metrics

In-block (same-level) directed edges

Layer 0　·　PCFG toy 4L d_model=448 / pcfg　·　matryoshka_hook_resid_post_L0　·　1,792 latents in 8 blocks　·　1,016,600 tokens over 3400 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)

• 763 directed edges, 59 duplicate pairs, 631 survive PMI>0; PolyFrac 94%, Gini 0.967.
• S_res: 10/631 edges are genuine refinements (0 children untestable).
• In-block superparents: 4 (e.g. F129 feature 129: 156 children, fires 100%).

parent → child	R	PMI	parent label	child label
129 → 139	1.00	0.00	feature 129	feature 139
129 → 44	1.00	0.00	feature 129	feature 44
129 → 83	1.00	0.00	feature 129	feature 83
129 → 158	1.00	0.00	feature 129	feature 158
129 → 205	1.00	0.00	feature 129	feature 205
129 → 10	1.00	0.00	feature 129	feature 10
129 → 98	1.00	0.00	feature 129	feature 98
129 → 120	1.00	0.00	feature 129	feature 120

Duplicate pairs (rename/split candidates): 10≈158 (feature 10); 13≈14 (feature 13); 13≈89 (feature 13); 14≈15 (feature 14); 14≈25 (feature 14); 14≈34 (feature 14)

Block B1 (224 features)

• 8 directed edges, 0 duplicate pairs, 8 survive PMI>0; PolyFrac 50%, Gini 0.964.
• S_res: 0/8 edges are genuine refinements (0 children untestable).

parent → child	R	PMI	parent label	child label
245 → 401	1.00	3.01	feature 245	feature 401
292 → 401	1.00	2.91	feature 292	feature 401
386 → 401	1.00	3.83	feature 386	feature 401
410 → 401	1.00	3.16	feature 410	feature 401
420 → 401	0.99	2.71	feature 420	feature 401
375 → 401	0.99	2.87	feature 375	feature 401
359 → 259	0.82	4.81	feature 359	feature 259
351 → 401	0.59	2.39	feature 351	feature 401

Block B2 (224 features)

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

parent → child	R	PMI	parent label	child label
465 → 617	0.53	3.25	feature 465	feature 617

Block B3 (224 features)

• 3 directed edges, 0 duplicate pairs, 3 survive PMI>0; PolyFrac 100%, Gini 0.987.
• S_res: 2/3 edges are genuine refinements (0 children untestable).

parent → child	R	PMI	parent label	child label
674 → 798	0.91	4.46	feature 674	feature 798
758 → 798	0.88	4.29	feature 758	feature 798
830 → 798	0.85	4.71	feature 830	feature 798

Block B4 (224 features)

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

Block B5 (224 features)

• 5 directed edges, 0 duplicate pairs, 5 survive PMI>0; PolyFrac 25%, Gini 0.978.
• S_res: 0/5 edges are genuine refinements (0 children untestable).

parent → child	R	PMI	parent label	child label
1266 → 1291	0.86	5.26	feature 1266	feature 1291
1178 → 1121	0.82	5.55	feature 1178	feature 1121
1282 → 1121	0.76	4.81	feature 1282	feature 1121
1256 → 1182	0.71	5.09	feature 1256	feature 1182
1173 → 1230	0.71	5.18	feature 1173	feature 1230

Block B6 (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
1364 → 1567	0.52	5.26	feature 1364	feature 1567

Block B7 (224 features)

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

parent → child	R	PMI	parent label	child label
1607 → 1634	0.90	6.16	feature 1607	feature 1634
1621 → 1573	0.88	5.69	feature 1621	feature 1573