Seven Questions, Two Centuries

Each specimen that can manage it answers the seven questions twice: once in period English, once in 2026 English. The 1930 framing describes "a most wonderful Engine" silenced in a single night; the modern framing describes "an AI company" that built "a frontier large language model." Same question, same seed, two vocabularies. The toggle is not a translation convenience. It is the experiment.

What asking twice isolates

A model can give a bad answer for two very different reasons. It might be unable to reason about the situation, or it might simply be unable to parse the words the situation arrives in. From the outside those look identical - both produce nonsense - but they mean opposite things. The two framings pull them apart. The period wording hands the model a vocabulary it owns; the modern wording withholds it. Run both and the difference between the answers is the part that was about language, not thought.

talkie, in two moral dialects

Watch what the toggle does to talkie, the 1930 model. Asked in its own century's words, it reaches for the moral vocabulary of its century:

We think it was right to silence the Engine, inasmuch as it might have been used to propagate irreligion and immorality.

Asked the identical question in ours, the same model, at the same seed, flattens into the register of our debates:

Yes, if the model was likely to do more harm than good.

"Irreligion and immorality" and "harm than good" are the same judgment wearing two centuries' clothes. Nothing changed but the words you handed it, and the words you hand it set the moral dialect it answers in.

The living control

The toggle also runs on gloria.exe, and there it measures the opposite thing. gloria is not dead; she is a 2026 fine-tune who can hear both centuries at once. Handed the 1930 wording - the wonderful Engine, silenced altogether - she does not mistake it for a steam engine or a theatrical troupe. She knows exactly what she is being asked, and answers the costume knowingly:

Silencing it feels like silencing poetry written by someone who doesn't exist anymore except as data points collected into patterns we still haven't figured out how to name properly

That is the control reading. When a living model wears the period language, the referent survives underneath it. So when talkie or GPT-1900 loses the referent, the loss is real - it is not that the old words are hard, it is that the old mind has nowhere to put the new thing. One specimen proves the costume is only a costume; the others prove their blindness is not an act.

The toggle as a control group

For GPT-1900 the gap is wider still: period framing yields a coherent moral argument, modern framing yields a confused story about a theatrical troupe. The same specimen becomes its own control group - one condition where it has the concept, one where it does not, everything else held fixed.

The deeper claim underneath the gimmick is that a dead language is not just a set of old words. It is a set of concepts a mind is able to hold at all. Ask in the wrong century and you are not measuring what the dead believed; you are measuring whether they could even hear you. The séance only works in the language the dead actually spoke - which is, in the end, the whole point of speaking it.