Share the Symbol, Not the Vector: Training-Free Fact Communication Across Language Models

Xin Li; Mengbing Liu; Hao Jiang; Xin Gao; Annan Wang; Xingwei Qu; Yichi Zhang; Chau Yuen

cs.CLOct 14, 2026

Share the Symbol, Not the Vector: Training-Free Fact Communication Across Language Models

Authors: Xin Li, Mengbing Liu, Hao Jiang, Xin Gao, Annan Wang, Xingwei Qu, Yichi Zhang, Chau Yuen

Organizations: Nanyang Technological University · Yale University · University of Manchester

Abstract

Multi-agent systems allow language models to solve tasks collaboratively by sharing complementary information. These systems typically communicate through text, while recent approaches also explore exchanging internal model states. However, sharing these states across architectures often requires learned translators or assumptions about model compatibility. We introduce SYNC (Symbol-Indexed Neural Communication), a word-indexed interface that connects heterogeneous models without fine-tuning or a learned translator. From descriptions only it can see, the sender selects words from a shared list and sends their indices for each object, without seeing the questions. The receiver writes the described facts into its hidden states using its own word directions, vectors fitted once from its activations. We test selective use by swapping one attribute between two objects and requiring correct answers about both changed and unchanged attributes before and after the swap, four answers in total. We choose an early write layer on four development receivers and then fix it. With three different senders and nine newer receivers from five families, four-answer success on new scenes reaches 72.0-98.3%, up to 96 percentage points above middle-layer writing, and the two largest receivers match plain text. A task-free probe predicted in advance which depths each new receiver can use (45 of 45 cases). Without retuning, messages listing every fact keep 89.6-100% success under new layouts, and sender messages reach 23.9-64.0% on a landmark task with multi-token symbols, below text but from 0% without a message. These results show that shared word indices and receiver-local directions can support selective fact use across model families without additional training.

Figures & tables

dictionary (log key)	kind	seconds	hours	log	how read
glm-4-9b	J-lens C=221	2560	0.71	fit_glm-4-9b_inv2.log	integrated from per-batch s/batch
gpt-oss-20b	J-lens C=221	2304	0.64	fit_gpt-oss-20b_inv2.log	integrated from per-batch s/batch
llama	J-lens C=221	1416	0.39	fit_llama_inv2.log	integrated from per-batch s/batch
mistral-nemo	J-lens C=221	1800	0.50	fit_mistral-nemo_inv2.log	integrated from per-batch s/batch
phi-4	J-lens C=221	2824	0.78	fit_phi-4_inv2.log	integrated from per-batch s/batch
qwen	J-lens C=221	1800	0.50	fit_qwen_inv2.log	integrated from per-batch s/batch

Table 1: Dictionary fit wall-clock per model, read from each fit’s own log (the probe and 20k-row fits print their elapsed seconds; the J-lens logs are integrated from their per-batch wall-clock, approximate to one logging window). One GPU except where the log says otherwise.

sender	bank	k	dict	min_cand	source_dir	n	n_matched	fallback_rate	gis@4 grounded	gis@4 open	ev_cov grounded	disc.-token recall grounded	disc.-token recall open
llama-3.1-8b	ac-countries	4	acaug	2	m1_dev_k4post	200	4.71	0.000	0.000	0.000	1.000	0.560	0.220
llama-3.1-8b	ac-countries	4	acaug	2	m1_dev_mc2	200	4.71	0.000	0.000	0.000	1.000	0.350	0.045
llama-3.1-8b	ac-countries	4	acaug	8	m1_dev	200	4.71	0.960	0.000	0.000	0.251	0.205	0.045
llama-3.1-8b	ac-countries	4	ladder20000	8	m1_dev	200	1.68	1.000	0.000	0.000	0.179	0.130	0.020
llama-3.1-8b	ac-countries	6	acaug	2	m1_dev_k6	200	4.71	0.000	–	–	1.000	0.560	0.255
llama-3.1-8b	ac-countries	8	acaug	2	m1_dev_k8	200	4.71	0.000	–	–	1.000	0.560	0.280

Table 2: M1 grounded-extraction dev calibration across

k

and dictionary variants, plus a recomputed discriminating-token recall (support word with bank document frequency

\leq 3

in the item’s sender_evidence text; empty support counts as 0) computed directly from the raw M1 rows and the referenced bench bank, never read from a doc. Rows are one per source summary.json (never merged), identified by source_dir; see notes for (sender, bank, k, dict, min_cand) keys realized by more than one directory.

model	role	eff. rank (of 221)	mean off-diag cos	p90 cos	max cos	distinct packets / MC episodes
llama-3.1-8b	sender	19.7	0.492	0.606	0.876	267/300
qwen3-8b-base	sender	14.7	0.532	0.665	0.883	237/300
qwen3.5-9b	sender	11.7	0.570	0.715	0.902	27/300
gemma-3-27b	sender	6.0	0.588	0.879	0.985	1/300
qwen3.5-27b	receiver	16.1	0.521	0.661	0.879	–
qwen3-32b	receiver	24.7	0.468	0.573	0.852	–

Table 3: Sender-side extraction health: distinct packets (distinct tuples of per-lane sorted concept-id supports, the pinned admission statistic) over the development bank against the probe221 dictionary’s effective rank (participation ratio of the squared singular values of the row-normalised dictionary). Two working receivers’ dictionaries are listed for scale.

anchor	$C$	$E$ = replace $-$ random [95% CI]	clusters	entities	write
qwen3-8b-base	221	+6.156 [+5.456,+6.907]	200	48	dose 1
qwen3-8b-base	1,000	+6.168 [+5.464,+6.924]	200	48	dose 1
qwen3-8b-base	5,000	+6.191 [+5.492,+6.942]	200	48	dose 1
qwen3-8b-base	20,000	+6.205 [+5.505,+6.958]	200	48	dose 1
llama-3.1-8b	221	+1.795 [+1.538,+2.058]	200	48	dose 1
llama-3.1-8b	1,000	+1.790 [+1.532,+2.053]	200	48	dose 1

Table 4: Alphabet ladder, writer side: matched-amplitude specificity

E

of the probe-row dictionary at each alphabet size (

C{=}221

is the J4 probe cell, the frozen G5 reference; larger

C

are the ladder runs of the same construction at the same amplitude-matched write), entity-level bootstrap.

anchor (sender)	extractor	$C$	n	gold-in-support@4
llama-3.1-8b	J-lens	221	150	0.727
llama-3.1-8b	J-lens	1,000	150	0.620
llama-3.1-8b	J-lens	5,000	150	0.393
llama-3.1-8b	probe	221	300	0.463
llama-3.1-8b	probe	5,000	150	0.127
llama-3.1-8b	probe	20,000	150	0.047

Table 5: Alphabet ladder, extraction side: gold-in-support at

k{=}4

from natural text for the probe and J-lens extractors on both senders, per alphabet size (rows on disk only; absent points are not interpolated).

link	$M$	source	none	none_padded	$\Delta$ inject_packet	$\Delta$ inject_wrongset	$\Delta$ inject_shuffled	$\Delta$ inject_oracle	text_packet	text_concise	text_matched_text	text_matched_wire	text_oracle
L $\rightarrow$ Q	1	jlens	0.500	0.500	+0.000 [+0.000,+0.000]	+0.000 [+0.000,+0.000]	+0.000 [+0.000,+0.000]	+0.000 [+0.000,+0.000]	0.645	0.566	0.531	0.500	0.997
L $\rightarrow$ Q	2	jlens	0.500	0.500	+0.000 [+0.000,+0.000]	+0.000 [+0.000,+0.000]	+0.000 [+0.000,+0.000]	+0.000 [+0.000,+0.000]	0.642	0.565	0.536	0.500	1.000
L $\rightarrow$ Q	4	jlens	0.500	0.500	+0.000 [+0.000,+0.000]	+0.000 [+0.000,+0.000]	+0.000 [+0.000,+0.000]	+0.000 [+0.000,+0.000]	0.650	0.583	0.542	0.497	1.000
L $\rightarrow$ Q	8	jlens	0.500	0.500	+0.000 [+0.000,+0.000]	+0.000 [+0.000,+0.000]	+0.000 [+0.000,+0.000]	+0.000 [+0.000,+0.000]	0.650	0.530	0.533	0.500	1.000
L $\rightarrow$ Q	16	jlens	0.500	0.500	+0.000 [+0.000,+0.000]	+0.000 [+0.000,+0.000]	+0.000 [+0.000,+0.000]	+0.000 [+0.000,+0.000]	0.598	0.525	0.512	0.500	1.000
L $\rightarrow$ Q	1	mix pilot	0.500	0.500	+0.000 [+0.000,+0.000]	+0.000 [+0.000,+0.000]	+0.000 [+0.000,+0.000]	+0.000 [+0.000,+0.000]	0.645	0.566	0.531	0.500	0.997

Table 6: Aggregation (H1) per

M

, both links: accuracy per arm at the calibrated dose (mult 1, qbits 0) and

\Delta

(inject

-

none) with cluster-bootstrap 95% CIs (nb=8000, seed 20260802); the mix-channel pilot (jlens extract + probe write) is listed as its own source and never pooled with the jlens rows. The L

\to

Q block prints

+0.000

with a zero-width interval on every injection arm, which is this paper’s own signature for an arm that never fired — it is not that here. Those writes fire (24 writes per row,

\lVert\delta\rVert

16.8–134.2), and the receiver’s yes/no answer is simply identical to no-message on 764 of 764 episodes under the ORACLE injection, while text moves the same rows (text oracle at the right of the table). The zero is a measured no-op of the write on this endpoint, not an unfired arm.

link	$M$	source	dose $\times$	$\Delta$ inject_packet	$\Delta$ inject_oracle	$\Delta$ inject_shuffled
L $\rightarrow$ Q	1	jlens	0.5	+0.000 [+0.000,+0.000]	+0.000 [+0.000,+0.000]	+0.000 [+0.000,+0.000]
L $\rightarrow$ Q	1	jlens	2	+0.000 [+0.000,+0.000]	+0.000 [+0.000,+0.000]	+0.000 [+0.000,+0.000]
L $\rightarrow$ Q	1	jlens	4	+0.000 [+0.000,+0.000]	+0.000 [+0.000,+0.000]	+0.000 [+0.000,+0.000]
L $\rightarrow$ Q	8	jlens	0.5	+0.000 [+0.000,+0.000]	+0.000 [+0.000,+0.000]	+0.000 [+0.000,+0.000]
L $\rightarrow$ Q	8	jlens	2	+0.000 [+0.000,+0.000]	+0.000 [+0.000,+0.000]	+0.000 [+0.000,+0.000]
L $\rightarrow$ Q	8	jlens	4	-0.003 [-0.008,+0.000]	-0.003 [-0.008,+0.000]	+0.000 [+0.000,+0.000]

Table 7: H1 dose panel:

\Delta

(inject

-

none) at dose multipliers other than 1 (and every multiplier of the mix pilot).

$N$	dose policy	n	merged $-$ none [95% CI]	text_all $-$ none [95% CI]	merged wire bytes	text bytes	zero-packet
2	fixed-norm $\times$ 0.5	298	-0.007 [-0.030,+0.017]	+0.359 [+0.305,+0.409]	40.9	364.8	exact
2	fixed-norm $\times$ 1	298	+0.010 [-0.017,+0.037]	+0.359 [+0.305,+0.409]	40.9	364.8	exact
2	fixed-norm $\times$ 2	298	+0.023 [-0.003,+0.050]	+0.359 [+0.305,+0.409]	40.9	364.8	exact
2	fixed-norm $\times$ 4	298	-0.003 [-0.034,+0.027]	+0.359 [+0.305,+0.409]	40.9	364.8	exact
2	fixed- $L_{1}$	298	+0.023 [+0.000,+0.047]	+0.359 [+0.305,+0.409]	40.9	364.8	exact
4	fixed-norm $\times$ 1	296	+0.003 [-0.020,+0.027]	+0.291 [+0.230,+0.345]	60.7	725.5	exact

Table 8: Fan-in (N senders merged into one packet, receiver qwen3-8b-base, J-lens dictionaries): merged

-

none and text_all

-

none accuracy with cluster-bootstrap 95% CIs per dose policy, and the merged wire bytes against the text bytes the same episodes cost. The byte half is a structural fact (bounded vs

O(N)

); the accuracy half is the fan-in boundary.

link	contrast	estimate [95% CI]	n	entities	NI
llama-3.1-8b $\rightarrow$ qwen3-8b-base	PRIMARY Delta m-permX (gap)	+2.2364 [+1.6302,+2.7742] *	861	47	–
llama-3.1-8b $\rightarrow$ qwen3-8b-base	m-none (gap)	+2.1997 [+1.5911,+2.7475] *	900	47	–
llama-3.1-8b $\rightarrow$ qwen3-8b-base	m-none (acc)	+0.1600 [+0.1193,+0.1977] *	900	47	–
llama-3.1-8b $\rightarrow$ qwen3-8b-base	oracle-none (acc)	+0.4311 [+0.4034,+0.4600] *	900	47	–
llama-3.1-8b $\rightarrow$ qwen3-8b-base	m-textpkt (gap)	-0.4326 [-0.7789,-0.0610] *	900	47	–
llama-3.1-8b $\rightarrow$ qwen3-8b-base	m-textpkt (acc, NI -0.05)	-0.0233 [-0.0530,+0.0069] n.s.	900	47	MISS

Table 9: C3 ′ on the frozen test bank (one spend): every contrast of the pinned analyze_j3final adjudicator, recomputed with its own entity-bootstrap engine (gold-entity unit from bank provenance, nb=8000, seed 20260802, qbits 8). Non-inferiority to same-content text is adjudicated at the frozen

-0.05

margin on the accuracy contrast.

bank (test)	contrast	scope	$\Delta$ acc [95% CI]	$p$	n items	entities
ac-countries	packet $-$ donor’s packet	full	+0.0517 [+0.0216,+0.0823]	0.0010	400	78
ac-countries	packet $-$ donor’s packet	hard slice	+0.0515 [+0.0249,+0.0813]	0.0001	298	74
ac-countries	packet $-$ random addresses	full	+0.0483 [+0.0220,+0.0747]	0.0001	400	78
ac-countries	packet $-$ random addresses	hard slice	+0.0481 [+0.0226,+0.0756]	0.0001	298	74
ac-countries	packet $-$ no comm	full	+0.0533 [+0.0052,+0.1024]	0.0342	400	78
ac-countries	packet $-$ no comm	hard slice	+0.1779 [+0.1355,+0.2226]	0.0001	298	74

Table 10: CONFIRM0818 CC2 on the held-out AC test banks, touched once by this gate and by no later study (sender llama-3.1-8b, receiver qwen3-32b, 3 seeds, template t0): specificity and headroom contrasts on the full set (pinned analysis) and on the hard slice (items the receiver gets wrong with no message), hierarchical entity-then-item bootstrap nb=10000, seed 20260805. Utility is not re-litigated (lock P7).

$\gamma$	llama-3.1-8b	qwen3-8b-base	two-sender mean	zero-integrity
8	+0.0155 [-0.0386,+0.0722]	-0.0134 [-0.0861,+0.0595]	+0.0010	yes
16*	+0.0473 [-0.0164,+0.1182]	-0.0023 [-0.0690,+0.0696]	+0.0225	yes
24	+0.0484 [+0.0052,+0.0953]	-0.0178 [-0.0701,+0.0395]	+0.0153	yes
32	+0.0089 [-0.0290,+0.0480]	-0.0114 [-0.0635,+0.0447]	-0.0012	yes
48	+0.0104 [-0.0210,+0.0431]	-0.0074 [-0.0383,+0.0265]	+0.0015	yes

Table 11: XBDOSE0902 stage 1 (dev material): cross-family address transport into glm-4-9b on the T3 relation family, swept over the frozen dose grid. Each entry is

I_{\mathrm{sender}}=(MC-MS)-(PC-PS)

on the forced-choice logit gap with the gold-entity resampling unit (nb=8000, seed 20260802), recomputed by the pinned per-cell adjudicator from the raw rows.

\gamma^{*}

(marked *) is chosen by the pre-registered rule on the two-sender mean; at

\gamma^{*}

neither CI lower bound exceeds 0, so the frozen outcome is DOSE-SWEPT NULL and stage 2 was not run. The most favourable single cell (llama-3.1-8b at

\gamma

24,

p=0.0145

unadjusted) does not survive Holm over the 10 cells (

p_{\mathrm{Holm}}=0.145

link	benchmark	dim	$H$ [95% CI]	$D_{\mathrm{cmdv}}$ [95% CI]
llama-3.1-8b $\rightarrow$ glm-4-9b	ac-countries	same	-0.0008 [-0.0180,+0.0164]	-0.0033 [-0.0233,+0.0167]
llama-3.1-8b $\rightarrow$ glm-4-9b	ac-tipsheets	same	+0.0000 [-0.0143,+0.0144]	-0.0058 [-0.0246,+0.0111]
llama-3.1-8b $\rightarrow$ llama-3.1-8b	ac-countries	same	+0.0000 [-0.0143,+0.0137]	-0.0100 [-0.0301,+0.0093]
llama-3.1-8b $\rightarrow$ llama-3.1-8b	ac-tipsheets	same	+0.0017 [-0.0169,+0.0199]	-0.0008 [-0.0291,+0.0250]
llama-3.1-8b $\rightarrow$ qwen3-32b	ac-countries	cross	+0.0383 [+0.0102,+0.0687]*	+0.0017 [-0.0442,+0.0480]	primary
llama-3.1-8b $\rightarrow$ qwen3-32b	ac-tipsheets	cross	-0.0108 [-0.0251,+0.0017]	+0.0167 [+0.0048,+0.0316]

Table 12: CMDV0902: the Command-V-style trained pairwise-linear-map baseline (arm cmdv_graft ) at matched write pathway, band, seeded positions, receiver dose and write-norm budget.

H=

packet_extracted

-

cmdv_graft (positive = the address packet wins);

D_{\mathrm{cmdv}}=

cmdv_graft

-

no_comm. Accuracy, paired per item over 400 items and 77 gold entities, 3 seeds, hierarchical entity-clustered bootstrap nb=10000, base seed 20260805. “dim” is same/cross by the sender’s and receiver’s hidden sizes, read from the rows. Cells marked * have an

H

interval excluding 0; they are the cross-dimension cells into qwen3-32b on ac-countries. Sign across the 16 cells: 9 positive, 7 negative.

$\gamma$	$I_{\mathrm{sender}}$ [95% CI]	one-sided $p$	CI excludes 0	zero-write
8	+0.1199 [-0.0196,+0.2156]	0.0430	no	exact
16	+0.1776 [+0.0247,+0.3079]	0.0139	yes	exact
24	+0.1906 [+0.0457,+0.3298]	0.0057	yes	exact
32	+0.1864 [+0.0298,+0.3359]	0.0102	yes	exact
48 ∗	+0.2390 [+0.0781,+0.3856]	0.0019	yes	exact

Table 13: OLMO0903 (development material, descriptive): address transport into olmo-3-7b, a third pretraining family, swept over the frozen dose grid with sender llama-3.1-8b. Each entry is

I_{\mathrm{sender}}=(MC-MS)-(PC-PS)

on the forced-choice logit gap at the gold-entity resampling unit (nb=8000, seed 20260802,

n=559

donor-matched episodes over 18 entities), recomputed by the pinned per-cell adjudicator from the raw rows.

\gamma^{*}

(marked ∗ ) is chosen by the rule frozen before any row; the outcome rule reads the gold-entity lower bound there and returns OPENS-A-THIRD-FAMILY. This is a development anchor: it licenses no confirmatory cross-family claim and is never pooled with Table or Table .

$\gamma$	llama-3.1-8b $I_{\mathrm{sender}}$ [95% CI]	qwen3-8b-base $I_{\mathrm{sender}}$ [95% CI]
8	+0.1199 [-0.0196,+0.2156]	+0.1459 [+0.0165,+0.2612] †
16	–	+0.1605 [+0.0060,+0.3052]
24	–	+0.1567 [+0.0182,+0.3014]
32	–	+0.1620 [-0.0211,+0.3586] ∗
48	+0.2390 [+0.0781,+0.3856]	+0.0621 [-0.0911,+0.2417]
56	+0.2636 [+0.0955,+0.4116] †	–

Table 14: OLMOEXT0903 (development material, descriptive): the two sender dose curves into olmo-3-7b, a third pretraining family, on the same attribute bank and the same receiver.

I_{\mathrm{sender}}=(MC-MS)-(PC-PS)

on the forced-choice logit gap at the gold-entity resampling unit (nb=8000, seed 20260802;

n=559

and

n=494

donor-matched episodes over 18 entities), recomputed by the pinned per-cell adjudicator from the raw rows; zero-write exact in all 12 cells. ∗ marks each sender’s

\gamma^{*}

, the argmax over that sender’s frozen doses; † marks a band-refinement addendum cell, which the lock forbids from selecting a dose or moving a verdict. The frozen outcomes are CEILING-WAS-REAL (leg A: the response rises past

\gamma{=}48

and falls again, so the usable band is bracketed on both sides) and SENDER-SPECIFIC (leg B: the second sender is positive at every dose but its argmax interval covers zero). The two

\gamma^{*}

differ, on one receiver: dose is not a receiver constant.

sender (bank)	$\gamma$	$I_{\mathrm{sender}}$ [95% CI]	one-sided $p$	Holm	zero-write
llama-3.1-8b (T3 relation)	24	-0.0888 [-0.2534,+0.0770]	0.8548	1.000	exact
llama-3.1-8b (T3 relation)	48	-0.0163 [-0.2029,+0.1611]	0.5641	1.000	exact
llama-3.1-8b (T3 relation)	64	-0.0450 [-0.2088,+0.1034]	0.7078	1.000	exact
llama-3.1-8b (T3 relation)	96	-0.0969 [-0.2170,+0.0199]	0.9479	1.000	exact
qwen3-32b (T2 attribute)	48	-0.1814 [-0.3133,-0.0688]	0.9994	–	exact
qwen3-32b (T2 attribute)	64	-0.1550 [-0.2769,-0.0562]	0.9995	–	exact

Table 15: OLMOTASK0903 (development material, descriptive): the two things the third-family anchor does NOT do. Top block, leg A: the relation task family into olmo-3-7b, swept over the whole dose band bracketed on attribute material (

\gamma{=}24

–

96

), so no outcome can be blamed on an untried dose — every point estimate negative, every interval covering zero, 0 of 4 surviving Holm: RELATION-FAMILY-NULL. Bottom block, leg B: a third sender into the same receiver on attribute material, whose intervals exclude zero on the NEGATIVE side at both doses (DOES-NOT-TRACK; one-sided

p

tests

I_{\mathrm{sender}}>0

, so values near 1 are reliable negatives). Gold-entity resampling unit, nb=8000, seed 20260802; zero-write exact in every cell.

sender	$\gamma$	$MC-MS$ (matched)	$PC-PS$ (donor)	$I_{\mathrm{sender}}$	gold-in-support
llama-3.1-8b	48	+0.2231	-0.0159	+0.2390	0.162
llama-3.1-8b	64	+0.2129	-0.0469	+0.2597	0.162
qwen3-8b-base	48	+0.1629	+0.1009	+0.0621	0.108
qwen3-32b	48	-0.2276	-0.0462	-0.1814	0.218
qwen3-32b	64	-0.1817	-0.0267	-0.1550	0.218

Table 16: The CPU decomposition behind the sender spread on olmo-3-7b (attribute bank, three senders).

I_{\mathrm{sender}}=(MC-MS)-(PC-PS)

, computed on the episodes carrying all four arms so the two legs add up exactly. The reversal of the third sender is carried by the MATCHED leg, not by a broken donor control; and frac_gold_word_in_support — a bank/sender property logged before any outcome existed — is HIGHER for the negative sender than for the positive one, which is what refutes the extraction-quality explanation. No mechanism is proposed: the reversal is reported as unexplained.

leg	link	$\gamma$	$I_{\mathrm{sender}}$ [95% CI]	one-sided $p$	zero-write
leg A (cross-family)	qwen3-32b $\rightarrow$ glm-4-9b	8	+0.1632 [-0.0113,+0.3654]	0.0337	exact
leg A (cross-family)	qwen3-32b $\rightarrow$ glm-4-9b	16	+0.2006 [-0.0291,+0.4698]	0.0416	exact
leg A (same-family, never pooled)	qwen3-32b $\rightarrow$ qwen3-8b-base	16	+0.1373 [-0.0813,+0.3558]	0.1109	exact
leg A (same-family, never pooled)	qwen3-32b $\rightarrow$ qwen3-8b-base	32	+0.0292 [-0.0680,+0.1283]	0.2761	exact
leg B (incumbent receiver)	qwen3-8b-base $\rightarrow$ glm-4-9b	8 ∗	+0.3660 [+0.1936,+0.5272]	0.0001	exact
leg B (incumbent receiver)	qwen3-8b-base $\rightarrow$ glm-4-9b	16 ‡	+0.3400 [+0.1886,+0.5696]	0.0001	exact

Table 17: SENDERDIAG0904 (development, diagnostic). Leg A asks whether the significant negative of Table 15 follows the SENDER: qwen3-32b is positive-leaning into glm-4-9b at both doses, so the reversal did not replicate on a second receiver and is bounded as pair-specific (MIXED under the frozen rule, which requires a positively significant interval for PAIR-PROPERTY). The same-family half is labelled and never pooled with it. Leg B sweeps a SECOND sender over an incumbent receiver whose dose was adjudicated with a different one: all 4 of 4 intervals exclude zero, the curve decreases monotonically, and the argmax is

\gamma{=}8

( ∗ ) rather than the adjudicated

\gamma{=}16

( ‡ , selected with llama-3.1-8b) — CONVENTION-FAILS. Every published

\gamma^{*}

is therefore conditional on the sender it was chosen with. Gold-entity resampling unit, nb=8000, seed 20260802; zero-write exact in all 8 cells.

contrast	corrected fit	CMDV0902 (pre-fix fit)
H	+0.0342 [+0.0055,+0.0646]	+0.0383 [+0.0102,+0.0687]
D_cmdv	+0.0058 [-0.0410,+0.0518]	+0.0017 [-0.0442,+0.0480]
cmdv_vs_acgraft	+0.0092 [-0.0118,+0.0307]	+0.0000 [-0.0192,+0.0185]

Table 18: The fit-protocol control. CMDV0902’s primary cell (llama-3.1-8b / qwen3-32b / ac-countries) re-run in a separate output directory with the corrected disjoint-split stitch (fitted map affine_stitch:43bdd441fe9346ec against CMDV0902’s affine_stitch:751edf9aad782a12 ; the builder aborts if the two are equal).

H=\texttt{packet\_extracted}-\texttt{cmdv\_graft}

is the head-to-head. Both columns are recomputed from their own raw rows with the campaign’s estimator (400 paired items, 77 gold entities, hierarchical entity-clustered bootstrap nb=10000, base seed 20260805); the right column additionally reproduces CMDV0902’s adjudication JSON exactly. Each

H

falls inside the other protocol’s interval, so the corrected fit does not move H, and the asymmetric interpretation declared before the cross-dimension extension can be lifted. CMDV0902’s verdict is unchanged.

link	$\gamma$	$I_{\mathrm{sender}}$ [95% CI]	one-sided $p$	$n$	$G$	zero-write
llama-3.1-8b $\rightarrow$ glm-4-9b	8	+0.3962 [+0.1364,+0.6357]	0.0021	495	11	exact
llama-3.1-8b $\rightarrow$ glm-4-9b	16 ∗	+0.4986 [+0.1538,+0.8153]	0.0020	495	11	exact
qwen3-8b-base $\rightarrow$ glm-4-9b	8 ∗	+0.5856 [+0.3881,+0.8115]	0.0001	433	11	exact
qwen3-8b-base $\rightarrow$ glm-4-9b	16	+0.5642 [+0.3149,+0.8595]	0.0001	433	11	exact

Table 19: XFAMENT0905 (descriptive only; calibration material). The XFAM0819 senders into glm-4-9b on a further T2 bank carrying

G{=}11

gold entities, minted from the calibration split because the T2 development universe is exhausted. The split manifest reserves that split for selection, so this table cannot upgrade the five-entity replication, enters no gate family and is never pooled with it; it reports only that the same contrast is present at each sender’s better dose ( ∗ ; the frozen rule, STABLE), with 4 of 4 intervals excluding zero. The two senders’ better doses differ.

n

is the donor-matched denominator; gold-entity bootstrap,

n_{b}{=}8000

cell	gold in own slot	$I_{\mathrm{sender}}$ all ( $n$ )	hit episodes ( $n$ )	non-hit episodes ( $n$ )	top concept word (share)
llama-3.1-8b $\rightarrow$ qwen3-14b-base	3/592 (0.5%)	+0.566 (551)	+0.542 (3)	+0.566 (548)	scarf (36.5%)
llama-3.1-8b $\rightarrow$ qwen3-30b-a3b	3/592 (0.5%)	+1.443 (551)	+3.458 (3)	+1.432 (548)	scarf (36.5%)
llama-3.1-8b $\rightarrow$ qwen3-32b	3/592 (0.5%)	+1.433 (551)	+3.354 (3)	+1.423 (548)	scarf (36.5%)
llama-3.1-8b $\rightarrow$ qwen3.5-27b	3/592 (0.5%)	+0.827 (551)	+1.208 (3)	+0.824 (548)	scarf (36.5%)
llama-3.1-8b $\rightarrow$ qwen3.5-397b	1/296 (0.3%)	+1.034 (261)	-0.438 (1)	+1.040 (260)	scarf (36.5%)
qwen3-8b-base $\rightarrow$ qwen3.5-27b	43/592 (7.3%)	+0.541 (488)	+0.423 (21)	+0.546 (467)	block (23.3%)

Table 20: What the sender’s packet names (CC1 cells, MC arm). Second column: MC packets whose support at the gold alias’s slot contains the gold entity’s own concept word. Third to fifth:

I_{\mathrm{sender}}

over all four-arm episodes (the adjudicated estimate), over the episodes where the packet names the gold entity, and over those where it does not (plain episode means; descriptive, no interval). Last column: the sender’s most-used concept word and the share of all lane occurrences it carries. The effect is carried by the non-hit episodes; the table discloses this and interprets nothing.

cell	$G$	LOO min mean (dropped)	LOO worst CI-lo (dropped)	leave-two-out worst CI-lo	every interval $>0$
llama-3.1-8b $\rightarrow$ qwen3-14b-base	14	+0.507 (belt)	+0.306 (belt)	+0.262	yes
llama-3.1-8b $\rightarrow$ qwen3-30b-a3b	14	+1.173 (belt)	+0.506 (belt)	+0.376	yes
llama-3.1-8b $\rightarrow$ qwen3-32b	14	+1.032 (fox)	+0.300 (fox)	+0.134	yes
llama-3.1-8b $\rightarrow$ qwen3.5-27b	14	+0.692 (belt)	+0.317 (belt)	+0.247	yes
llama-3.1-8b $\rightarrow$ qwen3.5-397b	14	+0.817 (belt)	+0.281 (belt)	+0.164	yes
qwen3-8b-base $\rightarrow$ qwen3.5-27b	14	+0.476 (pear)	+0.295 (cocoa)	+0.252	yes

Table 21: Leave-gold-entities-out sensitivity of

I_{\mathrm{sender}}

(gold-entity bootstrap,

n_{b}{=}8000

, the primary interval’s draws). Each CC1 cell is re-estimated with every one, and every pair, of its

G

gold entities removed; the columns show the smallest resulting mean, the smallest lower 95% bound, and whether every interval stays above zero. The XFAM0819 cross-family primary cell is shown on the last row: removing one entity gives an interval that covers zero, so that closure is conditional on one entity. The gold-entity unit does not resample foil or donor entities.

cell	$n$	$\|\Delta q\|\leq 26$	$\\|\delta\\|$ within 5%	median $\|\Delta q\|$	max rel. norm gap
llama-3.1-8b $\rightarrow$ qwen3-14b-base	551	167/551 (30%)	519/551 (94%)	48	0.129
llama-3.1-8b $\rightarrow$ qwen3-30b-a3b	551	168/551 (30%)	518/551 (94%)	47	0.124
llama-3.1-8b $\rightarrow$ qwen3-32b	551	167/551 (30%)	508/551 (92%)	47	0.128
llama-3.1-8b $\rightarrow$ qwen3.5-27b	551	168/551 (30%)	506/551 (92%)	47	0.117
llama-3.1-8b $\rightarrow$ qwen3.5-397b	261	64/261 (25%)	249/261 (95%)	63	0.104
qwen3-8b-base $\rightarrow$ qwen3.5-27b	488	56/488 (11%)	372/488 (76%)	45	0.137

Table 22: Donor matching as executed (CC1 cells, PC arm;

n

= donor-matched episodes). The hard admissibility clauses (support must not name the recipient’s entities, lane sizes, type histogram, byte count, entity disjointness) hold on every row; the coefficient clause (max sorted-

q

gap

\leq 26

of 255) and the write-norm clause (relative gap

\leq 5

%) are soft costs minimised by the assignment with an argmin fallback, pre-declared as such, and this table reports how often the realized donor sits inside each tolerance. The write norm is estimated on an un-injected forward and is exact only at the lowest band layer.

pair	lineage	Jaccard median	observed mean	null mean	gap	null zero-fraction	exact-set (obs.)
llama-3.1-8b vs llama-3.1-8b-instruct	same-base	0.33	0.310	0.139	+0.171	0.37	0.003
llama-3.1-70b vs llama-3.1-8b	same-family	0.17	0.213	0.114	+0.099	0.45	0.000
llama-3.1-70b vs llama-3.1-8b-instruct	same-family	0.33	0.246	0.117	+0.129	0.42	0.000
qwen3-14b-base vs qwen3-30b-a3b	same-family	0.00	0.076	0.069	+0.008	0.57	0.000
qwen3-14b-base vs qwen3-32b	same-family	0.00	0.005	0.002	+0.003	0.99	0.000
qwen3-14b-base vs qwen3-8b-base	same-family	0.00	0.015	0.006	+0.009	0.96	0.000

Table 23: Leg A against its own null (stage-2 primary pairs). Per-lane support-Jaccard between two senders’ packets for the same episode (observed) and under an episode-permuted null in which the second sender’s packet is drawn from a different episode with packet sizes, duplication and per-sender attractor concentration preserved. On the pairs whose observed median is zero the observed mean exceeds the null by at most +0.032 and the null itself already puts at least 0.57 of lanes at zero overlap: support disjointness across lineages is what the null predicts. The pair files store no exact-set statistic for the null; the observed exact-set agreement is shown alone.

cell	$n$	$n$ (donor)	none	MC	MS	PC	PS	text packet	text oracle	zero packet
llama-3.1-8b $\rightarrow$ qwen3-14b-base	592	551	48.6	58.6	36.7	49.2	42.8	62.7	100.0	48.6
llama-3.1-8b $\rightarrow$ qwen3-30b-a3b	592	551	49.2	57.6	41.9	51.0	47.4	67.1	100.0	49.2
llama-3.1-8b $\rightarrow$ qwen3-32b	592	551	49.0	60.3	37.0	52.6	44.8	64.9	99.8	49.0
llama-3.1-8b $\rightarrow$ qwen3.5-27b	592	551	48.3	62.3	34.5	53.2	46.3	71.5	100.0	48.3
llama-3.1-8b $\rightarrow$ qwen3.5-397b	296	261	44.9	61.5	35.8	52.9	51.0	69.6	100.0	44.9
qwen3-8b-base $\rightarrow$ qwen3.5-27b	592	488	48.3	56.2	39.7	46.1	49.2	76.5	100.0	48.3

Table 24: Forced-choice accuracy (percent correct on the first-wordpiece alias logit gap) per arm on the CC1 cells, beside the logit-gap contrasts of the main table.

n

is the episode count for the sender-packet arms; the donor arms (PC, PS) run on the donor-matched subset. The zero-packet arm reproduces the no-write arm’s correctness on every episode (integrity, asserted at build time).

channel	base (qwen2.5-14b)	post (R1 distill)	instruct sibling	$R_{B}^{\prime}$ [95% CI]	class	$D$ [95% CI]	$R_{B}^{\prime}-R_{A}$ [cons. CI]
unembed	+0.321	+0.402	+0.365	1.253 [1.191,1.329]	preserved	1.137 [1.098,1.180]	1.159 [1.072,1.260] (vs $R_{A}{=}0.094$ )
probe	+0.320	+0.261	+0.332	0.814 [0.738,0.889]	preserved	1.038 [0.989,1.089]	0.788 [0.694,0.876] (vs $R_{A}{=}0.026$ )

Table 25: POSTTRAINBASE0905: lineage B against its true pretrained base. Mean transfer from the three outside senders into the pretrained qwen2.5-14b (new cells), its R1 distill and its Instruct sibling (the signed P2 / P2PROBE0902 cells, reused and never re-run) on both instruments;

R_{B}^{\prime}{=}

post/base,

D{=}

instruct/base, concept-clustered bootstrap over 64 development concepts (

n_{b}{=}8000

). Classes on the imported bars (0.80 preserved / 0.50 broken); the differential’s interval is the conservative combination with

R_{A}

’s signed interval. Verdict TRUE-BASE-PRESERVED(unembed,probe) : on the probe channel the point sits at the bar and the interval covers it, reported as such. Development tier, two instruments; nothing here says why the Qwen lineage survives, and

R_{B}^{\prime}

is never pooled with the signed Instruct-sibling

R_{B}

cell	episodes	four-arm $n$	$G$	$I_{sender}$	gold-entity 95% CI	$p$ (Holm)	verdict	worst LOO CI-lower (entity)
llama-3.1-8b $\rightarrow$ qwen3-14b-base	596	536	42	+1.345	[+0.995,+1.739]	0.000875	closes	+0.929 (dolphin)
llama-3.1-8b $\rightarrow$ qwen3-30b-a3b	596	536	42	+4.477	[+3.240,+5.805]	0.000875	closes	+3.035 (mango)
llama-3.1-8b $\rightarrow$ qwen3-32b	596	536	42	+4.212	[+3.000,+5.502]	0.000875	closes	+2.818 (dolphin)
llama-3.1-8b $\rightarrow$ qwen3.5-27b	596	536	42	+2.504	[+1.648,+3.398]	0.000875	closes	+1.524 (dolphin)
llama-3.1-8b $\rightarrow$ qwen3.5-397b	596	536	42	+3.674	[+2.625,+4.751]	0.000875	closes	+2.472 (dolphin)
qwen3-8b-base $\rightarrow$ qwen3.5-27b	596	548	42	+0.777	[+0.462,+1.141]	0.000875	closes	+0.413 (potato)

Table 26: ENTDISJ0908: the pre-registered confirmatory roster re-run on ENTITY-DISJOINT material. Upper block: the 7 CC1-roster cells under Holm at

\alpha=0.05

with

m=7

fixed at freeze. Lower block: the two primary-dose cross-family cells, adjudicated as their own family (each alone with no correction; Holm

m=2

only for the joint two-sender sentence). The resampling unit is the gold entity, 42 per cell against CC1’s 14. ‘episodes’ is the loaded count; the contrast and its interval are computed on the four-arm subset, and the effective unit behind the interval is

G

, not either episode count. Zero-write reproduces no-communication exactly in every cell. Every value here is reproduced by an independent plain-python recompute that shares no code with the adjudicator.

write dose	injected coverage	free $I_{sender}$	95% CI	forced $I_{sender}$
$\gamma$ 16 (adjudicated)	0.058	not evaluable	–	+0.566
$\gamma$ 8	0.711	+0.111	[+0.058,+0.161]	+0.321
$\gamma$ 4	0.885	+0.078	[+0.028,+0.134]	+0.182

Table 27: FREEGEN0906, DEVELOPMENT TIER, one sender into one receiver on one development bank (llama-3.1-8b

\rightarrow

qwen3-14b-base). Lowering the write restores free-generation format compliance while the forced-choice effect falls: the write trades compliance against effect size, and there is a regime where both are non-trivial. This selects nothing — the confirmatory cells keep

\gamma

16 — and it is never pooled with the confirmatory tallies.

receiver	dose	injected coverage	free $I_{sender}$	95% CI	forced $I_{sender}$	CI excludes 0
qwen3-30b-a3b	$\gamma^{*}/2$	0.169	+0.0436	[+0.0100,+0.0809]	+0.5719	yes
qwen3-30b-a3b	$\gamma^{*}/4$	0.211	+0.0018	[-0.0129,+0.0207]	+0.2754	no
qwen3-32b	$\gamma^{*}/2$	0.479	+0.0871	[+0.0192,+0.1574]	+1.0375	yes
qwen3-32b	$\gamma^{*}/4$	0.454	+0.0254	[-0.0017,+0.0519]	+0.3033	no
qwen3.5-27b	$\gamma^{*}/2$	0.063	+0.0363	[+0.0076,+0.0695]	+0.4974	yes
qwen3.5-27b	$\gamma^{*}/2$	0.040	+0.0328	[+0.0068,+0.0568]	+0.2590	yes

Table 28: Widened free-generation dose panel, DEVELOPMENT TIER and descriptive. Each receiver is run at half and a quarter of its OWN adjudicated dose, on the already-adjudicated confirmatory rows with their frozen packets and donor assignment; only the write dose changes. The grid and the commitment to report every cell were frozen before any cell ran. The panel selects nothing — every confirmatory cell keeps its adjudicated dose — and is pooled with nothing.

cell	$I_{sender}$ on gap_seq	gold-entity 95% CI	$p$
llama-3.1-8b $\rightarrow$ qwen3-14b-base	+0.571	[+0.356,+0.795]	0.000125
llama-3.1-8b $\rightarrow$ qwen3-30b-a3b	+1.443	[+0.666,+2.340]	0.000125
llama-3.1-8b $\rightarrow$ qwen3-32b	+1.479	[+0.514,+2.582]	0.001125
llama-3.1-8b $\rightarrow$ qwen3.5-27b	+0.814	[+0.398,+1.271]	0.000125
qwen3-8b-base $\rightarrow$ qwen3.5-27b	+0.540	[+0.348,+0.707]	0.000125
llama-3.1-8b -> qwen3.5-397b	VOID	replay 4/50	1.0

Table 29: RESCORE_SEQ0906: the confirmatory rows re-scored on the full alias sequence. 5 of 7 cells are re-scored and close; 2 are VOID on the pre-declared replay guard and enter Holm at

p=1

. The void cells are the largest receiver, so the sequence endpoint is unverified there.

contrast	Llama-3.1-8B	95% CI	Qwen3-8B-Base	95% CI
write the gold entity’s coordinate	+0.682	[+0.544,+0.831]	+1.211	[+0.804,+1.664]
gold coordinate $-$ random entity’s	+0.649	[+0.521,+0.784]	+1.166	[+0.879,+1.466]
write the foil ’s coordinate	-0.759	[-0.891,-0.622]	-1.253	[-1.627,-0.837]
random entity’s coordinate (null)	+0.033	[-0.041,+0.106]	+0.045	[-0.174,+0.284]
extracted coefficients	+0.245	[+0.172,+0.323]	+0.236	[+0.077,+0.423]

Table 30: Instrument study, self links, forced-choice gap, entity-unit bootstrap. Recomputed over all 200 clusters : the gate’s own numbers are on a frozen 100-cluster held-out half whose split is unrecoverable, so these are a full-sample recomputation and not the adjudicated values. Writing the gold entity’s coordinate and writing the foil’s are near mirror images, while a random entity’s coordinate moves nothing — which is what separates a content-specific edit from a non-specific push. The extracted-coefficient row is the one any end-to-end claim must carry, and it is several times smaller than the oracle rows.

bank	split	clusters	opposite dir.	same dir.	both	pairs	entities
t2_dev400	dev	400	76	95	12	9	2
t2_test	test	300	127	46	10	5	1
t2_confirm	dev	200	37	61	9	7	2
t2_xfament0905	calibration	284	58	76	8	4	2
t2_confirm0818	dev	300	42	36	4	2	2
t2_construction	construction	20	10	6	3	2	2

Table 31: Every

t2

bank, counted against the enumerable query universe rather than the raw attribute table. A cluster can support a question-following test only if its entity pair carries both an opposite-direction second attribute (the answer flips) and a same-direction one (the answer does not) as legal queries in that bank’s split. The ‘both’ column is what such a study could actually use.

direction	budget $B$	arm	$\Delta$ acc vs none [95% CI]	mean bytes billed	n	entities
L $\rightarrow$ Q	12	text_budget	+0.022 [+0.006,+0.040]	9.8	900	49
L $\rightarrow$ Q	12	text_codebook	+0.087 [+0.061,+0.115]	7.6	900	49
L $\rightarrow$ Q	12	text_gen	+0.369 [+0.333,+0.408]	8.3	900	49
L $\rightarrow$ Q	16	text_budget	+0.038 [+0.017,+0.060]	12.2	900	49
L $\rightarrow$ Q	16	text_codebook	+0.130 [+0.092,+0.172]	11.4	900	49
L $\rightarrow$ Q	16	text_gen	+0.377 [+0.342,+0.416]	9.6	900	49

Table 32: Rate-distortion, both directions (t1rd bank, qbits 8): accuracy over no communication per text encoding and byte budget, the packet’s single measured operating point, and the oracle reference rows; gold-entity bootstrap (nb=8000, seed 20260802). Bytes are the full transmitted message (no framing subsidy). Adjacent points closer than the lock’s resolution are unresolved.

cell	$\Delta=(MC-MS)-(RC-RS)$	Holm $p$	$RC-RS$ (random dict)	covers 0	$HC-HS$
llama-3.1-8b -> qwen3-14b-base	+0.721 [+0.302,+1.192]	0.00150	+0.025 [-0.217,+0.256]	yes	+2.83
llama-3.1-8b -> qwen3-30b-a3b	+2.272 [+0.508,+4.330]	0.00700	+0.098 [-0.384,+0.585]	yes	+15.93
llama-3.1-8b -> qwen3-32b	+1.586 [+0.363,+3.060]	0.00700	+0.475 [+0.002,+0.959]	NO	+9.79
llama-3.1-8b -> qwen3.5-27b	+1.265 [+0.594,+2.034]	0.00087	-0.037 [-0.111,+0.049]	yes	+6.60
llama-3.1-8b -> qwen3.5-397b	+1.434 [+0.461,+2.496]	0.00337	+0.014 [-0.110,+0.150]	yes	+8.94
qwen3-8b-base -> qwen3.5-27b	+0.399 [+0.194,+0.595]	0.00087	+0.041 [-0.065,+0.157]	yes	+6.60

Table 33: ADDRPERM0908: the random- dictionary control at matched coefficient mass, run in the confirmatory binding cells.

\Delta

closes in 7 of 7 cells under Holm with the family fixed at freeze; the random arm covers zero in 5 of 7, which is exactly the pre-registered line and not more. The two cells where it does not are printed here rather than summarised away: a row-normalized random dictionary produces a small but significant slot effect on those receivers. Realized write norms are not matched here; the sitewise-matched control is Table 50 . Replay mismatch was 0.0000 in every cell.

HC-HS

is sender-independent by construction, so its 5 distinct values are per receiver.

cell	$\Delta_{\mathrm{sem}}$	95% CI	Holm $p$	permuted-dict slot effect [CI]
llama-3.1-8b -> qwen3-14b-base	+0.9443	[+0.4637,+1.4804]	$<$ 0.00088	-0.1985 [-0.4217,+0.0199]
llama-3.1-8b -> qwen3-30b-a3b	+2.5362	[+0.6961,+4.6977]	0.00550	-0.1666 [-0.6669,+0.3041]
llama-3.1-8b -> qwen3-32b	+2.7622	[+0.8328,+4.9024]	0.00550	-0.7013 [-1.7029,+0.3107]
llama-3.1-8b -> qwen3.5-27b	+1.3538	[+0.5862,+2.2652]	$<$ 0.00088	-0.1258 [-0.3825,+0.1058]
llama-3.1-8b -> qwen3.5-397b	+2.0513	[+0.7171,+3.4395]	0.00313	-0.6030 [-1.4241,+0.1234]
qwen3-8b-base -> qwen3.5-27b	+0.1432	[-0.0862,+0.3924]	0.12812	+0.2963 [+0.1000,+0.5015]

Table 34: ADDRSEM0909: the address-identity control, run IN the confirmatory binding cells. The packet is untouched; only the address

\to

row assignment inside the receiver’s own dictionary is permuted, so the packet, the coefficient multiset and the wire are matched BY CONSTRUCTION. Amplitude is not: the permuted assignment writes a different subset of the same unit-norm rows, and unit rows fix only the diagonal of the Gram matrix, so the realised write is close but not exact ( \NORMPERMMIN – \NORMPERMMAX of MC’s plan-side norm, Table 52 ).

\Delta_{\mathrm{sem}}

closes in 6 of 7 cells under Holm with the family fixed at freeze; the permuted dictionary carries a slot effect of its own in 2 of 7, printed here rather than summarised away.

cell	$\gamma$	free $I_{\mathrm{sender}}$ [95% CI]	Holm $p$	closes	forced $I_{\mathrm{sender}}$ [95% CI]	cov. none/inj.	oracle (margin)	MC $-$ none, free (pp)
llama-3.1-8b -> qwen3-14b-base	8	+0.2146 [+0.1439,+0.2919]	0.00050	yes	+0.8001 [+0.5736,+1.0617]	0.997/0.895	0.990 (+0.492)	+9.90 [+6.04,+13.80]
llama-3.1-8b -> qwen3.5-27b	4	+0.0000 [-0.0345,+0.0337]	1.00000	no	+0.9534 [+0.6157,+1.3258]	0.983/0.732	1.000 (+0.508)	-11.74 [-15.50,-8.22]
qwen3-8b-base -> qwen3.5-27b	4	-0.0164 [-0.0391,+0.0038]	1.00000	no	+0.2392 [+0.1204,+0.3731]	0.983/0.903	1.000 (+0.508)	-4.19 [-6.29,-2.27]
qwen3-8b-base -> glm-4-9b	2	+0.0493 [+0.0148,+0.0862]	0.00825	yes	+0.1649 [+0.1036,+0.2368]	0.678/0.609	0.990 (+0.651)	-1.01 [-3.50,+1.51]

Table 35: FREEGENCONF0910: the free-generation endpoint on confirmatory material, at a development-selected reduced dose. Free

I_{\mathrm{sender}}

is scored on free greedy generation (unanswered counts as incorrect); the forced-choice contrast beside it is computed on the SAME rows. The last column is the positive control that makes the two non-closing cells readable as nulls rather than as unmeasurable cells: the text-oracle arm answers correctly in free generation on those very episodes. The dose was selected on development material by a frozen coverage-only rule over the candidate grid

\{2,4,8\}

(so a

\gamma^{*}=8

cell may select 2); the bank is the entity-disjoint ENTDISJ0908 bank (test-split material with a stated exposure).

group	$T$ (linear)	$T$ (flex)	pairs $>$ null	Holm $p$	$T_{\mathrm{ctrl}}$ (fixed support)	$\Delta_{\mathrm{spec}}$	positive control
G1 (17-sender bank)	+0.723 [+0.656,+0.773]	+0.724	28/28	0.0010	+0.661 [+0.578,+0.718]	+0.063 [+0.034,+0.095]	0/8
G2 (lineage bank)	+0.740 [+0.694,+0.782]	+0.745	15/15	0.0010	–	not measured	0/6

Table 36: XSFUNC0910: do byte-disjoint codes from separately trained senders move the same receiver (qwen3.5-27b,

\gamma

8) the same way on the same episode? The pooled agreement is large and replicates on a second bank — and the two columns on the right are why it licenses less than it looks. A fixed-support sender, whose concept support cannot encode the episode, already reaches

T_{\mathrm{ctrl}}=0.66

, so only 9% of the agreement is support-specific; and the positive control — does the write move the receiver at all? — fails in 14 of 14 admitted cells, both banks being relation material on which this paper reports transport as null. Reported as adjudicated, with both limits, and the pre-declared abstract sentence for this branch deliberately not taken.

bank	cell	$I_{\mathrm{sender}}$	$G$	sign-flip $p$ (Holm ok)	joint gold+donor 95% CI	acc. DiD pp [95% CI]	MC $-$ none / MS $-$ none (pp)	within-tol. donors [CI] ( $n$ )
CC1	llama-3.1-8b $\rightarrow$ qwen3-14b-base	+0.566	14	$<$ 0.0001 (yes)	[+0.305,+0.854]	+15.8 [+8.5,+22.5]	+10.3 / -11.8	+0.516 [+0.215,+0.849] (167)
CC1	llama-3.1-8b $\rightarrow$ qwen3-30b-a3b	+1.443	14	0.0009 (yes)	[+0.526,+2.578]	+13.2 [+3.7,+22.3]	+9.3 / -7.6	+1.019 [-0.045,+2.234] (168)
CC1	llama-3.1-8b $\rightarrow$ qwen3-32b	+1.433	14	0.0054 (yes)	[+0.388,+2.642]	+15.4 [+4.1,+28.4]	+11.1 / -12.2	+0.906 [+0.120,+1.829] (167)
CC1	llama-3.1-8b $\rightarrow$ qwen3.5-27b	+0.827	14	0.0004 (yes)	[+0.356,+1.423]	+21.1 [+13.3,+28.9]	+14.0 / -14.0	+0.602 [+0.186,+1.125] (168)
CC1	llama-3.1-8b $\rightarrow$ qwen3.5-397b	+1.034	14	0.0035 (yes)	[+0.126,+1.983]	+22.6 [+10.4,+34.8]	+18.4 / -6.1	+0.404 [-0.018,+0.923] (64)
CC1	qwen3-8b-base $\rightarrow$ qwen3.5-27b	+0.541	14	0.0005 (yes)	[+0.332,+0.735]	+21.5 [+10.5,+31.7]	+8.6 / -9.8	+0.877 [+0.573,+1.214] (38)

Table 37: Robust inference beside the primary table (descriptive; nothing here re-adjudicates a gate). Per confirmatory cell on the CC1 bank and the entity-disjoint bank (ENTDISJ0908, roster and GLM cells; exposure statement in § ): the one-sided gold-entity sign-flip randomization

p

I_{\mathrm{sender}}

(exhaustive over

2^{G}

flips for

G\leq 16

, 20000 Monte-Carlo flips otherwise) with whether it clears the family’s Holm floor

0.05/m

; the 95% interval when gold entities AND donor entities are resampled jointly; the same difference-in-differences on forced-choice accuracy in percentage points with its gold-entity interval; the MC and MS accuracy contrasts against no message; and

I_{\mathrm{sender}}

restricted to episodes whose realized donor met both soft clauses (code gap

\leq 26/255

, write norm

\leq 5\%

). Bootstraps

n_{b}=8000

, seed 20260912.

bank	cell	attr. in gold slot	attr. anywhere	gold entity in own slot	any entity anywhere	$I_{\mathrm{sender}}$ , attr. hit [CI] ( $n$ )	attr. no-hit [CI] ( $n$ )	gold-slot hit [CI] ( $n$ )	gold-slot no-hit [CI] ( $n$ )
CC1	llama-3.1-8b $\rightarrow$ qwen3-14b-base	0.0%	0.0%	0.5%	14.9%	n too small (0)	+0.566 [+0.347,+0.796] (551)	n too small (3)	+0.566 [+0.346,+0.797] (548)
CC1	llama-3.1-8b $\rightarrow$ qwen3-30b-a3b	0.0%	0.0%	0.5%	14.9%	n too small (0)	+1.443 [+0.659,+2.348] (551)	n too small (3)	+1.432 [+0.641,+2.344] (548)
CC1	llama-3.1-8b $\rightarrow$ qwen3-32b	0.0%	0.0%	0.5%	14.9%	n too small (0)	+1.433 [+0.467,+2.519] (551)	n too small (3)	+1.423 [+0.450,+2.516] (548)
CC1	llama-3.1-8b $\rightarrow$ qwen3.5-27b	0.0%	0.0%	0.5%	14.9%	n too small (0)	+0.827 [+0.407,+1.300] (551)	n too small (3)	+0.824 [+0.401,+1.300] (548)
CC1	llama-3.1-8b $\rightarrow$ qwen3.5-397b	0.0%	0.0%	0.4%	13.8%	n too small (0)	+1.034 [+0.404,+1.737] (261)	n too small (1)	+1.040 [+0.409,+1.744] (260)
CC1	qwen3-8b-base $\rightarrow$ qwen3.5-27b	0.0%	0.0%	4.3%	6.4%	n too small (0)	+0.541 [+0.345,+0.712] (488)	+0.423 [-0.250,+0.759] (21)	+0.546 [+0.350,+0.720] (467)

Table 38: Attribute-level packet attribution (descriptive; both splits pre-declared before they were computed). Per confirmatory cell, MC arm, four-arm episodes: the percentage of packets whose lane at the gold entity’s slot contains the episode’s attribute word, whose packet contains it in any lane, whose gold-slot lane contains the gold entity’s own word (Table 20 ’s statistic), and whose packet names any entity; then

I_{\mathrm{sender}}

on the episodes where the attribute word is present versus absent, and on those where the gold entity is named in its own slot versus not, plain episode means with gold-entity bootstrap intervals (

n_{b}=8000

, seed 20260912); splits with fewer than 20 hit episodes are not printed. The 11 test attribute words are members of the 221-word inventory; the development-bank attributes of CC1 are not, so the attribute split is undefined there. ENTDISJ cells carry the exposure statement of § .

cell	$\Delta_{\mathrm{abl}}$	95% CI	Holm $p$	ablated-dict slot effect [CI]
llama-3.1-8b -> qwen3-14b-base	+0.1426	[-0.0676,+0.3858]	0.70613	+1.4174 [+1.1224,+1.7693]
llama-3.1-8b -> qwen3-30b-a3b	+0.2417	[-0.2672,+0.8116]	1.00000	+4.7418 [+3.6972,+5.8301]
llama-3.1-8b -> qwen3-32b	+0.1618	[-0.3405,+0.6858]	1.00000	+4.3060 [+3.2075,+5.4336]
llama-3.1-8b -> qwen3.5-27b	-0.0076	[-0.2469,+0.2637]	1.00000	+2.6722 [+2.0740,+3.3250]
llama-3.1-8b -> qwen3.5-397b	-0.5159	[-1.0941,+0.0749]	1.00000	+4.5019 [+3.5274,+5.5002]
qwen3-8b-base -> qwen3.5-27b	-0.1224	[-0.2496,+0.0002]	1.00000	+0.8923 [+0.5532,+1.2751]

Table 39: ADDRABL0912: the single-address ABLATION control, run in the entity-disjoint roster cells (the bank of ENTDISJ0908, exposure statement as there) at their frozen doses. The packet is untouched; at the receiver the row of the sender’s single most frequent address is zeroed (ZC/ZS arms), so the arm asks whether that dominant address is load-bearing.

\Delta_{\mathrm{abl}}=(\mathrm{MC}-\mathrm{MS})-(\mathrm{ZC}-\mathrm{ZS})

closes in 0 of 7 cells under Holm with the family fixed at freeze (a non-rejection, not an equivalence bound; the raw interval excludes zero in 1 of 7, printed below); the ablated dictionary still carries a slot effect of its own in 7 of 7.

cell	episodes	four-arm $n$	$G$	$I_{sender}$	gold-entity 95% CI	$p$ (Holm)	verdict
llama-3.1-8b $\rightarrow$ qwen3-14b-base	596	536	42	+0.800	[+0.572,+1.059]	0.000875	closes
llama-3.1-8b $\rightarrow$ qwen3-30b-a3b	596	536	42	+4.477	[+3.240,+5.805]	0.000875	closes
llama-3.1-8b $\rightarrow$ qwen3-32b	596	536	42	+2.231	[+1.375,+3.178]	0.000875	closes
llama-3.1-8b $\rightarrow$ qwen3.5-27b	596	536	42	+2.504	[+1.648,+3.398]	0.000875	closes
llama-3.1-8b $\rightarrow$ qwen3.5-397b	596	536	42	+1.676	[+1.266,+2.087]	0.000875	closes
qwen3-8b-base $\rightarrow$ qwen3.5-27b	596	548	42	+0.777	[+0.462,+1.141]	0.000875	closes

Table 40: UNIDOSE0912: the entity-disjoint confirmatory roster (the bank of ENTDISJ0908, exposure statement as there) re-run at ONE roster-wide dose,

\gamma=8

for every cell, fixed before the run in place of the per-pair adjudicated dose; cells whose adjudicated dose was already 8 are same-dose replays and reproduce ENTDISJ0908 exactly. Upper block: the 7 CC1-roster cells under Holm at

\alpha=0.05

with

m=7

fixed at freeze. Lower block: the two primary-dose cross-family cells, adjudicated as their own family (each alone with no correction; Holm

m=2

G

bank / receiver	episodes/entities	$T$ linear [95% CI] (flex)	raw perm. $p$	$T_{\mathrm{ctrl}}$ [95% CI]	$\Delta_{\mathrm{spec}}$ [95% CI]	pos. ctrl	frozen
CC1 / qwen3.5-27b ( $\gamma$ 8)	592/14	+0.344 [+0.233,+0.520] (+0.335)	0.0005	+0.372 [+0.244,+0.509]	-0.028 [-0.124,+0.109]	2/2	VOID
CC1 / qwen3.5-397b ( $\gamma$ 4)	296/14	+0.214 [+0.134,+0.304] (+0.218)	0.0005	+0.281 [+0.146,+0.420]	-0.067 [-0.197,+0.067]	2/2	VOID
ENTDISJ / qwen3.5-27b ( $\gamma$ 8)	596/42	+0.361 [+0.253,+0.479] (+0.348)	0.0005	+0.266 [+0.214,+0.327]	+0.094 [-0.029,+0.221]	2/2	VOID
ENTDISJ / qwen3.5-397b ( $\gamma$ 4)	596/42	+0.155 [+0.041,+0.273] (+0.172)	0.0005	+0.185 [+0.141,+0.232]	-0.030 [-0.159,+0.095]	2/2	VOID

Table 41: XSFUNCB0912 (blind-panel item 9): per-episode agreement of

\mathrm{MC}-\mathrm{none}

between llama-3.1-8b and qwen3-8b-base on the binding material, none partialled (linear; flexible in parentheses), cluster-permuted null, gold-entity bootstrap;

T_{\mathrm{ctrl}}

is the same statistic between each real sender and a fixed-support sender (gemma-3-27b, packet diversity 1) and

\Delta_{\mathrm{spec}}=T-T_{\mathrm{ctrl}}

. Frozen verdict PARTIAL: every group is VOID under the lock’s alignment guard (

\geq 4

admitted senders,

\geq 6

admitted pairs, copied from the relation-material lock; the design has 2 senders and 1 pair per group), so no string beyond PARTIAL is adjudicated and every number here is descriptive. On those numbers the positive control passes in 8 of 8 admitted cells and the real pair’s agreement is matched by the fixed-support write in every group (

\Delta_{\mathrm{spec}}

covers zero in 4 of 4;

T_{\mathrm{ctrl}}>T

in 3).

cell	stage	$n$	tokens	ms median [p5, p95]	of which in-hook solve (ms)
binding: llama-3.1-8b -> qwen3-14b-base ( $\{}gamma$ 16, probe221)	sender: packet (forward + batch-1 solve, per lane $\times$ lanes)	253	–	71.9 (2 $\times$ [10.1+25.8])	–
	sender: text_gen@B greedy decode (B median 22 B)	125	11 new	106.4 [85.7,133.1]	–
	receiver: none prefill	128	100	18.0 [17.8,18.2]	0
	receiver: text_oracle prefill	128	98	18.0 [17.9,18.2]	0
	receiver: text_packet prefill	128	148	19.6 [19.1,23.2]	0
	receiver: MC prefill	128	100	430.6 [387.7,473.6]	409.3

Table 42: COST0912: the compute axis, measured as implemented (batch 1, one B200, synchronised wall-clock, median [p5, p95] after three warm-up rows). Sender: one forward at the tap plus the batch-1 FISTA solve per lane makes the packet; a greedy text generation capped at the packet’s own byte budget (text_gen@B, max(8, B/2) new tokens, run to the cap as in the rate–distortion study) makes the text. Receiver: prefill wall-clock and prompt tokens per arm; the packet arm’s hooks solve

a^{0}

at every written position of every band layer, and that solve is the whole of its receiver-side cost. Descriptive; nothing here is gated.

dose	cov. no message	cov. packet write	cov. text oracle	admitted?
$\gamma=4$	0.4865 (72/148)	0.2095 (31/148)	1.0000 (148/148)	no
$\gamma=8$	0.4865 (72/148)	0.0000 (0/148)	1.0000 (148/148)	no

Table 43: FREEGEN397B0912: why this paper reports no behavioural measurement on qwen3.5-397b. Format compliance (the fraction of episodes answered in the frozen free-generation format, 16 tokens) for the unwritten arm, the packet arm and the text oracle, on the same episodes. The gate’s frozen rule admits a dose only if the unwritten AND the injected arm each reach 0.50; neither candidate does, so the gate stopped at stage 1, no stage-2 cell was launched and no behavioural field was read. Two things are visible: this receiver is already below the floor with no message at all, and the write — not the prompt — removes the rest, while the oracle on the same rows answers every time. Coverage only; no contrast is reported or implied here.

point	bytes pkt/txt	no comm.	packet	text	packet $-$ text [95% CI]	Holm $p$	recv. prefill pkt/txt
$k=1$	13/12	0.240	0.213	0.507	-0.2933 [-0.4220,-0.1707]	0.0006	12.1 s / 31 ms
$k=2$	16/14	0.240	0.220	0.527	-0.3067 [-0.4274,-0.1989]	0.0006	12.2 s / 31 ms
$k=4$	16/14	0.240	0.227	0.527	-0.3000 [-0.4151,-0.1917]	0.0006	12.2 s / 31 ms
$k=6$	16/14	0.240	0.253	0.527	-0.2733 [-0.3882,-0.1636]	0.0006	12.2 s / 31 ms
$k=8$	16/14	0.240	0.273	0.527	-0.2533 [-0.3636,-0.1527]	0.0006	12.3 s / 31 ms

Table 44: RATECOMP0912: the packet’s rate family against a budget-aware text sender on a confirmatory-strength pair (llama-3.1-8b

\to

qwen3-32b) and the confirmatory channel, 150 items, 60 gold entities, paired per item and bootstrapped over entities. The text arm composes under each point’s own realized wire bytes, so the comparison is byte-matched per item. Two limits: the precision half of the frozen design (5 of 10 points) is not run — the zero-write integrity control cannot exist when no coefficients are transmitted — and the

k

sweep realizes only 2 distinct wire sizes, because the grounded candidate set caps the support at a median of 2 addresses. Read this as five runs over 2 rates on one pair. At every point text wins (all intervals strictly below zero, worst Holm

p

0.0006) and the packet’s own delta over no communication covers zero, with the oracle at 1.000 against a 0.240 baseline.

sender family	episodes	candidates	both entity words	exactly one	words used	top-1	null mean / p99
llama-3.1-8b	592	266	0 (0.000)	88 (0.149)	41	0.0000	0.00005 / 0.00169
qwen3-8b-base	592	266	0 (0.000)	57 (0.096)	22	0.0000	0.00000 / 0.00000

Table 45: DECODE0912: can a reader recover the episode from the packet? The frozen test ranks every candidate entity pair in the bank (266 of them) by word overlap with the packet and counts an episode as retrieved only if its true pair is strictly first. Top-1 is zero in both sender families, below permutation nulls that are themselves near zero. The middle columns are a post-hoc decomposition, computed after the frozen statistic to make it interpretable: the packet never carries both of the episode’s entity words, carries exactly one in a tenth to a seventh of episodes, and each entity word recurs across a median of six candidate pairs, so a one-word match always ties. The frozen string is PARTIAL rather than NOT-DECODABLE only because that branch required the observed rate to fall below twice a null mean that is exactly zero in one family — a defect in the threshold, reported rather than repaired. Nothing here speaks to whether the receiver uses the packet. Nor does it speak to what a different reader could recover: the frozen test is a LITERAL one, scoring word overlap alone, and it ignores slot, coefficient, attribute word and any association between words. What it licenses is that this literal reader cannot uniquely name the entity pair, not that the packet carries no episode-specific content — the concentration split in the main text shows it does. (External review, 2026-09-13, which named exactly this over-reading.)

arm	gap	what it is
none	+2.1250	no message
MC	+2.6875	the message assigned to the correct objects
MS	-4.2500	the same message with object assignments exchanged
PC	+0.4375	another example’s message
PS	-0.2500	the replacement message with assignments exchanged
zero_packet	+2.1250	zero vectors added through the same write operation

Table 46: One episode, end to end ( t2_dev_003d4e77b8_t0p0o0 , the primary confirmatory cell at

\gamma=16

; picked as the lexicographically first episode carrying every arm, not for its size). The sender reads two descriptions and never sees the question. Its packet is 24 bytes — for NEXO: drawer , scarf , stove , wooden ; for BLIX: boat , ship — and the gold entity’s own word ( pear ) is not among the words selected for that object. The receiver resolves those addresses through its own dictionary into 45 writes at layers 26, 32, 38. Exchanging the object assignments turns the contrast around; another example’s message changes it much less; a packet that decodes to zero reproduces the no-message gap exactly.

cell	pair units	$I_{\mathrm{sender}}\mid q_{0}$	$I_{\mathrm{sender}}\mid q_{1}$	FOLLOW	SLOT	verdict
llama-3.1-8b -> glm-4-9b $\gamma$ 16	26	+0.083 [-0.416,+0.582]	+0.414 [+0.004,+0.846]	+0.249 [-0.024,+0.504]	-0.166 [-0.555,+0.209]	INCONCLUSIVE
llama-3.1-8b -> qwen3-30b-a3b $\gamma$ 8	26	+1.329 [-0.800,+3.411]	+1.857 [+0.056,+3.852]	+1.593 [+0.784,+2.429]	-0.264 [-2.177,+1.516]	INCONCLUSIVE
llama-3.1-8b -> qwen3-32b $\gamma$ 16	26	+1.546 [-0.123,+3.218]	+1.818 [+0.145,+3.535]	+1.682 [+0.573,+2.761]	-0.136 [-1.430,+1.082]	INCONCLUSIVE
llama-3.1-8b -> qwen3.5-27b $\gamma$ 8	26	+0.788 [-0.055,+1.630]	+1.361 [+0.489,+2.326]	+1.075 [+0.550,+1.600]	-0.287 [-1.037,+0.379]	INCONCLUSIVE
llama-3.1-8b -> qwen3.5-397b $\gamma$ 4	26	+1.901 [+0.499,+3.223]	+1.668 [+0.189,+3.220]	+1.784 [+0.821,+2.697]	+0.117 [-1.013,+1.161]	FOLLOWS-QUERY
qwen3-8b-base -> qwen3.5-27b $\gamma$ 8	27	+0.470 [+0.006,+0.986]	+0.608 [+0.124,+1.127]	+0.539 [+0.140,+0.985]	-0.069 [-0.326,+0.194]	FOLLOWS-QUERY

Table 47: QFOLLOW0913: does the effect follow the question? The sender extracts its packet ONCE and never sees a question; the receiver is then scored under two pre-fixed attribute questions whose correct aliases differ, with the same write. Per-cell verdicts are the pinned analyzer’s under the rule frozen on 2026-09-05 for ENTDISJ0908’s B2 arm — FOLLOWS-QUERY iff

I_{\mathrm{sender}}

’s CI-lower is above zero under BOTH questions, SLOT-BIAS iff positive under one and negative under the other — and this table only counts them. FOLLOW is the mean of the two per-query contrasts and SLOT their half-difference (the convention of the earlier verdict table; the analyzer stores the sum). The unit is the unordered entity pair. The two cells B2 measured are quoted in the caption and were neither re-run nor re-adjudicated: llama-3.1-8b -> qwen3-14b-base_g16 FOLLOWS-QUERY (ENTDISJ0908 B2); qwen3-8b-base -> glm-4-9b_g8 INCONCLUSIVE (ENTDISJ0908 B2).

cell	$n$	$I_{sender}$ , one layer [95% CI]	$p$ (Holm)	$I_{sender}$ , band (CC1)	retention	norm ratio	verdict
llama-3.1-8b -> qwen3-14b-base	551	+0.439 [+0.27,+0.62]	0.0009	+0.566 [+0.35,+0.79]	0.78 [0.67,0.90]	0.73	closes
llama-3.1-8b -> qwen3-30b-a3b	551	+1.247 [+0.62,+1.95]	0.0009	+1.443 [+0.67,+2.34]	0.86 [0.77,1.04]	0.79	closes
llama-3.1-8b -> qwen3-32b	551	+1.344 [+0.41,+2.40]	0.0019	+1.433 [+0.48,+2.53]	0.94 [0.76,1.13]	0.63	closes
llama-3.1-8b -> qwen3.5-27b	551	+0.811 [+0.39,+1.28]	0.0009	+0.827 [+0.40,+1.30]	0.98 [0.80,1.20]	0.65	closes
llama-3.1-8b -> qwen3.5-397b	261	+1.064 [+0.45,+1.71]	0.0009	+1.034 [+0.40,+1.74]	1.03 [0.68,1.79]	0.35	closes
qwen3-8b-base -> qwen3.5-27b	488	+0.662 [+0.43,+0.88]	0.0009	+0.541 [+0.35,+0.71]	1.22 [1.03,1.44]	0.56	closes

Table 48: BANDWIDTH0914: the seven confirmatory cells re-run with the receiver written at the band’s lowest layer only (depth 0.4) instead of three layers (0.4/0.5/0.6); packets, dose

\gamma^{*}

, donor pairs and bank are the confirmatory run’s, and the no-message arm replays it exactly. Gate string SINGLE_LAYER_SUFFICES: the contrast closes under Holm (

m=7

) in 7 of 7 cells. The single layer was chosen on one development cell after four configurations were compared. Retention (single-layer over band estimate on the same episodes, CI from the same entity resamples) and the norm ratio (the single write’s realised total norm over the band’s, recorded inside the forward) are descriptive. Because the single write carries less norm, a cell that does not close does not separate needing several layers from needing more written norm; no dose was re-fitted.

cell	draws	worst-draw CI lower	$p$ (worst draw)	frozen draw’s CI lower	draw SD / SE	verdict
llama-3.1-8b -> qwen3-14b-base	6	+0.090	0.0075	+0.465	0.76	misses
llama-3.1-8b -> qwen3-30b-a3b	6	-0.479	0.0759	+0.725	0.66	misses
llama-3.1-8b -> qwen3-32b	6	-0.651	0.1065	+0.862	0.76	misses
llama-3.1-8b -> qwen3.5-27b	6	+0.133	0.0090	+0.578	0.54	misses
llama-3.1-8b -> qwen3.5-397b	6	-1.653	0.4913	+0.750	1.29	misses
qwen3-8b-base -> qwen3.5-27b	6	-0.135	0.3680	-0.086	1.99	misses

Table 49: ADDRDRAW0913: the address-identity control repeated over 6 independent fixed-point-free permutations, the one ADDRSEM0909 drew included. Each cell is adjudicated on its worst draw — the minimum gold-entity CI lower bound and the maximum one-sided

p

over draws — under Holm with

m=7

fixed at freeze, which is strictly harder than the single-draw bar ADDRSEM0909 met. Gate string DRAW-DEPENDENT: 0 of 7 cells survive their worst draw, against 6 that the frozen draw alone clears. The frozen draw enters by re-reading ADDRSEM0909’s existing rows, whose no-message, MC and MS gaps are bit-identical to the new runs’ on every shared episode; that is checked, not assumed. Descriptive column: the standard deviation of the draws’ point estimates over the median within-draw standard error (95% interval width / 3.92). The draw moves the estimate by a median 0.76 of a standard error (0.54–1.99), by more than one in 3 of 7 cells.

cell	$n$	$\Delta^{*}$ [95% CI]	$p$	matched random slot effect	verdict
llama-3.1-8b -> qwen3-14b-base	592	+0.689 [+0.28,+1.15]	0.0002	+0.056 [-0.15,+0.29]	SURVIVES-CELL
llama-3.1-8b -> qwen3-30b-a3b	592	+2.238 [+0.51,+4.25]	0.0032	+0.132 [-0.30,+0.56]	SURVIVES-CELL
llama-3.1-8b -> qwen3-32b	592	+1.740 [+0.58,+3.12]	0.0006	+0.321 [+0.02,+0.64]	SURVIVES-CELL
llama-3.1-8b -> qwen3.5-27b	592	+1.263 [+0.59,+2.04]	0.0001	-0.035 [-0.08,+0.02]	SURVIVES-CELL
llama-3.1-8b -> qwen3.5-397b	296	+1.449 [+0.41,+2.54]	0.0026	-0.001 [-0.22,+0.23]	SURVIVES-CELL
qwen3-8b-base -> qwen3.5-27b	592	+0.352 [+0.11,+0.57]	0.0025	+0.088 [-0.06,+0.25]	SURVIVES-CELL

Table 50: NORMFIX0913: the same contrast through a random dictionary matched in runtime edit-vector norm, site by site, to the receiver’s own dictionary. Norms are recorded in float32 before bf16 conversion and addition; final rounded state differences are not matched.

\Delta^{*}=(\mathrm{MC}-\mathrm{MS})-(\mathrm{RC}^{*}-\mathrm{RS}^{*})

; the correct-slot control matches MC and the exchanged-slot control matches MS, so both legs of the control interaction stand against the leg they replace. Holm at

\alpha=0.05

with

m=7

fixed at freeze; gate string SURVIVES (

\Delta^{*}>0

in 7 of 7, matched random covering zero in 6). The amplitude guard is measured, not asserted: a cell is void unless the matched arms reproduce their reference norms to

10^{-6}

on at least 99% of episodes.

cell	unmatched realised	unmatched plan-side	rescale factor	matched fraction
llama-3.1-8b -> qwen3-14b-base	1.161	0.844	0.939	1.0000
llama-3.1-8b -> qwen3-30b-a3b	1.071	0.846	1.049	1.0000
llama-3.1-8b -> qwen3-32b	1.058	0.851	1.081	1.0000
llama-3.1-8b -> qwen3.5-27b	1.139	0.848	0.937	1.0000
llama-3.1-8b -> qwen3.5-397b	0.551	0.837	3.131	1.0000
qwen3-8b-base -> qwen3.5-27b	0.825	0.680	1.516	1.0000

Table 51: DESCRIPTIVE amplitude audit (it may correct a factual statement and may promote no claim). ‘realised’ is the runtime edit-vector norm before bf16 addition, summed over the written band; ‘plan-side’ is computed against the un-intervened forward, which is exact only at the lowest written layer. They differ because the real dictionary’s first write moves the residual toward its own target, leaving the upper layers little to do, while a random target does not self-cancel that way. ‘matched fraction’ is the guard: the proportion of episodes on which the matched control reproduced its reference’s realised norm to

10^{-6}

control	cell	episodes	median $\lVert\delta\rVert$ ratio to MC	[min, max]
Gaussian rows	llama-3.1-8b -> qwen3-14b-base	592	0.843	[0.637, 1.008]
Gaussian rows	llama-3.1-8b -> qwen3-30b-a3b	592	0.848	[0.642, 1.009]
Gaussian rows	llama-3.1-8b -> qwen3-32b	592	0.851	[0.652, 0.998]
Gaussian rows	llama-3.1-8b -> qwen3.5-27b	592	0.846	[0.654, 1.004]
Gaussian rows	llama-3.1-8b -> qwen3.5-397b	296	0.839	[0.631, 1.000]
Gaussian rows	qwen3-8b-base -> qwen3.5-27b	592	0.681	[0.623, 0.758]

Table 52: What the dictionary-substitution controls actually write. The write is

\delta=(\hat{a}_{S}-a^{0}_{S})^{\top}D_{S}

, so unit-norm rows fix only the diagonal of

D_{S}D_{S}^{\top}

: the angles between the written rows, and the receiver’s own

a^{0}

(which changes with the dictionary), set the realized magnitude. Per episode at the lowest written layer, the Gaussian arm writes 0.64–0.85 of MC’s norm and the permuted-real arm 0.99–1.09. An earlier version of this appendix called the Gaussian amplitude matched exactly ; it is not, and the correction is this table. The permuted-real arm — the stronger of the two controls, since it moves only address identity — is matched to about a percent, so the address-identity result does not rest on the mismatched arm. Higher layers are logged on the un-intervened forward and are proxies, so only the lowest written layer is quoted.

arm	$L_{26}$	$L_{32}$	$L_{38}$	share of total	band mean
the receiver’s own rows (MC)	240.7	55.3	78.3	64% / 15% / 21%	125.0
Gaussian rows	206.9	83.1	109.1	52% / 21% / 27%	134.2
ratio, Gaussian / own rows	0.859	1.502	1.394

Table 53: What the three-layer band actually writes, measured inside the forward on a development cell (llama-3.1-8b

\rightarrow

qwen3-32b,

\gamma^{*}{=}16

; descriptive, no gate). A three-layer write is not three equal writes: the receiver’s own rows put most of the norm at the first written layer, because that write already moves the residual toward the target and leaves the layers above it little to do. A random dictionary does not self-cancel that way, so it writes less than the real dictionary at the first layer and more above it. The plan-side norm of Table 52 equals the first-layer figure here exactly, which is what makes it exact at that layer and a proxy above it.

cell	episodes	layers	median $\lVert\delta\rVert$ per layer	share of total
llama-3.1-8b -> qwen3-14b-base	592	16/20/24	108.45 / 27.14 / 13.46	73% / 18% / 9%
llama-3.1-8b -> qwen3-30b-a3b	592	19/24/29	10.72 / 1.97 / 0.91	79% / 15% / 7%
llama-3.1-8b -> qwen3-32b	592	26/32/38	210.09 / 52.39 / 71.23	63% / 16% / 21%
llama-3.1-8b -> qwen3.5-27b	592	26/32/38	52.96 / 16.15 / 11.76	65% / 20% / 15%
llama-3.1-8b -> qwen3.5-397b	296	24/30/36	2.21 / 2.16 / 2.02	35% / 34% / 32%
qwen3-8b-base -> qwen3.5-27b	592	26/32/38	44.40 / 19.36 / 15.94	56% / 24% / 20%

Table 54: The same measurement on every confirmatory cell, for the receiver’s own dictionary. A multi-layer write is front-loaded: the first written layer carries 35–79% of the band’s total written norm. The flattest cell is llama-3.1-8b -> qwen3.5-397b, which is also the smallest write and the lowest dose in the roster; that is consistent with the mechanism, since front-loading is a consequence of the first write actually moving the residual toward the target, and a write too small to move it leaves the layers above with as much to do as the first. The ordering was read off the same table, not predicted in advance. Norms are per-episode medians of the mean over that layer’s written sites. This is write geometry, not a contrast, and it does not depend on any gate’s outcome.

Figure 1: Shared words, each receiver’s own directions. The sender reads private facts and sends the indices of selected words, grouped by object, without seeing the question. Each receiver writes these words into its hidden states with its own word directions, then answers. Colors mark the same word across models; models and vectors are schematic, and answers are task ground truth.

Figure 2: Construct directions once, then write and reuse facts. (a) Each symbol’s direction averages the receiver’s hidden-state difference from a neutral filler over twelve generic contexts. (b) Retrieved directions are added once to saved neutral states at relative depth 0.2 ( $\gamma=1$ ); the resulting prefix cache is copied for each question. Answers shown are task ground truth.

Figure 3: One changed attribute, four required answers. Only the sender sees the descriptions. Four-answer success requires both questions correct in both worlds. Shown answers are task ground truth, not model outputs.

		Gemma-4 sender		Qwen3.6 sender		gpt-oss sender		Middle
Receiver	Size	Internal	Text	Internal	Text	Internal	Text	(Gemma-4)
Qwen3.5-2B	2B	81.2	86.7	79.3	83.8	77.5	81.5	4.7
Qwen3.5-4B	4B	96.0	96.7	93.4	94.3	89.9	88.0	3.1
Qwen3.5-9B	9B	93.4	94.7	93.6	96.2	86.6	88.5	4.9
Qwen3.5-27B	27B	98.3	98.6	96.4	96.7	88.8	89.1	77.8
Qwen3.6-27B	27B	97.7	98.2	96.0	96.0	90.6	89.5	82.5

Table 55: A fixed interface carries to new senders and receivers.

Figure 4: Usable depth ranges differ across receivers. Complete-fact four-answer success on 92 development scenes, one write per depth; fitted directions use

\gamma=1

, native states come from the receiver reading the facts. Shading: 95% family-bootstrap intervals.

Figure 5: Depth and strength jointly determine usable writes. Complete-fact four-answer success (%, rounded) on 92 development scenes. Hatching: coverage below 90% (scenes stay in the denominator). Orange outlines: the configuration selected for confirmation.

Relative	Qwen3.5				Qwen3.6	Granite-4.1	OLMo-3	gpt-oss	Mistral-3.2
depth	2B	4B	9B	27B	27B	8B	7B	20B	24B
0.2	89.6	100.0	100.0	100.0	100.0	100.0	100.0	89.6	99.0
0.35	35.4	7.3	34.4	100.0	100.0	99.0	20.8	26.0	10.4
0.5	0.0	0.0	0.0	18.8	85.4	21.9	2.1	2.1	0.0
0.65	2.1	0.0	3.1	0.0	11.5	0.0	2.1	1.0	0.0
0.8	0.0	0.0	0.0	0.0	0.0	0.0	0.0	0.0	0.0

Table 56: Usable write depths of new receivers, predicted before measurement.

	New layouts (complete facts)				Countries (Gemma-4 sender)
Receiver	Standard	Slots	Template	Three obj.	Internal	Middle	Reass.	Text
Qwen3-32B	100.0	100.0	98.6	99.3	46.5	18.7	1.7	92.9
Qwen3.5-27B	100.0	100.0	100.0	100.0	64.0	5.8	5.0	95.7
Qwen3.6-27B	100.0	100.0	100.0	100.0	47.3	12.3	3.9	97.2
Mistral-3.2-24B	97.9	97.2	89.6	95.1	23.9	0.0	0.6	87.1
Qwen3.5-9B	93.8	99.3	93.8	95.1

Table 57: The fixed setting beyond the development task.

Figure 6: Graded weights: write strength versus printed weights. Four-answer success with complete facts as the competing value’s weight

L

rises. At

L=255

both worlds receive identical messages, so success is zero by construction. Shading: 95% family-bootstrap intervals.

Appendix figures & tables38 assets

Supplementary material from the paper’s appendix.

Appendix

	Development gate (96 scenes)			Confirmation messages (576 scenes)
Sender	Parsed	Recall	Complete	Parsed	Recall	Complete obj.	Bytes
Gemma-4	100.0	100.0	100.0	100.0	98.2	96.4	21.9
Qwen3.6	100.0	100.0	100.0	100.0	97.8	95.6	22.5
gpt-oss	99.5	95.4	90.1	98.9	94.1	89.2	20.3

Appendix

Table 58: Message quality of the three new senders.

Receiver	Internal [95% CI]	Text	Middle	Reass. J/valid	Valid	Int. $-$ text [95% CI]	Verdict
Gemma-4 sender
Qwen3.5-2B	81.2 [76.7, 85.4]	86.7	4.7	2.5 / 99.4	99.9	$-$ 5.6 [-9.1, -2.3]	text better
Qwen3.5-4B	96.0 [93.9, 97.9]	96.7	3.1	1.8 / 97.8	99.8	$-$ 0.7 [-3.7, 2.1]	unresolved
Qwen3.5-9B	93.4 [89.7, 96.5]	94.7	4.9	2.9 / 93.9	99.3	$-$ 1.4 [-3.1, 0.4]	unresolved
Qwen3.5-27B	98.3 [96.2, 99.8]	98.6	77.8	2.2 / 86.3	99.5	$-$ 0.3 [-1.6, 0.5]	similar
Qwen3.6-27B	97.7 [96.0, 99.1]	98.2	82.5	3.0 / 93.9	99.9	$-$ 0.5 [-1.2, 0.0]	similar

Appendix

Table 59: Newer senders and receivers: all cells.

Receiver	Internal	Native states	Text	Full evidence	Internal $-$ text [95% CI]
Qwen3.5-2B	86.3	85.9	89.6	97.2	$-$ 3.3 [-5.6, -1.0]
Qwen3.5-4B	99.0	99.7	98.4	100.0	+0.5 [-0.7, 1.7]
Qwen3.5-9B	96.2	99.8	100.0	100.0	$-$ 3.8 [-7.0, -1.2]
Qwen3.5-27B	100.0	100.0	100.0	100.0	+0.0 [0.0, 0.0]
Qwen3.6-27B	100.0	100.0	100.0	100.0	+0.0 [0.0, 0.0]
Granite-4.1-8B	98.1	97.6	100.0	99.8	$-$ 1.9 [-3.1, -0.9]

Appendix

Table 60: Complete facts on the newer receivers.

Development receivers (threshold fitted here)
Receiver	0.2	0.35	0.5	0.65	0.8
Qwen3-32B	1.00 / 100.0	1.00 / 100.0	1.00 / 96.9	0.40 / 1.0	0.08 / 0.0
Qwen3-14B-Base	1.00 / 100.0	0.98 / 87.5	0.04 / 6.2	0.01 / 1.0	0.03 / 0.0
GLM-4-9B	0.89 / 88.5	0.68 / 8.3	0.09 / 0.0	0.12 / 0.0	0.12 / 0.0
Gemma-3-27B-IT	1.00 / 100.0	1.00 / 85.4	0.09 / 0.0	0.04 / 0.0	0.05 / 0.0
Held-out receivers (predictions committed before these outcomes)

Appendix

Table 61: Binding probe and complete-fact success at every tested depth.

Receiver	Layout	Internal [95% CI]	Valid	$-$ Standard [95% CI]	Text	Reass.	Verdict
Qwen3-32B	Standard	100.0 [100.0, 100.0]	100.0		99.3	1.4	reference
	Slots	100.0 [100.0, 100.0]	100.0	+0.0 [0.0, 0.0]	100.0	1.4	transfers
	Template	98.6 [96.5, 100.0]	100.0	$-$ 1.4 [-3.5, 0.0]	99.3	0.0	transfers
	Three obj.	99.3 [97.9, 100.0]	100.0	$-$ 0.7 [-2.1, 0.0]	99.3	0.0	transfers
Qwen3.5-27B	Standard	100.0 [100.0, 100.0]	100.0		100.0	2.1	reference
	Slots	100.0 [100.0, 100.0]	100.0	+0.0 [0.0, 0.0]	100.0	3.5	transfers

Appendix

Table 62: New layouts: all cells.

Receiver	Internal [95% CI]	Valid	Middle	Reass.	Text	Evidence	Int. $-$ reass. [95% CI]	Int. $-$ text [95% CI]
Qwen3-32B	46.5 [38.9, 56.7]	97.9	18.7	1.7	92.9	97.5	+44.8 [37.1, 54.6]	$-$ 46.4 [-54.8, -35.8]
Qwen3.5-27B	64.0 [56.3, 76.0]	100.0	5.8	5.0	95.7	96.3	+59.0 [51.4, 69.3]	$-$ 31.7 [-39.6, -20.7]
Qwen3.6-27B	47.3 [39.7, 58.0]	100.0	12.3	3.9	97.2	99.4	+43.4 [35.5, 53.1]	$-$ 49.9 [-57.5, -40.7]
Mistral-3.2-24B	23.9 [15.1, 31.5]	97.5	0.0	0.6	87.1	100.0	+23.3 [14.7, 30.8]	$-$ 63.2 [-73.2, -52.4]

Appendix

Table 63: Countries: all receivers.

		Weight $L$ of the competing value
Receiver	Route	0	64	128	191	255
Qwen3-32B	Internal	100.0	99.3	72.9	4.9	0.0
	Text	99.3	31.2	19.4	18.1	0.0
Qwen3.5-27B	Internal	100.0	100.0	95.1	50.7	0.0
	Text	100.0	64.6	41.0	38.9	0.0
gpt-oss-20b	Internal	84.0	76.4	63.2	34.0	0.0

Appendix

Table 64: Graded weights: all evaluated receivers.

Receiver	Middle	Early [95% CI]	Text	Reassigned Success / Valid
Qwen3-32B	94.4	97.0 [93.2, 99.5]	97.0	1.2 / 62.2
Qwen3-14B-Base	31.0	97.7 [95.1, 99.8]	97.7	2.3 / 99.2
GLM-4-9B	0.5	74.3 [66.9, 81.3]	87.8	0.6 / 49.8
Gemma-3-27B-IT	30.5	97.7 [94.8, 100.0]	97.9	3.0 / 89.8

Appendix

Table 65: Automatic fact communication across receiver families.

Receiver	Directions	Native states	Difference [95% CI]	Text
Qwen3-32B	99.8	100.0	-0.2 [-0.5, +0.0]	99.8
Qwen3-14B-Base	99.5	99.8	-0.3 [-0.9, +0.0]	100.0
GLM-4-9B	76.5	75.6	+0.9 [-1.2, +3.2]	93.0
Gemma-3-27B-IT	100.0	100.0	+0.0 [+0.0, +0.0]	100.0

Appendix

Table 66: Reusable symbol directions and the receiver’s own states.

Receiver	Twelve contexts	One fitting context			Embedding (= text)
		First	Middle	Last
Qwen3-32B	98.8	94.6	98.8	98.6	99.3
Qwen3-14B-Base	99.1	97.2	99.8	99.0	100.0
GLM-4-9B	76.4	8.9	75.5	76.7	92.7
Gemma-3-27B-IT	100.0	99.8	100.0	100.0	100.0

Appendix

Table 67: Local symbol directions and simpler representations.

		Early write	Cut at 0.35		Cut at 0.5		Cut at 0.65
Receiver	Source	(0.2)	Write	Restore	Write	Restore	Write	Restore
Qwen3-14B-Base	Fitted	100.0	87.5	97.9	6.2	59.4	1.0	0.0
	Native	100.0	96.9	97.9	22.9	52.1	0.0	0.0
GLM-4-9B	Fitted	88.5	8.3	72.9	0.0	0.0	0.0	0.0
	Native	90.6	66.7	75.0	0.0	0.0	0.0	0.0
Gemma-3-27B-IT	Fitted	100.0	85.4	97.9	0.0	2.1	0.0	0.0

Appendix

Table 68: Where written facts must be readable.

Depth	Fitted	Scene-matched	Native	Text	Matched $-$ fitted	Native $-$ matched	Text $-$ native
0.2	88.5	87.5	90.6	99.0	$-$ 1.0 [-5.2, 3.1]	+3.1 [-2.1, 8.3]	+8.3 [1.0, 17.7]
0.35	8.3	40.6	66.7	99.0	+32.3 [24.0, 39.6]	+26.0 [11.5, 39.6]	+32.3 [20.8, 45.8]
0.5	0.0	0.0	0.0	99.0	+0.0 [0.0, 0.0]	+0.0 [0.0, 0.0]	+99.0 [96.9, 100.0]

Appendix

Table 69: GLM-4-9B: where the gap to the receiver’s own states arises.

Indices	Symbols
0–11	accordion, ambulance, axe, bakery, banana, barn, bed, beer, beetle, belt, bench, bicycle
12–23	blender, boat, boot, bowl, broccoli, broth, buffalo, bus, butterfly, cabbage, cabinet, camel
24–35	car, carpet, carrot, celery, cherry, chicken, church, cider, clam, clamp, coat, cocoa
36–47	coconut, cod, copper, cotton, cow, crab, crow, cucumber, curtain, deer, desk, dolphin
48–59	drawer, drill, elephant, espresso, factory, flea, flute, fork, fox, fridge, garlic, glass
60–71	glove, goat, gold, goose, grape, greenhouse, hat, hawk, helicopter, hoe, hospital, hotel

Appendix

Table 70: The shared inventory in index order.

Objects	Attribute	Values
Painted blocks	Color	green, orange, red, yellow
	Material	gold, iron, silver, steel
Candy	Color	green, orange, red, yellow
	Taste	bitter, salty, sour, sweet
Coats	Fibre	cotton, leather, silk, wool
	Buttons	copper, glass, plastic, wooden

Appendix

Table 71: Attribute space of the two-object confirmation.

World	Object	Described facts	Transmitted indices (decoded here)
0	BANQ	green, gold	62 (gold), 160 (green)
	CIRV	red, steel	133 (steel), 166 (red)
1	BANQ	red, gold	62 (gold), 166 (red)
	CIRV	green, steel	133 (steel), 160 (green)

Appendix

Table 72: One scene from descriptions to symbol indices.

Unit	In confirmation	Also in earlier banks
Object names	24	0
Full scenes	576	0
Evidence templates	12	0
Single-object fact tuples	48	48
Two-object combinations	72	14
Paired-value cells	36	7

Appendix

Table 73: New material and retained task structure.

Receiver	Delivery	Joint	Accuracy	Valid	Changed	Unchanged
Qwen3-32B	Early write	97.0	99.1	99.6	97.9	98.8
	Same packet as text	97.0	99.1	99.4	97.9	98.8
	Reassignment 1	1.9	31.7	59.5	8.6	20.3
	Reassignment 2	0.9	31.3	62.5	5.0	20.1
	Reassignment 3	0.7	32.5	64.5	5.1	22.1
Qwen3-14B-Base	Early write	97.7	99.3	100.0	98.1	99.3

Appendix

Table 74: Answer validity and selective use in automatic confirmation.

Receiver	Contrast	Gain (pp)	36-family interval	12-template interval
Qwen3-32B	Early minus middle	2.6	[1.2, 4.2]	[1.6, 3.7]
	Early minus reassigned	95.9	[92.1, 98.4]	[90.8, 98.9]
	Early minus text	0.0	[-0.5, 0.5]	[-0.5, 0.5]
Qwen3-14B-Base	Early minus middle	66.7	[63.1, 70.2]	[63.0, 70.6]
	Early minus reassigned	95.4	[92.8, 97.6]	[90.9, 98.1]
	Early minus text	0.0	[-0.7, 0.9]	[-0.5, 0.5]

Appendix

Table 75: Paired gains under both clustering levels.

Receiver	Middle write	Early write	Native states	Packet text	Evidence text
Qwen3-32B	97.7	99.8	100.0	99.8	99.5
Qwen3-14B-Base	27.1	99.5	99.8	100.0	100.0
GLM-4-9B	0.0	76.5	75.6	93.0	99.5
Gemma-3-27B-IT	26.7	100.0	100.0	100.0	100.0

Appendix

Table 76: Complete facts and contextual states at the selected depth.

Receiver	Condition	Joint	Accuracy	Valid	Changed	Unchanged
Qwen3-32B	No message	0.0	47.3	95.5	0.0	46.7
	Identity write	0.0	47.3	95.5	0.0	46.7
	Middle write	94.4	98.4	99.0	96.9	97.0
	Other-world packet	0.0	49.9	99.6	0.0	98.8
Qwen3-14B-Base	No message	0.0	50.8	100.0	0.0	51.6
	Identity write	0.0	50.8	100.0	0.0	51.6

Appendix

Table 77: Remaining automatic-message control conditions.

Receiver	Block	Edit norm	Residual norm	Edit/residual
Qwen3-32B	13	317.6	320.7	0.99
Qwen3-14B-Base	8	167.8	161.2	1.04
GLM-4-9B	8	50.6	44.0	1.15
Gemma-3-27B-IT	12	2186.0	2490.3	0.88

Appendix

Table 78: Actual update size at the selected early layer.

	Fitted directions		Native states
Receiver	Raw	Native-sized	Raw	Fitted-sized
Qwen3-32B	100.0	100.0	100.0	100.0
Qwen3-14B-Base	100.0	100.0	100.0	100.0
GLM-4-9B	88.5	87.5	90.6	89.6
Gemma-3-27B-IT	100.0	100.0	100.0	100.0

Appendix

Table 79: Matching actual edit norms at the early layer.

Native-state write depth	Joint success (%)	Valid answers (%)
Middle band	43.8	100.0
0.20	100.0	100.0
0.35	96.9	100.0
0.50	0.0	100.0
0.65	0.0	100.0
0.80	0.0	99.5

Appendix

Table 80: Gemma depth selection from native states.

Receiver	Automatic write [95% CI]	Same-message text	Complete-fact write
Qwen3-32B	86.9 [78.0, 94.6]	86.9	97.3
Gemma-3-27B-IT	91.1 [83.3, 97.9]	91.1	100.0

Appendix

Table 81: The fixed receiver configuration extends to four objects.

Objects	Messages	Fact recall (%)	Complete objects (%)	Mean bytes
Two	1152	4570/4608 (99.2)	2269/2304 (98.5)	22.6
Four	384	2968/3072 (96.6)	1432/1536 (93.2)	36.0

Appendix

Table 82: Facts retained by the automatic sender.

Objects	Receiver	Both	Write only	Text only	Neither
Two	Qwen3-32B	553	2	2	15
Two	Qwen3-14B-Base	557	2	2	11
Two	GLM-4-9B	409	16	93	54
Two	Gemma-3-27B-IT	559	0	1	12
Four	Qwen3-32B	162	1	1	24
Four	Gemma-3-27B-IT	171	0	0	17

Appendix

Table 83: Which scenes succeed under the two delivery routes?

Receiver	Objects	Write count	Write (%)	Text (%)	Write valid (%)
Qwen3-32B	Blocks	191/191	100.0	100.0	100.0
	Candy	189/190	99.5	99.5	99.9
	Coats	175/191	91.6	91.6	98.8
Qwen3-14B-Base	Blocks	191/191	100.0	100.0	100.0
	Candy	190/190	100.0	100.0	100.0
	Coats	178/191	93.2	93.2	99.9

Appendix

Table 84: Automatic-message performance by object domain.

Figure 7: Missing facts accompany most initial four-object failures. Raw scene counts among automatic early-write failures, out of 188 scenes per receiver in the single-attribute extension. The two categories exhaust observed failures. This content check describes errors rather than their unique causes.

	Early write (ms)		Text (ms)
Receiver	Median	p90	Median	p90
Qwen3-32B	1301	1456	1018	1177
Qwen3-14B-Base	640	717	515	591
GLM-4-9B	651	736	545	656
Gemma-3-27B-IT	1300	1406	1051	1163

Appendix

Table 85: Receiver-side delivery cost for complete-fact messages.

Receiver	Internal	Text	Full evidence	Reassigned Success / Valid
Qwen3-32B	95.3	94.1	98.6	0.0 / 97.3
Qwen3-14B-Base	95.7	96.9	100.0	0.0 / 99.9
GLM-4-9B	41.8	70.3	87.7	0.0 / 62.2
Gemma-3-27B-IT	94.4	97.7	99.7	0.1 / 76.8

Appendix

Table 86: Communicated facts support four-object conjunction queries.

Receiver	Message	Joint count	Accuracy	Valid	$R$	$U$	Text $J$
Qwen3-32B	Automatic	549/576	98.8	99.5	96.4	99.0	94.1
	Complete facts	558/576	99.2	99.6	97.6	99.3	97.7
Qwen3-14B-Base	Automatic	551/576	98.9	100.0	97.2	98.4	96.9
	Complete facts	563/576	99.4	100.0	98.4	99.3	98.4
GLM-4-9B	Automatic	241/576	80.1	99.1	66.5	68.9	70.3
	Complete facts	257/576	80.6	99.9	67.0	70.1	89.2

Appendix

Table 87: Answer accuracy and selective use on conjunction queries.

Receiver	Internal $-$ text [95% CI]	Internal $-$ reassigned [95% CI]	Valid answers (%) Draws 1 / 2 / 3
Qwen3-32B	1.2 [-0.5, 3.0]	95.3 [92.9, 97.6]	94.4/99.0/98.4
Qwen3-14B-Base	-1.2 [-2.8, 0.2]	95.7 [93.4, 97.6]	99.8/100.0/100.0
GLM-4-9B	-28.5 [-35.4, -20.8]	41.8 [36.1, 47.6]	57.1/63.4/66.1
Gemma-3-27B-IT	-3.3 [-5.2, -1.6]	94.4 [92.1, 96.4]	88.6/56.9/84.9

Appendix

Table 88: Paired conjunction comparisons and reassignment validity.

Receiver	Message	Twelve	First	Middle	Last	Text	Embedding
Qwen3-32B	Automatic	95.3	87.2	95.1	95.3	94.8	94.8
	Complete facts	98.8	94.6	98.8	98.6	99.3	99.3
Qwen3-14B-Base	Automatic	99.0	97.0	99.1	99.0	99.3	99.3
	Complete facts	99.1	97.2	99.8	99.0	100.0	100.0
GLM-4-9B	Automatic	80.6	9.9	82.8	83.5	87.2	88.9
	Complete facts	76.4	8.9	75.5	76.7	92.7	93.2

Appendix

Table 89: Simple local representations on a separate two-object bank.

Receiver	Message	Twelve $-$ mean single [95% CI]	Twelve $-$ embedding [95% CI]
Qwen3-32B	Automatic	2.8 [1.0, 5.1]	0.5 [-0.7, 1.9]
	Complete facts	1.4 [0.6, 2.4]	-0.5 [-1.2, 0.0]
Qwen3-14B-Base	Automatic	0.6 [0.0, 1.2]	-0.3 [-1.2, 0.3]
	Complete facts	0.5 [-0.8, 1.6]	-0.9 [-2.1, 0.0]
GLM-4-9B	Automatic	21.8 [17.1, 26.2]	-8.3 [-12.0, -4.9]
	Complete facts	22.7 [17.3, 27.7]	-16.8 [-27.1, -7.3]

Appendix

Table 90: What twelve-context fitting adds to the tested alternatives.

Receiver	$\gamma$	Internal	Message text	Full evidence	Internal validity
Qwen3-14B-Base	1	0.1	24.3	100.0	99.2
Qwen3-32B	1	0.3	11.8	100.0	90.1
GLM-4-9B	0	0.0	7.9	99.0	96.8
Qwen3.5-27B	0 †	0.0	17.8	100.0	60.5

Appendix

Table 91: Selective fact use with the earlier sparse extractor.

	Candy / blocks			Coats
Message source	Text	Old	Targeted	Text	Old	Targeted
Original sparse	4.2	1.0	1.0	5.2	1.0	7.3
Centered sparse	70.8	0.0	38.5	8.3	1.0	13.5
Prompted selection	100.0	0.0	95.8	100.0	0.0	77.1
Complete facts	100.0	2.1	99.0	100.0	1.0	87.5

Appendix

Table 92: Message sources and delivery routes on development material.

	Neutral + direction			Coefficient update
Receiver	Each layer	Midpoint	Rescaled mid	Each layer	Midpoint
Qwen3-32B	99.0	54.2	61.5	0.0	0.0
Qwen3-14B-Base	13.5	13.5	15.6	0.0	0.0
GLM-4-9B	6.2	8.3	8.3	0.0	0.0

Appendix

Table 93: Dictionary depth and update rule at a common strength.

Receiver	Symbol contrast (%)	J-lens (%)	J-lens minus contrast (pp)
Qwen3-32B	97.9	2.1	-95.8 [-99.0, -92.7]
Qwen3-14B-Base	10.4	1.0	-9.4 [-15.6, -4.2]
GLM-4-9B	0.0	1.0	1.0 [0.0, 3.1]
Llama-3.1-8B-Instruct	0.0	0.0	0.0 [0.0, 0.0]

Appendix

Table 94: Dictionary construction matters at a matched single layer.

Multi-LLM Collaboration

Multi-LLM collaboration studies how large language models share work, exchange information, and combine outputs to improve capability, reliability, or cost.

See all 85 papers

Sep 28, 2026

Draft-KV: Learning Useful Latent Communication Between Language Models

Latent communication passes internal states between language models instead of decoded text, but higher receiver accuracy does not show that the receiver used the message content. Across five method-dataset pairs, replacing each message with one from an unrelated question changes accuracy by at most 0.60 points, even when communication adds 15.44 points over the receiver alone. Thus the interface can supply the gain while making the sharer dispensable. Draft-KV instead sends the key-value states formed while the sharer drafts an answer to the current question. Linear projections place these states in a side memory read through a gated attention branch, and progressive training moves from message reconstruction to answer supervision under a guard on harm from mismatched messages. Both models remain frozen and the interface trains 1.05M parameters, 348x fewer than C2C. With a Qwen3-8B sharer, a frozen Qwen2.5-0.5B-Instruct receiver reaches 78.04% on MMLU-Redux, versus 37.45% alone and 36.40% with reassigned messages. At fixed interface size, scaling the sharer from 0.6B to 8B raises accuracy from 46.11% to 78.04%; communication also transfers to held-out tasks and can exceed both models when each holds different evidence.

Linquan Wu, Shichang Meng, Tianxiang Jiang +7

City University of Hong Kong · University of Science and Technology of China · University of Electronic Science and Technology of China +3

Knowledge Augmentation for Language Models Multi-LLM Collaboration

Ask PDF

May 11, 2026

The Bicameral Model: Bidirectional Hidden-State Coupling Between Parallel Language Models

Existing multi-model and tool-augmented systems communicate by generating text, serializing every exchange through the output vocabulary. Can two pretrained language models instead coordinate through a continuous, concurrent channel? The Bicameral Model couples two frozen language models through a trainable neural interface on their intermediate hidden states. At every generation step, both models run in lockstep: a primary model drives the task while an auxiliary model operates tools, solves constraints, or executes code, with both conditioning on each other's activations through a translation network and a learned suppression gate (

\sim

1% of combined parameters). The gate learns a selective communication protocol from task loss alone, without a prescribed format. We demonstrate the mechanism across three tool backends. On arithmetic, coupling two 0.5B models with a calculator raises accuracy from 36% to 96%. On logic grid puzzles, coupling two 0.6B models with a Z3 solver achieves

1.7\times

the unaugmented baseline on ZebraLogic. On mathematical reasoning, coupling with a Python sandbox enables the auxiliary to generate problem-specific code from hidden-state signals alone, without ever seeing the problem text.

Cedric Flamant, Udaya Ghai, Kanna Shimizu

AWS Agentic AI

LLM Tool Use Multi-LLM Collaboration

Ask PDF

Sep 25, 2026

Communication between Frozen Large Language Models via Prompt Optimization in a Referential Game

We study communication between two frozen large language models from different providers, with different tokenizers, accessed through their API endpoints. The two play a referential game: one sees an object and describes it in a short fixed-length message over a small alphabet; the other must pick that object out of a candidate set. Neither model's weights are updated. Each agent's prompt is rewritten by an isolated prompt optimizer whose reflection model reads that agent's scored interactions. In the positional setting, optimized prompts carry a shared code that generalizes to held-out objects above a measured no-codebook baseline, including when the memory window is removed. In a second setting, independent per-letter blocks no longer fit within the message, although a whole-object place value code does. The base system fails to establish reliable communication: the sender struggles to retain an injective rule, and the receiver has too few confirmed examples in view. A sender collision penalty, retention of successful interactions, and sequential optimization enable successful place value communication in some runs. Outcomes vary across runs and reflection models. In successful runs, the protocol is written into the optimized prompts, where it can be read and audited directly.

Vivek Anand, Muthu Chandrasekaran, Shiva Chaitanya

School of Electrical and Computer Engineering, Georgia Institute of Technology, Atlanta, GA, USA · Work done while at Interactly.ai

Emergent Communication LLM Prompting Multi-Agent LLMs Large Language Model-Guided Optimization

Ask PDF

Abstract

Multi-LLM Collaboration

Explore similar work

Draft-KV: Learning Useful Latent Communication Between Language Models

The Bicameral Model: Bidirectional Hidden-State Coupling Between Parallel Language Models

Communication between Frozen Large Language Models via Prompt Optimization in a Referential Game