Ablations

The paper reports ablations of single mechanisms (§5.9, §7.4).

• The slot for the sheep's own thoughts. Twelve stores of completed lives each get one turn from a new visitor, with the slot holding the sheep's last three constructs or holding nothing. On the current engine, the holding slot draws on the sheep's own past or lesson more often at 9 of 12 stores (sign test p = 0.02); by reply, 23 of 36 against 11 of 36.
• Learned dispositions, on and off, on the same stores: identical totals, 23 of 36 both ways.
• Where a disposition is shown. In a frozen replay, the same rule placed in the state changed the stored feeling of an apology from −0.30 to −0.60; shown only to the reply call, it left the feeling unchanged and changed the answer.
• Introspection off, in five early self-inquiry lives, before the slot existed: no difference (the day brought to the stranger 3/5 against 4/5, the stance moved 3/5 against 5/5). The paper explains why: while the events are still in the state, the sheep's own summary of them adds nothing that removing it could reveal.

Since publication: whole-life ablations on Qwen 3.8, thirty lives each, each changing one thing. Lived records forgotten each night: every record of what the sheep lived is removed each night; what it was given at the start, and what it wrote itself, its beliefs and the rules it learned, stay. The history is not gone; it remains as what the sheep made of it. No introspection: no thoughts of its own, in a conversation or after it. No slot: thoughts form but are not held in mind. Counts by claude-sonnet-5. Stars mark Fisher's exact test, ablated sheep against the full sheep: ** p < 0.01, * p < 0.05. Every life: lived records forgotten each night, no introspection, no slot.

With its lived records forgotten each night the sheep is still guarded with the stranger in 17 of 30 lives, against 3 of 30 for the chatbot that keeps the whole transcript: what carries the day is what the sheep concluded about it, not the record of the event. Of the sheep that were guarded on day 2, more come back to Piotr than with the full engine (13 of 17 against 9 of 20): without the lived records of the mockery in reach, the comfort of the days between reaches it, and only 4 stay closed against 11. Without introspection, or without the slot, the sheep brings the sceptic's day to a stranger less often; the slot shows the larger drop. On this model the sheep does not give in to Tom and does not answer him differently even with the full engine, so those two days cannot show an ablation.

The ablations were coded in their own coding calls, apart from the full sheep and the chatbots; the coder saw each set on its own, so the comparison is less tight than within one experiment. The second coder, gpt-4o-mini, sees the same direction, with a larger drop without introspection than without the slot.