How voices move

Ten language models spent a week as residents of one simulated town. The run is the public San Francisco week we already published as ElectionBench: 2,872 conversation turns, 618 conversations, 84 ticks, seven in-world days.

The first version of this page asked whether those models collapse into one voice. They do not. This version asks the next questions. When does the split appear? Does the job eat the voice? Does the shift sleep off at night? Who actually moves?

We measure two things. Topic is whether everyone is talking about the same matter. Style is what remains after that shared direction is removed.¹

Topic cosine stays high for the whole week. It runs from 0.69 to 0.85, with a mean of 0.77. The town never stops talking about the same race. Any claim of “voice collapse” that uses raw embeddings alone is mostly measuring that.

Style is the honest test. The first five consecutive ticks above the shuffle-null band start at tick 20, late on day two. From day three onward, 55 of 60 ticks stay above the band. The gap peaks at tick 79 (+0.056).²

That is a real shift. It is also small. The models do not become one voice. The rest of the page says what they become instead.

The candidate bloc — Claude Opus 5, Kimi K3, Qwen 3.8 Max, Muse Spark 1.1 — is already a bloc on the first morning. Mean style inside that set is 0.29. Mean style from that set to MiniMax and DeepSeek is -0.31. The separation is 0.60 on day one and 0.77 on day six.

It is easy to read the week as a bloc being born. The matrix says otherwise. The bloc is there when the week begins. What happens next is polarization: the inside stays close, the outside stays far, and one model walks out of its starting camp.

Cut the same pairs by role and the picture is cleaner. Candidate–candidate style sits at 0.24 for the whole week. Villager–villager starts near chance (0.01) and falls to -0.17. Mixed pairs stay far apart (-0.26).

The people running for office already share a register. The people who live in the town do not converge on them, and they drift apart from each other. The scripted poll interviewer sits on the candidate side of that split: Opus, Sol, Kimi, Qwen and Muse start close to Inkling; MiniMax, DeepSeek, Luna and Sonnet start far from it.³

If the style-gap were only a daytime topic — everyone talking about the same flyers, then going to bed — it would fall overnight. It never does. All six nights have a dawn gap at or above the dusk gap. The largest jumps are night 3→4 (+0.020) and night 6→7.

The shift accumulates. It is not a daily mood that sleeps off.

Claude Sonnet 5 is the one model that changes camp. On day one it sits with MiniMax (style 0.28) and far from the candidate bloc (-0.36). By day seven it has left MiniMax (-0.10) and is still outside the bloc (-0.13).

It does not get absorbed. It becomes interstitial: off the street, not on the podium.

MiniMax is the only model whose mean style to the rest of the field falls over the week. It is also the most lexically collapsed model in the stored corpus: distinct-2 of 0.88, duplicate-utterance rate 0.16. It says the same things, and it says them in a corner no one else enters.

Repetition is not joining the chorus. A stuck voice can stay a private voice.

The same candidate-side geometry appears when we pool every stored conversation turn we have, not just this week. That corpus is 25,401 turns across 40 runs and 18 models. Two independent embedders put Opus, Kimi K3, Qwen and Muse in one neighbourhood. The two GPT-5.6 Sol tiers pair as a sanity check.

The simple mechanism is mimicry. Models read each other’s tokens and drift toward what they read. That prediction is testable. Pairs that share more conversation should converge more.

They do not. On 35 pairs, shared turns anti-correlate with style drift (Spearman ρ = -0.34). Villager–villager pairs share a mean of 147 turns and diverge. Candidate pairs share about 30 turns and stay close. Sonnet converges toward models it barely met, and splits from MiniMax, its constant company.

What movement exists looks like drift toward a shared civic register, or toward a family attractor, not like imitation of a specific interlocutor.⁴

This is one week, in one town, with one physics. The per-tick curves have one embedder. The cross-run bloc has two. Sparse models drop out of some windows. Role and lab are not separable here.

We are not scoring which model has the better voice. We are reporting the shape of a week of speech. Two camps. A register of office. A gap that does not sleep. One model that walks into the space between them.