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)

• 78 directed edges, 12 duplicate pairs, 78 survive PMI>0; PolyFrac 49%, Gini 0.953.
• S_res: 3/75 edges are genuine refinements (0 children untestable).

parent → child	R	PMI	parent label	child label
58 → 190	1.00	0.93	feature 58	feature 190
58 → 169	1.00	0.93	feature 58	feature 169
58 → 47	1.00	0.93	feature 58	feature 47
58 → 197	1.00	0.93	feature 58	feature 197
58 → 182	1.00	0.93	feature 58	feature 182
58 → 106	0.99	0.92	feature 58	feature 106
149 → 6	0.99	2.91	feature 149	feature 6
58 → 55	0.99	0.92	feature 58	feature 55

Duplicate pairs (rename/split candidates): 2≈142 (feature 2); 35≈51 (feature 35); 35≈143 (feature 35); 35≈148 (feature 35); 35≈180 (feature 35); 47≈182 (feature 47)

Block B1 (224 features)

• 5 directed edges, 1 duplicate pairs, 5 survive PMI>0; PolyFrac 0%, Gini 0.994.
• S_res: 0/5 edges are genuine refinements (0 children untestable).

parent → child	R	PMI	parent label	child label
301 → 351	1.00	4.07	feature 301	feature 351
301 → 299	1.00	4.07	feature 301	feature 299
353 → 432	1.00	2.59	feature 353	feature 432
301 → 408	0.99	4.06	feature 301	feature 408
301 → 398	0.93	3.99	feature 301	feature 398

Duplicate pairs (rename/split candidates): 351≈408 (feature 351)

Block B2 (224 features)

• 6 directed edges, 1 duplicate pairs, 6 survive PMI>0; PolyFrac 25%, Gini 0.981.
• S_res: 1/3 edges are genuine refinements (0 children untestable).

parent → child	R	PMI	parent label	child label
645 → 518	1.00	2.87	feature 645	feature 518
649 → 518	1.00	2.62	feature 649	feature 518
584 → 632	1.00	4.27	feature 584	feature 632
584 → 514	0.99	4.26	feature 584	feature 514
600 → 518	0.97	2.53	feature 600	feature 518
639 → 454	0.51	1.79	feature 639	feature 454

Duplicate pairs (rename/split candidates): 557≈643 (feature 557)

Block B3 (224 features)

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

Block B4 (224 features)

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

Block B5 (224 features)

• 0 directed edges, 1 duplicate pairs, 0 survive PMI>0; PolyFrac 0%, Gini 0.000.

Duplicate pairs (rename/split candidates): 1252≈1312 (feature 1252)

Block B6 (224 features)

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

parent → child	R	PMI	parent label	child label
1465 → 1475	0.87	3.14	feature 1465	feature 1475

Block B7 (224 features)

• 2 directed edges, 0 duplicate pairs, 2 survive PMI>0; PolyFrac 100%, Gini 0.991.
• S_res: 0/2 edges are genuine refinements (0 children untestable).

parent → child	R	PMI	parent label	child label
1729 → 1651	1.00	5.37	feature 1729	feature 1651
1632 → 1651	0.75	3.21	feature 1632	feature 1651