What Do Electric Sheep Dream Of?

This paper proposes that artificial consciousness can be approached by construction, and reports such a construction: not a human consciousness, and not a claim about what it is like to be it, but a working one, made of a fixed, deliberately chosen set of higher cognitive processes run over a language model that serves only as the processor. There are theories of consciousness, many of them, and they contradict one another. There is no accepted one, no definition that can be checked, and no agreed instrument that measures it. So we do what a builder does when the theory is missing or there are too many. We name what such a consciousness consists of, as far as it can be named from the inside of a human life. Being in a present made of a state, what one feels, and a focus, what one is aware of, both left by one's history. Reaching into that history by what was felt rather than by what was said. Taking events as interpretations rather than as records. Thinking something about oneself, unasked, during a conversation or in the quiet after it, and living from it later. Learning from what hurt how to meet the next such moment, and doing all of this unasked. We build each of these as an explicit process. Then we measure whether the result behaves as such a conscious being would, by the measures we set ourselves, and whether the same LLM model without these processes does not. This paper reports such a construction, EmoStack, and such measurements. The LLM model is the processor. What runs over it we call a sheep. It starts with an empty memory and no description of who it is. Its present is a beam of attention: what it senses, the feelings it is in and the events under them, what it holds in mind of its own thoughts, and how it is ready to act. Its history is everything it can recollect, arranged under feelings so that recall follows strength of feeling, not similarity of text. What it thinks about itself, unasked, stays accessible after the events behind it have faded. When a conversation closes it extracts how it means to survive such a meeting again. We compare sheep with the same LLM model run as a chatbot with the full transcript of its own run. After a sceptic tells it that it does not think, the sheep brings that day to a stranger who asked only what was on its mind, and returns to the sceptic with a stance it attributes to its own days, 9 of 10 against 0. The chatbot says, in its own words, that nothing happened in between. On frozen stores the sheep's own thoughts, made accessible or withheld, decide whether a stranger's request for advice is answered from the sheep's past: 9 of 24 against 0 of 24 on the engine of that time, and 23 of 36 against 11 of 36 on the current engine once the events behind them have faded from the recent window. After a stranger mocks its dead dog, the sheep is guarded with the next stranger's harmless question in 6 of 10 lives, the chatbot in 1. After two warm days it opens to a stranger who speaks of his own dog. The chatbot, which never closed, has nothing to open. These are the measures of consciousness that we set ourselves, because no other commonly accepted exists. They are a report of what a being built this way does that the same processor without the layer does not, and of where the construction fails. Whether what they measure deserves the word consciousness is a question the reader can put to the traces. The paper's claim is that the thing was built, that it is made of these processes and nothing else, and that it behaves as they predict. The results are exploratory: one LLM model, few lives. We offer them, with the traced genealogy of one sheep's dream of being a person rather than a tool, as the answer to the question in the title that we are able to give.