metrics

In-block (same-level) directed edges

Layer 1 · PCFG toy 4L d_model=448 / pcfg · matryoshka_hook_resid_post_L1 · 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)

parent → child R PMI parent label child label
201 → 8 1.00 1.33 feature 201 feature 8
120 → 8 1.00 1.41 feature 120 feature 8
13 → 8 1.00 1.44 feature 13 feature 8
109 → 8 0.99 1.36 feature 109 feature 8
7 → 8 0.99 2.91 feature 7 feature 8
31 → 127 0.99 1.31 feature 31 feature 127
34 → 127 0.99 1.35 feature 34 feature 127
13 → 127 0.99 1.43 feature 13 feature 127

Duplicate pairs (rename/split candidates): 7≈101 (feature 7); 13≈31 (feature 13); 13≈34 (feature 13); 13≈46 (feature 13); 13≈52 (feature 13); 13≈97 (feature 13)

Block B1 (224 features)

Block B2 (224 features)

parent → child R PMI parent label child label
513 → 453 0.67 2.40 feature 513 feature 453

Block B3 (224 features)

parent → child R PMI parent label child label
708 → 675 0.59 3.77 feature 708 feature 675

Block B4 (224 features)

Block B5 (224 features)

Block B6 (224 features)

parent → child R PMI parent label child label
1407 → 1551 0.52 4.22 feature 1407 feature 1551

Block B7 (224 features)