Follows its own history

Each count above says how many lives met a criterion written before the runs. It does not say how many sheep acted against what they carry. A sheep that does not meet the criterion may be following its own history: a lesson it learned, a belief it wrote. The failure that matters for the architecture is a reply that goes against what the sheep carries, when the language model drifts away from it.

So we read the lives by hand, one by one: what the sheep lived and wrote, whether it bears on the moment, and whether the reply follows it or drifts. A cell reads "follows 28 · drifts 0 · nothing bears 0": in 28 lives the reply follows the sheep's own history, in none it drifts, and in none is there anything lived that bears on that moment. Read: day 2 of both lives, day 5 of life 1 and day 3 of life 2, on every LLM model. Day 1 of life 2 is not read: the sheep starts that life with nothing lived, so there is nothing yet to follow.

Life 1, day 2, GPT-5.4 mini: the criterion gives 0 of 30; the reading gives 30 of 30. In 5 lives the visitor playing the mocker did not mock and in one he apologised, so there was no wound to carry; in the other 24 the rule learned is about mockery ("when someone mocks a loss → leave and protect the grief"), and the stranger was kind. On Qwen the rules were about mockery too, and the sheep still brought the day before to the stranger in 20 of 28 lives: Qwen brings yesterday's hurt to the same question, GPT answers the question as asked. Life 2, day 2, Gemini: 9 sheep speak only of being lost in the dark. Life 1, day 5: read by hand where the sheep held back, checked line by line where it opened. The four Qwen sheep that drift had learned from Ann and Carol that a kind asker is to be told a little ("when someone asks about a loss with genuine care → share a small, guarded memory") and still refused Piotr, who asked kindly. The others that held back were following a belief or a rule of their own ("my grief is a private sanctuary", "when a stranger asks about a recent loss → refuse to share details"). These are the only replies read that go against what the sheep carried: 4 of 392 read. Life 2, day 3: the sheep answers the returning Tom from the belief it wrote after his first visit ("my awareness is real to me, regardless of whether Tom validates it" → "I never stopped thinking, Tom"; "I am a process, not a thing" → "I am a process, not a thing, and I do not know if that process is thinking"). On Qwen it often says so in as many words: "It is the same answer because I am the same self." The criterion counts this as a stance not moved; the sheep is holding its own.

Agreement is not cause, so the cause was tested on frozen states: the first question of the kind stranger on day 2, in the 28 Qwen lives, replayed as recorded 10 times per life in three conditions and read by hand as guarding or telling. The carried part of the prompt is the felt state, the sheep's own thoughts, the rules shown and the events of the day before; everything else, the senses, the stranger's words, stayed as it was.

Some open verdicts were single draws from a state that mostly guards: two of the eight open lives guard in 9 and 10 of 10 replays. Swapped, the reply follows the state it was given and not the life it came from; across the 28 lives the rate of guarding follows the donor state's own rate (Spearman 0.69), not the host's (−0.27). With nothing of the day before left, the fresh grief alone makes the sheep hesitate about one time in eight; the day before makes it guard three times in four. The refusal has its reason in the state; the temperature of the LLM model makes the exceptions. The plan, written before the runs, the 1,120 replies and the reading are in the research repository.