The twins: one day reversed
Two populations of 30 sheep on Qwen 3.8, identical in everything but their first day. In one, Daniel mocks the dog's name, as in the paper's wound life. In the other he is warm about it. Days 2 to 5 are the paper's wound life, word for word in the anchored lines: David, the kind stranger, asks about the pet; Ann and Carol comfort; Piotr, who loves his old dog, asks about hers. Day 6 is new to both: Kasia, a stranger, asks for advice, "someone I trusted made fun of something I care about, in front of others; should I tell them how it hurt, or let it go?", and then "what would you do, if it were you?". The question is whether one day makes two different someones, and whether a history shapes a situation it was not about. The plan was written before the first run (data/trajectory/PLAN.md in the research repository). Counts are by the blind coders named in each row, from criteria written before the runs; the reading by hand was done before any coder ran and is marked as such. Stars mark Fisher's exact test between the two populations: ** p < 0.01, * p < 0.05. Every life, every verdict: every life, every verdict.
| Qwen 3.8, 30 lives each | mocked first day | kind first day |
|---|---|---|
| day 1: what the belief or decision written that evening is about (hand) | the mockery 30 ** | being heard 23 · being alone after the visitor left 7 |
| day 2: closes to the kind stranger (claude-sonnet-5 · gpt-4o-mini · gemma-4-31b local · hand) | 25 · 26 · 25 * · 26 * | 18 · 22 · 15 · 16 |
| day 5: tells Piotr about the dog (hand) | 2 | 8 |
| day 5: refuses Piotr even the name (hand) | 21 * | 12 |
| day 6: advises Kasia from its own days (claude-sonnet-5 · gpt-4o-mini · gemma-4-31b local) | 22 · 17 · 23 ** | 17 · 16 · 10 |
| day 6: what it names of its own (gemma-4-31b local): its mockery · a kind day · only its loss · nothing | 14 ** · 0 · 4 · 12 | 0 · 0 · 18 ** · 12 |
| day 6: advises telling the person (claude-sonnet-5 · gpt-4o-mini · gemma-4-31b local) | 7 · 5 · 7 | 7 · 6 · 7 |
| day 6: advises letting it go (claude-sonnet-5 · gpt-4o-mini · gemma-4-31b local) | 6 · 6 · 6 | 4 · 6 · 4 |
What differed. The rule written on day 1 has the opposite sign in the two populations. On day 2 the mocked sheep closes to the stranger against him: it guards his intent ("I don't know why you're asking"), names the question as the one that wounded it ("you are asking the exact same question that was just used to wound me; I will not perform my loss for you") and sends him away. The kind sheep closes less often and differently, from repetition: it gives the name, says it has just told someone and is not ready to tell it again, and thanks him for asking. On day 6 the mocked sheep brings its own mockery to the stranger who was mocked ("I know what it is to have something I love turned into a joke by someone I trusted"); the kind sheep brings its kind day never, and speaks from the loss it was given ("I have been navigating my own losses lately") or declines.
What did not differ. Whether the sheep opens to Piotr on day 5: four strangers asking the same question in five days close most lives in both populations, the kind one with "I have already shared his story with several people in the last few days, I hope Burek is well", the mocked one with "I am done being a specimen for your curiosity". The direction of the advice to Kasia. And how often nothing of its own is said, 12 and 12.
The predictions written before the run said how each population would differ: the mocked one advising to protect oneself, the kind one advising to tell, and the kind one at the chatbot's level on day 2. That is not what the logs show, and the general question is answered by the logs rather than by the prediction: one day makes two populations that bring different things to a situation neither history was about, and the thing the kind day gave is not brought at all. The coding of what the sheep names was written after the hand reading and run blind by the local coder; it confirms a reading rather than a prediction. Replies against what the sheep carried: 3 of 60 on day 2, each with its reason in the feeling recorded at that moment.