metrics

In-block (same-level) directed edges

Layer 2　·　PCFG toy 4L d_model=448 / pcfg　·　matryoshka_hook_resid_post_L2　·　1,792 latents in 8 blocks　·　656,000 tokens over 2000 docs　·　edge: reverse coverage ≥ 0.5, both endpoints fire ≥ 20

Parent→child within a block (asymmetric containment); co-extensive pairs are reported as duplicates (renames/splits), never edges.

Block B0 (224 features)

• 881 directed edges, 49 duplicate pairs, 811 survive PMI>0; PolyFrac 54%, Gini 0.888.
• S_res: 22/811 edges are genuine refinements (0 children untestable).
• In-block superparents: 2 (e.g. F94 feature 94: 186 children, fires 99%).

parent → child	R	PMI	parent label	child label
94 → 61	1.00	0.01	feature 94	feature 61
94 → 152	1.00	0.01	feature 94	feature 152
94 → 115	1.00	0.01	feature 94	feature 115
40 → 173	1.00	1.69	feature 40	feature 173
94 → 172	1.00	0.01	feature 94	feature 172
94 → 105	1.00	0.01	feature 94	feature 105
94 → 19	1.00	0.01	feature 94	feature 19
129 → 27	1.00	1.70	feature 129	feature 27

Duplicate pairs (rename/split candidates): 8≈184 (feature 8); 17≈46 (feature 17); 17≈59 (feature 17); 17≈178 (feature 17); 17≈212 (feature 17); 26≈46 (feature 26)

Block B1 (224 features)

• 30 directed edges, 4 duplicate pairs, 30 survive PMI>0; PolyFrac 62%, Gini 0.959.
• S_res: 3/30 edges are genuine refinements (0 children untestable).

parent → child	R	PMI	parent label	child label
420 → 292	0.99	2.25	feature 420	feature 292
403 → 292	0.99	2.04	feature 403	feature 292
250 → 292	0.99	2.55	feature 250	feature 292
372 → 292	0.99	4.51	feature 372	feature 292
277 → 292	0.99	5.03	feature 277	feature 292
266 → 394	0.99	4.79	feature 266	feature 394
266 → 231	0.98	4.78	feature 266	feature 231
273 → 414	0.92	2.67	feature 273	feature 414

Duplicate pairs (rename/split candidates): 224≈414 (feature 224); 231≈415 (feature 231); 266≈415 (feature 266); 310≈414 (feature 310)

Block B2 (224 features)

• 12 directed edges, 0 duplicate pairs, 12 survive PMI>0; PolyFrac 43%, Gini 0.969.
• S_res: 2/12 edges are genuine refinements (0 children untestable).

parent → child	R	PMI	parent label	child label
654 → 482	0.96	3.97	feature 654	feature 482
520 → 617	0.83	2.65	feature 520	feature 617
520 → 635	0.73	2.53	feature 520	feature 635
615 → 545	0.70	1.89	feature 615	feature 545
612 → 545	0.68	1.99	feature 612	feature 545
521 → 466	0.66	2.24	feature 521	feature 466
600 → 466	0.62	2.13	feature 600	feature 466
520 → 609	0.60	2.33	feature 520	feature 609

Block B3 (224 features)

• 1 directed edges, 0 duplicate pairs, 1 survive PMI>0; PolyFrac 0%, Gini 0.996.
• S_res: 0/1 edges are genuine refinements (0 children untestable).

parent → child	R	PMI	parent label	child label
707 → 894	0.57	3.60	feature 707	feature 894

Block B4 (224 features)

• 1 directed edges, 0 duplicate pairs, 1 survive PMI>0; PolyFrac 0%, Gini 0.996.
• S_res: 0/1 edges are genuine refinements (0 children untestable).

parent → child	R	PMI	parent label	child label
971 → 1022	0.85	4.64	feature 971	feature 1022

Block B5 (224 features)

• 0 directed edges, 0 duplicate pairs, 0 survive PMI>0; PolyFrac 0%, Gini 0.000.

Block B6 (224 features)

• 0 directed edges, 0 duplicate pairs, 0 survive PMI>0; PolyFrac 0%, Gini 0.000.

Block B7 (224 features)

• 0 directed edges, 0 duplicate pairs, 0 survive PMI>0; PolyFrac 0%, Gini 0.000.