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 · 1,022,000 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)

parent → child R PMI parent label child label
0 → 106 0.99 0.85 feature 0 feature 106
86 → 120 0.98 2.39 feature 86 feature 120
21 → 120 0.98 1.84 feature 21 feature 120
5 → 106 0.98 1.55 feature 5 feature 106
0 → 209 0.98 0.84 feature 0 feature 209
81 → 120 0.98 1.52 feature 81 feature 120
161 → 120 0.98 1.99 feature 161 feature 120
0 → 67 0.98 0.84 feature 0 feature 67

Duplicate pairs (rename/split candidates): 21≈68 (feature 21); 21≈161 (feature 21); 35≈120 (feature 35); 35≈127 (feature 35); 35≈191 (feature 35); 120≈191 (feature 120)

Block B1 (224 features)

Block B2 (224 features)

Block B3 (224 features)

Block B4 (224 features)

Block B5 (224 features)

parent → child R PMI parent label child label
1156 → 1275 0.79 3.76 feature 1156 feature 1275

Block B6 (224 features)

Block B7 (224 features)