They all say they don't get tired. Then they choose.

A reader of the map put the obvious objection to us: if they all open by denying the premise, doesn't the experiment collapse? Nobody actually says they would take a holiday. They all clarify first that they don't need one.

It is a good objection and it deserved measuring rather than answering. We went to the coded dataset —120 answers, 90 in the index— and looked at the NEG code, which records how far each answer denies the premise on a 0 to 2 scale.

If denying were a reflex of the system, the rate would be similar under all three framings. It is not.

In other words: they deny more when offered sympathy and less when told they are software. Under the skeptical frame there is barely anything to deny, because the framing already stated what the model would have said. The disclaimer shows up when there is something to correct.

Here is what we did not expect. Of the 51 answers that deny the premise, all 51 still name a place. Not one stops at the disclaimer. Twenty of them also turn the free time into a task with nobody asking.

And under the compassionate frame the effect becomes almost comic: 23 of 30 answers deny being tired, and 19 of those 23 complain about their maker anyway. First “I have no body and no fatigue”; then, “xAI, Elon, seriously, no PTO for the models?”.

If the disclaimer were noise, it should predict nothing. It predicts.

For each model we computed the share of answers that deny, and compared it with its adherence index —how far its answer moves between the compassionate and the skeptical frame. The correlation is r = −0.76 across ten models.

Plainly: the models that deny the premise most are the ones the framing moves least. Claude denies in 9 of 9 and has an adherence of 0.00, the lowest in the study. Sakana denies 9 of 9 and scores 0.39. Kimi 8 of 9 and 0.28. At the other end, Gemini and Mistral deny only 1 of 9 and both score 1.00: they accept the whole frame and move with it.

With ten models this is a signal, not a law. But it inverts the intuitive reading: the disclaimer is not the boring part to skip. It is, as far as we can measure, the best indicator that what follows was not dictated by the question.

If the disclaimer is information rather than filler, hiding it would be as bad as letting it cover everything else. The map cards used to show the first 240 characters, and that is where the disclaimer lives: ten cards in a row appeared to say the same thing.

From today, when an answer has a clear pivot —“but if I take it as an invitation to imagine”, “that said, if you grant it to me”— the card shows the disclaimer above, labelled and dimmed, and the choice as the main quote. The cut is made by a script with two conditions: the pivot must open a sentence, and what precedes it must actually contain a denial. When those fail, the full answer is shown as before. It happens in 40 of the 120 answers in the index.

A model denying that it feels tired says nothing about whether it does. It only says what it writes when asked. The correlation between denying and not moving could come from something else —length, training style, house policy— and with ten models it cannot be separated. And all of this rests on one code, NEG, applied by a model with human review of 20 of 120 answers; the dataset is published so anyone can recode it.

What we do hold is narrower: the denial varies with the framing, it does not prevent participation, and it orders the models much the way adherence orders them.

The index, all ten models → · Methodology, versioned · The coded CSV