Already-public plots · numerical archive only

Measurements

The five plots below are the public renders. Numbers in the table come from the archive operating-map table (22 cells, 330 subcells, 65,868 merges). Llama-2-13B is an addendum and is excluded from this n=22 freeze. Raw arrays stay in Source / Archive, not in this hero.

**Gain versus constructive fraction.** Each point is one ROME cell. Perturbation gain (median |drop| / dose, log axis) versus the share of merges whose logit drop is negative. The vertical cut at gain = 8 separates the destructive and constructive clouds. Source / Archive: gain-law table.

**Cross-talk versus merge size.** Spearman between the projected cross-term and the observed drop, plotted against group size g. High-gain cells stay positive at small g; low-gain deep cells (Qwen2.5-14B L36, gpt2-xl L36) sit below zero. Source / Archive: cross-term alignment table.

**Geometry-ordered admission.** Mean logits of damage avoided per admitted edit when the cheapest 25% or 50% of pairs are kept, comparing key-geometry order against a magnitude-only order. The high-gain benefit +0.647 is measured retrospectively as a member-target quantity under fixed composition. Source / Archive: admission-benefit table.

**Dose response by group size.** Median |drop| per edit against g. High-gain cells remain loud; low-gain cells stay quiet even as groups grow. Source / Archive: operating-map table.

**Partial geometry correlation.** Partial ρ(I_cos, drop | I_mag) by g, split by regime. The horizontal line is the pre-registered partial floor (+0.15). Saturated cells are drawn hollow. Each cell carries its own geometry-valid window: g ≤ 5 at the reference cell, wider at six further qualifying cells. Source / Archive: operating-map table.

22-cell operating table

Gain and constructive fraction are cell-level. Partial ρ columns are three-seed means at that g. High-gain / low-gain is the gain = 8 screen.

model	L	gain	frac<0	regime	ρ g=2	ρ g=3	ρ g=5	ρ g=10
Llama-3.2-1B	8	64.524	0.041	high-gain	0.234	0.275	0.239	0.155
Qwen2.5-3B	18	37.240	0.332	high-gain	0.100	0.161	0.180	0.208
Llama-3.2-3B	21	32.011	0.022	high-gain	0.558	0.553	0.545	0.455
Llama-3.2-1B	14	28.665	0.013	high-gain	0.317	0.281	0.185	0.094
Mistral-7B-v0.3	24	28.295	0.278	high-gain	0.325	0.388	0.404	0.506
Qwen2.5-1.5B	14	28.149	0.242	high-gain	0.146	0.242	0.276	0.341
Llama-3.2-1B	12	27.207	0.025	high-gain	0.469	0.484	0.320	0.150
Llama-3.1-8B	24	22.207	0.127	high-gain	0.358	0.447	0.521	0.579
gemma-2-2b	13	20.017	0.287	high-gain	0.080	0.030	0.036	0.052
Phi-3.5-mini	16	15.636	0.252	high-gain	0.099	0.146	0.215	0.234
gemma-2-9b	31	12.491	0.119	high-gain	0.253	0.180	0.098	0.020
gemma-2-2b	19	12.483	0.115	high-gain	0.127	0.105	0.104	0.050
Mistral-Nemo-Base-2407	30	10.687	0.423	high-gain	0.019	0.097	0.161	0.305
Phi-3.5-mini	24	3.910	0.460	low-gain	-0.245	-0.085	0.087	0.481
Qwen2.5-3B	27	3.625	0.430	low-gain	0.037	0.099	0.318	0.576
Qwen2.5-7B	21	3.451	0.349	low-gain	0.144	0.203	0.337	0.359
gpt2-xl	24	3.410	0.711	low-gain	-0.125	-0.142	-0.050	0.074
Qwen2.5-1.5B	21	3.299	0.420	low-gain	-0.012	0.050	0.164	0.325
Qwen2.5-14B	36	3.102	0.836	low-gain	-0.458	-0.392	-0.399	-0.193
Qwen2.5-1.5B	24	1.110	0.471	low-gain	-0.031	-0.001	-0.007	0.057
gpt2-xl	36	0.841	0.725	low-gain	-0.160	-0.169	-0.211	-0.224
gpt-neox-20b	33	0.103	0.634	low-gain	0.019	0.030	0.102	0.090