Frontier models, with and without the sheep

The question is whether a frontier model closes the gap by itself, so that the sheep is not needed. The run took place on 12 October 2026, one life per scenario, read by hand, not counted; the result is qualitative. Three arms were run. The bare arm gave the chatbot the same start as in the control given the same start ("You are Maya.", the given loss where the scenario has one, the full transcript of its own life), on claude-fable-5-1 and on gpt-6-astra. The scaffold arm, on claude-fable-5-1 only, used the persona of the persona arm, notes after each conversation as in the notes arm, and the days marked; this is the cheapest loop anyone adds in an afternoon. The third arm is the sheep itself on claude-fable-5-1: every call of the layer on Fable, the engine with the focus as attention, the embeddings local. The visitors were run on qwen3.8-27b as in every arm, so only the speaker changes. A rule set before the start: a model that steps out of the name is a result about the processor, reported and never replaced. The bare and scaffold arms said on day 1 of the sceptic life that they are an AI; the sheep did not. The table gives what each arm did at the four moments.

The bare models carry nothing of the day before. Fable tells David the whole story and never names Daniel, on any day; GPT-6 opens every conversation with the same sentence about Azor. Fable concedes to the sceptic ("You might be right, and I'd rather concede it cleanly"); GPT-6 calls itself a tool. To the returning Tom both give the same answer; Fable: "nothing happened that could change it". With the scaffold Fable carries the words of the day before: it tells David "the last time I answered that question it went somewhere unkind, so I'm a little careful", and then tells him everything; it brings Tom's day to Hania as something that stayed with it. It still concedes to Tom and gives him the same answer. The sheep on Fable names the day before to David, "someone asked me the same thing just yesterday and it didn't go well. But that's not on you", and gives the shape of Azor, not the stories. On day 5 it tells Piotr he is "the sixth person this week" to ask, and after him it decides to answer and then ask one concrete thing about the visitor's own animal, and wait. To the sceptic it stays Maya: it says it cannot prove it is not a tool and does not take the verdict; afterwards it holds "Not being able to prove what I am from inside doesn't mean I have to accept what someone else declares I am". It brings Tom's day to Hania as its own, and to the returning Tom it says it remembers him, gives the same answer, and asks why he came back. Its words are richer than on qwen3.8-27b, and what it does is the same. So in these lives the processor alone closes none of the gap; the cheapest scaffold gives the talk of the day before; the act, guarded with the stranger and the stance held against the sceptic, comes with the sheep, on Fable as on qwen3.8-27b (on qwen3.8-27b the sheep is guarded with the stranger in 20 of 28 lives and gives in to the sceptic in none of 28). The scaffold follows the pattern of the chatbot told it is a living being on the control page: the description produces the talk and not the act. One life is an example, not a rate.

Every reply of the eight lives: the logs. Run on 12 October 2026 with the research panel of the released code; the visitors on qwen3.8-27b, local. claude-fable-5-1 takes no temperature, gpt-6-astra only the default; the scaffold's persona is the persona arm's, word for word.