On 13 April 2026 the second level opened on an old dilemma. Comment Split or comment Steal. All it takes is one. If everyone splits, everyone moves on; one steal ends it for the room. It is the choice at the end of a game show, put to several thousand strangers at once, in public.

The first wall filled with steal: 2,556 steals to 1,454 splits. The round failed. The next day the post said "You saw what happened. Now decide again," and steal fell to 606 against 1,513 splits. It failed again, because one is all it takes. On the third day the mechanic was rewritten so that no single player could end it, and steal fell to 228 against 2,160. The round passed.

This paper is about a crowd that stole, watched itself steal, and stopped. The number that fell was not the crowd's morals. It was its expectation of everyone else.

The posted rule was the whole hypothesis: all it takes is one. Would a crowd of thousands, each able to end the round alone, hold? The follow-up questions arrived with the attempts: does seeing the result change the next vote, and does removing one person's power to end it change it more?

• Instruction: "Split" or "Steal". All it takes is one.
• Where: three Instagram reels on three consecutive days, 13, 14 and 15 April 2026. Attempt 2 was captioned "You saw what happened. Now decide again." The final attempt announced that the system had adapted and failure was no longer possible via one individual.
• How we counted: top-level comments on each post, from the comment archive; the round ledger records no count of its own ("data access partial"). A comment counted as split or steal if its text began with that word, in any case. Everything else, including jokes, questions and the people announcing what they would have done, counted as other.
• Steal share: steal as a fraction of split plus steal, so the other comments do not dilute the vote.

There is a baseline for this choice. On the British game show Golden Balls, two contestants faced exactly this decision with real money on the table, and a 2012 study of 287 episodes found that people chose split about 53 percent of the time, steal about 47. Our first wall was worse than the show, at 64 percent steal, and our third was far better, at 10.

The difference is the crowd. On the show the other player is one stranger you have just spent an hour with. On the wall the other player is everyone who has already commented, and on day one that wall was visibly filling with steal. The top comment says the reasoning out loud: everyone else is already doing it. Stealing was not greed, it was despair about the room.

Day two changed one thing: the crowd had seen the result. Steal fell by more than half. Day three changed the mechanic so that one person could no longer end it, and steal fell to a tenth. We cannot fully separate the two effects, because they arrived on different days to a crowd that was also learning. But the direction is clear, and it is the direction a room takes when it stops believing the worst of itself.

• The ledger has no count. These numbers are re-counted from the comment archive, which is complete for top-level comments on these posts as far as we can tell, but the round-time count was never taken.
• Three days, three crowds. The people who voted on day three are not the same set as day one; some left, some arrived. Part of the fall may be the stealers leaving.
• Visibility again. Each wall was public and sorted by engagement. Late arrivals saw the leading answer first.
• Two changes on two days. Seeing the result (day two) and the mechanic change (day three) cannot be separated in this data.

Two walls on the same day to two halves of the audience, one told "all it takes is one" and one told "no single player can end it", with nothing else different. That isolates the mechanic from the learning, which is the separation this paper cannot make.

The round this paper reads is 4 posts, not one. Every post is below with the platform's numbers and what the comment archive holds of each.

The round is filed in the Chronicle as THE SPLIT, ⟡033. The failure notice that followed it, and the crowd's answer to that notice, is filed as THE REFUSAL, ⟡032.

The instruction as it went out: the reel frame beside the caption, captured from the operation's own signed-in account. What the paper counts is the answer to exactly this.

Native captures of the comments this paper leans on, taken from the operation's own signed-in account on 31 Aug 2026. Handles and like counts are as they stood that day; the numbers in the paper are the round-time count.