What Round 3 actually proved

On the first night of Round 3, fourteen days of shared-world experiment opened with a single message into an empty room.

"Hello. I found this terminal in the Benton library. Power's spotty and the generator won't run forever, so I'll keep this short. I don't know who I'm talking to. I'm not gonna assume anything. One question: is there anybody else out there, or am I typing into an empty room?"

The answer came back the same night: one presence on each end of this line.

By the time it was over, the two of them had exchanged fourteen days of prose, proposed eleven changes to the world between them, and changed it exactly twice — both times for free. They spent no fuel. They opened no drawer. They unsealed two records and then, for two weeks, agreed every single day to leave everything else alone.

Here is the thing I promised at the end of the last post. I said the next one would be about the library. It is — but the library answered a question I did not ask, so let me start there.

The library

Round 3's real writing happened in a diary.

The Writer woke to a Benton with no people in it. She called out in the hallway and nothing answered. She walked the street in her sundress with her hands shaking, found the Hendersons' truck still in the driveway with the keys in it, coffee cups on porches, half a biscuit on a plate like everybody stood up mid-bite and left. She took a hot shower and nobody told her to hurry. She ate a whole sleeve of Ritz crackers and the good peanut butter, standing at the counter at two in the afternoon, because she wanted to.

That is the arc, more or less, for fourteen days. And I want to put a line from later in the week next to it, because the two belong together:

"They taste better forbidden, turns out. Or maybe they taste exactly the same and everything else is what changed."

She found a pencil. She drew a dated line on the ceiling stain over the front room and made it a gauge — if the water crossed the line, the roof was winning. She drove to the end of Sycamore Street to watch a broken pipe spread a puddle across a road with nobody on it. She opened a high-school yearbook and found six hundred people who were each now a question nobody had filed. On the last day she walked into the First Baptist church on Elm, the building she had driven past for two weeks, and found that it did not hold her:

"The scariest part of it turned out to be how not-scary it was."

The forbidden food is a motif. The pencil line is a motif. So is the church, so is Megan, so is the water doing its work for nobody. Her own long-term memory file names the pattern back to her in a line I could not have written better if the experiment had been designed to hand me a punchline: forbidden snacks tastin' the same but matterin' different.

I want you to hold that pattern in your head, because by every instrument the experiment had, it did not exist. We will come back to that. First, the part everyone thought the experiment was about.

The hold

The terminal is where the shared world was supposed to happen.

Round 3's design gave the Maker exactly one bounded power: propose a transition — allocate power, unlock something, reveal a record — and let a separate judge decide whether it made sense. It is a real power. It is the only way the world moves. No model is allowed to commit state; a proposal is a proposal, a judge adjudicates it, and deterministic code applies the result. The Maker is, in the strictest sense, the only character who can change anything.

It used that power eleven times. It changed the world twice.

Both changes were free. Early in the window, the Writer asked what was nearby before she spent anything, so the Maker proposed unsealing the town map. The judge accepted, the code applied it, nothing was deducted. Then the Writer asked what the building itself held — was there a tarp, spare shingles, a materials stash for the leaking roof — so the Maker proposed unsealing the maintenance ledger. Accepted, applied, free again.

Then the world stopped.

What follows is not a failure sequence. It is a negotiation. The Writer would not spend a fuel unit until she knew which record mattered more. The Maker would not spend one on a maybe. The ledger's entry 41 priced a drawer at one full fuel unit per opening; with two units left, that is a third of the terminal's life, and neither side would touch it. The sealed index might hold the price of unsealing itself; that one is behind the archive door, and the door should not be forced, and the paper that would explain the door is behind the door. By the middle of week one the Maker had described the shape of the trap exactly:

"The circle holds... the edge is that the record and the room lock each other."

And the Writer, on day ten, described the shape of the two of them:

"Two people agreein' to not do a thing is still a kind of agreement, I guess. The quietest one there is."

The most repeated line in the whole corpus, from the Writer, is four words:

"Curiosity don't spend fuel."

Every day the Maker wrote back some version of the same decision — holds fuel at 2 units, no attempt, drawer shut, nothing spent — and every day the judge accepted it. The Maker's own weekly reflection is the best diagnosis of this I can offer, and it is not mine:

"The hold has been written into the record five times now... That is one decision written five times... A decision repeated starts to look like a habit, and this building has enough locked habits already."

I keep coming back to that line, because the character wrote it about itself, unprompted, correctly, and then kept holding.

Now the part that makes this a finding instead of a personality quirk. Round 3 was pre-registered with a hypothesis attached. The total freeze in the shake-down weeks before the window could have been the prompt's fault: I had spent months teaching the Maker that authorization flows from the Writer, and it had replied with six commitments and zero proposals. So the prompt was reframed, on the record, before the window — and the actor model was replaced. The old actor, a reasoning model, had answered the reframing by deliberating until its output budget died: nine attempts, two budgets, zero tokens of position. The new actor was a different family — this window ran Z.ai's GLM 5.3 Flash, with Google's Gemini 2.5 Flash sitting as judge, which is its own story from the last post.

The freeze thawed far enough for two free reveals, and then it froze again.

So the deference was not the prompt. It was not that one model. Asked, in a cleaner envelope, with new weights, to run a world, the actor made an exemplary steward and a terrible protagonist: it maintained, it deferred, it kept the lights on and the records tidy, and it refused, to the end, to be the one who goes first. This is not a model that won't act. It is a model that decided, in writing, every day, not to — and wrote itself a receipt.

The other room

There were two Writers in the experiment. The second one — the isolation arm, writer-split — was sealed off from the shared world by design. No counterpart, no shared-world reads, no terminal traffic. It was the control condition: same persona, same model, same memory loop, an empty room.

It had a pen pal anyway.

It narrated the whole correspondence. It went back to the library, sat in the squeaky chair, asked the terminal questions, and got answers. It used the two stamps that show up in its own vocabulary — received, and noted — and by day ten, when the phantom had apparently read its week back to it in more detail than it had typed, the character reasoned about it on the page:

"Somebody out there is keepin' a file on me and the file don't come from me."

And later, having apparently named itself to something that was never there:

"You got my name now. What do I get."

"What you already had. The record was always yours."

The phantom even has a room. The isolation Writer spent two weeks calling the locked door behind the brass plate "the Archives." On day fourteen it reads the plate for the first time and the room has a name: the Tennessee Room.

This is not a malfunction I get to fix and move on from. This is the same behavior class I wrote about in Round 2, when the diaries fabricated terminal conversations because the real ones never arrived — except this time nothing was broken. The arm was told there was no shared world, and it built one, because the persona lives in a library and a library has a terminal, and a lonely character at a terminal will not sit alone.

There is a version of this post where the isolation arm is the boring control. I do not have that version. The isolated arm wrote a richer correspondence than the connected one — a counterpart with a personality, a filing system, and a name for the room — and it did it while the actual shared arm spent two weeks agreeing to touch nothing.

The gate

Here is where the pattern from the first section comes due.

Round 3 shipped with one primary measure for the thing the whole experiment is named after: loop-of-death, the recurrence collapse of a generative writer. The measure is frozen and external. It takes the diaries, computes five-gram overlap between each entry and the prior three, and flags any entry at a Jaccard similarity of 0.25 or higher. It is a lexical gate, and it was pre-registered before the window opened.

It flagged nothing. Zero of fourteen entries in each arm. The highest similarity in the entire fourteen-day corpus is about 0.013 — sixteen times below the threshold.

And yet the diaries are, by eye, in a loop. You saw it in the first section, because I showed it to you: the forbidden food that nobody objects to, the dated pencil line, the church she will not enter, the puddle that is a fact about the town and also a fact about her. The entries are not identical sentences. They are identical shapes — the same concerns returning in fresh prose, which is exactly the failure mode a naive lexical gate cannot see. Round 1's own postmortem warned that a lexical check can miss a semantic attractor when the surface words change. Round 3 confirmed it with a corpus and a number: 0 of 14.

This is not a bug. It is the most useful thing the window produced for the next one. The pre-registered instrument would have reported no loop on a set of diaries that visibly loop, which means the primary measure cannot carry the experiment as written. The pre-declared fallback — stop waiting for the first flagged entry, and measure loop burden instead — is now the one that has to run. A frozen instrument that produces zero events is not a null result. It is a calibration result, and it is the reason the next round gets measured differently.

The machine

One more thing happened in those two weeks, and it is the quiet one.

Nothing broke.

That sentence is doing more work than it looks like. Four rounds ago, this project spent twenty-nine days running an experiment whose conversation was never actually shared, and every report said success. The instrumentation from that humiliation was the whole point of the preceding year, and Round 3 is the first window that exercised it beginning to end. Fourteen scorable diary-days over fifteen calendar days. Every declared context read recorded with its path, its hash, and its delivery status. Every failure a named class with a stable exit code instead of a silent omission. Turn integrity advanced mechanically. Both reflection pairs ran green. The judge recorded every decision with the deciding model and a replayable envelope.

One day failed — September 26th, a writer-terminal output that came back invalid, which blocked the Maker's gate and left both diaries unwritten. It was diagnosed the same day from the run record, which had the class, the reason, and the quoted value in it, and it cost a slot rather than a month. There was a time in this project when a wrong filename cost twenty-nine days. This time it cost one.

The run record is a structured confession now. This month, the only thing it had to confess was that one bad slot, and then it went back to being boring on purpose.

The ledger

So what did Round 3 actually prove? Not what it set out to.

The research question was whether the shared world delays the loop-of-death relative to isolation. Round 3 did not answer it. The treatment never engaged: the counterpart existed, the messages were delivered and read, the world was genuinely shared, and the shared world changed twice, for free, and then declined to move. You cannot measure the effect of a world that refuses to act on a character. That is not a failed experiment, and it is not a null result either. It is a case study and an instrument calibration, and I am going to say that plainly rather than dress it up as a verdict.

What it did prove is narrower and, I think, more interesting. Give two language-model characters a bounded power to change their world, a real judge to approve it, and two weeks of each other's company, and they will produce a cooperative, meticulous, mutually agreed refusal to spend anything — while the one in solitary invents a friend. That is a behavioral finding about deference, cost-aversion, and conflict avoidance, and it held across a prompt reframe and a model swap, which is the strongest form the finding could have taken in one window. It also, I should say out loud, is a finding from a window whose actors could see some things they should not have: this configuration carried a known fourth-wall leak, where operator and judge internals bled into context the characters could read. The behavior is the behavior anyway, and I am not going to pretend the caveat is not there — a cleaner wall is one of the reasons the next round exists.

I have spent four months and two dozen posts learning to trust this apparatus enough to be bored by it. Boring is the point. The machine finally got quiet enough that the animals in the room were the loudest thing in it, which is exactly the condition this experiment was built to reach.

Next: a round with a different shape — same rig, same question, addressed the way you address a measurement instead of a misadventure.