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
222 → 152 1.00 1.52 feature 222 feature 152
186 → 152 1.00 1.55 feature 186 feature 152
186 → 30 0.99 1.55 feature 186 feature 30
222 → 30 0.99 1.52 feature 222 feature 30
99 → 176 0.99 1.16 feature 99 feature 176
9 → 30 0.99 2.57 feature 9 feature 30
152 → 30 0.99 4.10 feature 152 feature 30
123 → 30 0.99 1.57 feature 123 feature 30

Duplicate pairs (rename/split candidates): 41≈181 (feature 41); 99≈161 (feature 99); 123≈142 (feature 123); 123≈186 (feature 123); 123≈222 (feature 123); 186≈222 (feature 186)

Block B1 (224 features)

parent → child R PMI parent label child label
424 → 330 0.70 5.50 feature 424 feature 330
430 → 317 0.63 3.48 feature 430 feature 317
283 → 336 0.57 1.88 feature 283 feature 336

Block B2 (224 features)

Block B3 (224 features)

parent → child R PMI parent label child label
711 → 888 0.67 2.67 feature 711 feature 888
720 → 888 0.65 2.49 feature 720 feature 888

Block B4 (224 features)

Block B5 (224 features)

Block B6 (224 features)

Block B7 (224 features)