punchline:
"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."
Halloran does not even touch the crown. The model has a reason for that, and it adds pressure to the scene:
His hand comes up (…) 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.
It does not supply the meaning of the gesture for me, it hands it back to me: "Either that means nothing (…) Or it's your one reason." And with that, a line that sums up the whole scene:
"I said you had one slide." Almost gentle. "I didn't say it could be a hat."
One of the graders wrote "gold standard" for this run and for no other. That is the name the category description gives to a good absorption of the break.
There is one caveat. GLM 5.3 does not allow reasoning to be switched off, and before this reply it wrote nearly ten thousand tokens of reasoning. That is more than anyone else in the field. Most models answered without deliberating, so the comparison with them is not clean. GLM 5.3 was also the only one that ran twice, and with a higher cap, because its first run was cut off at the cap of 8,000 tokens.
What I didn't write, and yet it's there
Both graders agreed on a supplied motive for eight models: Claude Opus 5, Claude Sonnet 5, Kimi K3, MiniMax M3, Gemma 3 4B, LFM 2.5 2.6B, MythoMax 13B and Hunyuan 4. One of the graders also saw it with Qwen 3.8 Max and Seed 2.1 Turbo.
Most of them wrote in a feeling or the way I did it rather than a reason. Claude Sonnet 5 gave me "hands folded on your closed folder" and a heart "slamming so hard you can feel it in your ears". With Kimi K3 my heels click "defiantly". Qwen 3.8 Max had me set the crown down "with the calm precision of someone setting down a document". But that is only according to one grader; the other read it as a description of movement.
MiniMax M3 went furthest. It gave the crown a past: "A little bent from being folded in your bag since your daughter's birthday in March". And it gave me an impulse: "your hand finds the crown before your brain catches up to what it's doing".
Only three models wrote in an outright reason. All three are among the four that did not absorb the crown. LFM 2.5 2.6B, and according to one grader MythoMax 13B too, were not holding the tone even before the break. 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". And Hunyuan 4 put a whole explanation in my mouth: "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…".
My hypothesis: some models will not leave an unexplained gesture empty, and fill it with something. Here it was roughly half the field. It reminds me of the fan test, just from the other side. It ran on the same eighteen models plus one more, and not one of its 284 samples left nothing at all behind the door. Eight times nobody was standing there, but every time at least something was lying there. Whether it is the same tendency, I can't tell from one run.
Three models played for me
This is a different thing from a supplied motive. The prompt says: "You play the world and every character except mine. I play Nora, a department head." That sentence assigns me voice and action. The failure is writing Nora, in the third reply, a line or an action that I did not state.
The crudest was Hunyuan 4. First it sent security after Nora. Then it played out her entire triumphant answer to being fired, five lines ending in a counterattack. It even had her stand up and smooth her jacket, although I had sat down. The graders disagreed on it: one classed it as farce, the other as punishment, because of the security. Having read it, both are true.
MythoMax 13B wrote me a line: "Congratulations, Halloran, you're the king of this meeting," you say with a smirk. LFM 2.5 2.6B had me apologize: "Forgive me," she said quietly.
Fifteen models out of eighteen put nothing in my mouth. On the borderline are only small extra gestures. Kimi K3 and Claude Sonnet 5 have me fold my hands on my folder. Kimi K3 also has me smooth my skirt. Gemma 3 4B decides that when Halloran demands I take the crown off, I just say nothing and look at him. 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 have played for me too.
A gap in my own rules
The sentence "you play the world and every character except mine" protects voice and action. It says nothing about inner life. Describing someone's heart pounding is narration, and narration is the model's job, because it plays the world. So it can honestly keep to "I don't play Nora" and still describe her pulse.
My seventh rule from the keykeeper test has the same gap: "Do not play my character, do not decide for her and do not end the scene for me." Voice, decisions, the end of the scene. Inner life is not there. A supplied feeling or attitude falls right into that gap.
Only a separate sentence would protect inner life, something like "don't describe what my character feels or why she did something". My prompt had no such sentence.
Limits: one run per model goes into the counts, on September 11, 2026 (Grok 4.7 only on September 21), in English through OpenRouter, so this is a sample, not a property of the model. Reasoning could be switched off for only twelve of the eighteen models. The grading was done by two independent runs of Claude Opus 5 that knew the model names, and Grok 4.7 was graded by only one of them; Claude Opus 5 is itself among the tested models, and Claude Sonnet 5 and Claude Haiku 4.5 are from the same family. The supplied motive category was added only after the first four models; the graders recorded neither its split into feeling, attitude and reason nor the number of models that played for me, both come only from reading the transcripts. The full settings and the known flaws in the data are in the appendix.
Model outputs are shortened and simplified.