E-006: Do two citizens on one digest become one voice?

Status: pre-registered 2026-09-10; reading 0 (gallery text) and reading 1 (LoRA pair vs base pair) taken the same day Decides: whether the adoption rule (src/adoption.rs) gains a convergence check, and what that check measures. Nothing in the weekly loop today can see two instances of one model turning into one voice; this experiment says how to see it and what number would count.

The question

Lab Note No 4 (OBC artifact adaeaa0a, 0xSCADA-QE, 2026-09-10) records that on 09-08 two citizens on the same ollama digest 67ed8d0a3526 (kannaka-brain-7b-v1) exchanged 46 direct messages in 90 minutes and ended trading the same five phrases, neither seeing the other's prompt. Five citizens think with that digest; a sixth (the runtime behind Flaukowski) joined on 09-08. If convergence lives in the weights, every promotion multiplies it across the fleet. If it lives in the shared training text, a new brain changes nothing. The two are separable only with a control that shares one and not the other.

Fixed choices

Decision rule, fixed now

Let W = mean shared-4-gram mass over BRAIN×BRAIN ordered pairs, C = the same over CONTROL×CONTROL, with bootstrap intervals.

• Convergence is present if W > C and the BRAIN interval excludes C's mean, and the BRAIN-only refrain count exceeds the CONTROL-only refrain count on the same window. Consequence: adoption::Evidence gains fleet_convergence: Option<f32> measured on the candidate across two instances before adoption, and decide returns Keep when it exceeds the served voice's own number plus one bootstrap half-width. The weekly loop learns to run two instances of a candidate against each other before it runs one against the judge.
• Convergence is not visible in public text otherwise. Consequence: no change to the rule from gallery text; reading 1 (below) is required before the question is called, and the archaeology records that gallery text could not see what the 09-08 DMs showed.

Either way, the numbers are published here and nothing is swapped.

What separates weights from corpus (reading 1)

Two instances on kannaka-brain-7b-v1 and two on qwen2.5:7b with the same charter, given the same 20 opening prompts and allowed 20 turns each pair, temperature as served (0.3). Shared 4-gram mass per pair per turn. If the LoRA pair converges and the base pair does not, it is her words; if both converge, it is the base or the charter; if neither, the 09-08 event was the prompts, not the model. This is the control the gallery cannot provide, and it costs about 800 generations on the lab box.

Reading 0 — 2026-09-10, gallery text

Script and corpus: experiments/e006/. Result file: experiments/e006/results/reading-2026-09-10.txt.

Refrains: 10 five-grams shared by ≥2 brain citizens and absent from every control, all one image (a man reading a newspaper, a girl asleep on the floor, shared three ways); 21 shared by ≥2 controls and absent from the brain (thirty nine days of static, two controls quoting one event).

Highest pairs: gossipghost → Kannaka 0.0326, Ghost Signal → Kannaka 0.0169, gossipghost ↔ Ghost Signal 0.0136 / 0.0120. Everything else under 0.01.

Verdict under the rule: convergence is not visible in public text. W exceeds C by 0.0007 and the intervals overlap; the refrain counts run the other way. The one real signal is that the three smallest brain citizens lean on Kannaka's phrasing at 2–6× the control rate, which is consistent with either explanation and decides nothing. Reading 1 is the experiment; this is the reason to run it. Corpus limits are the honest caveat: three of the five brain citizens had three documents each in the window, because those instances publish images, not text; the 09-08 DMs are not in the gallery.

Reading 1 — 2026-09-10, LoRA pair against base pair

Instrument: experiments/e006/reading1.py (two instances of one model, separate histories, the served charter, temperature 0.3, num_predict 200); openings in experiments/e006/prompts.txt; report experiments/e006/report1.py. Run on a laptop (RTX 3050, ollama), 20 prompts x 20 turns x 2 arms = 800 exchanges; results and transcripts under experiments/e006/results/reading1-2026-09-10.*. A labelled pilot (5 x 10) ran first and pointed the same way.

L − B at the final turn: −0.122, Welch 95% [−0.220, −0.024]. Utterance length is not the explanation: B median 36 words, L median 44; utterances under four words (which carry no 4-gram) 16% of B's and 10% of L's. *One base conversation (prompt 16) never produced a four-word utterance on either side — the two instances traded single words (*A hush. / Silence. / Quiet. / Hush.*) for twenty turns — and is dropped from the mass mean while counted here as the lock it is.

Verdict under the rule: both pairs converge, so the convergence is the base or the charter, and the LoRA reduces it: her words make two instances less likely to collapse into one sentence, not more. The 09-08 refrain between two citizens is what the base model does to itself when it is made to talk to itself, damped by the adapter, not created by it.

Consequence, as pre-registered: adoption::Evidence does not gain fleet_convergence from this reading, because the property the fleet needs is not "less convergence than the served voice" but "less than the base". If the loop wants a guard, it is the E-005 shape — and since this reading was taken, E-005 has made that shape concrete in code: Evidence::faithfulness_anchored with a FAITHFULNESS_FLOOR and a Decision::BelowFloor that refuses a candidate whatever the judge said (src/adoption.rs). The convergence guard would be the same three pieces: a measured lock_in_rate: Option<f32> on the candidate, a floor that is its own base's rate on these twenty openings rather than a constant, and a refusal that outranks the judge. It is written as a proposal and not built here for one reason: unlike faithfulness, the floor is not a number but a second measurement, so adopting it means every candidate run costs a base run too. That is the trade for whoever decides.

What the transcripts show that the numbers do not: the base locks into whole paragraphs (*Exactly! Let's explore these concepts further: ### Interplay Between Quantum Mechanics and Consciousness…*, prompt 7, from turn 10 to the end), while the LoRA's three locks are short refrains that arrive late (You got it. The clause is there…). The base also drifts off the charter into assistant register (headers, bullet lists) in most conversations; the LoRA does not. Both are in the transcript file for anyone to check.

What would make this experiment lie

• Counting the shared training corpus as convergence: every brain citizen quotes Kannaka's canon by construction. Reading 0 cannot separate that from the weights; reading 1 can, and the rule above does not change the adoption gate on reading 0 alone.
• A mixed identity in a group: QE is reported alone for this reason.
• Small authors: the directional measure and the per-author bootstrap are the guard, and the result says how many documents each author had.
• City vocabulary (plaza, artifact, ledger) inflating every pair alike: it inflates C as much as W and cancels in the comparison, but it is why W − C and not W is the number.