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
185 → 165 1.00 1.20 feature 185 feature 165
185 → 184 0.99 1.19 feature 185 feature 184
185 → 102 0.99 1.19 feature 185 feature 102
102 → 184 0.99 2.25 feature 102 feature 184
193 → 119 0.99 1.12 feature 193 feature 119
125 → 119 0.99 1.08 feature 125 feature 119
185 → 24 0.99 1.18 feature 185 feature 24
185 → 17 0.97 1.17 feature 185 feature 17

Duplicate pairs (rename/split candidates): 84≈125 (feature 84); 84≈193 (feature 84); 125≈193 (feature 125); 154≈169 (feature 154)

Block B1 (224 features)

Block B2 (224 features)

parent → child R PMI parent label child label
500 → 617 0.71 3.53 feature 500 feature 617
598 → 617 0.61 2.39 feature 598 feature 617

Block B3 (224 features)

Block B4 (224 features)

Block B5 (224 features)

parent → child R PMI parent label child label
1141 → 1291 0.60 2.30 feature 1141 feature 1291
1264 → 1291 0.54 3.34 feature 1264 feature 1291

Block B6 (224 features)

Block B7 (224 features)