metrics

In-block (same-level) directed edges

Layer 3 · PCFG toy 4L d_model=448 / pcfg · matryoshka_hook_resid_post_L3 · 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
22 → 27 1.00 0.82 feature 22 feature 27
18 → 27 1.00 1.37 feature 18 feature 27
88 → 27 1.00 1.40 feature 88 feature 27
144 → 27 1.00 1.37 feature 144 feature 27
154 → 52 0.97 0.89 feature 154 feature 52
46 → 52 0.97 0.90 feature 46 feature 52
119 → 34 0.97 1.92 feature 119 feature 34
139 → 34 0.96 1.89 feature 139 feature 34

Duplicate pairs (rename/split candidates): 3≈17 (feature 3); 3≈46 (feature 3); 3≈76 (feature 3); 3≈99 (feature 3); 3≈154 (feature 3); 5≈12 (feature 5)

Block B1 (224 features)

parent → child R PMI parent label child label
286 → 240 0.96 3.81 feature 286 feature 240
254 → 291 0.76 3.17 feature 254 feature 291
304 → 240 0.65 2.16 feature 304 feature 240
225 → 240 0.52 2.08 feature 225 feature 240

Block B2 (224 features)

Block B3 (224 features)

Block B4 (224 features)

Block B5 (224 features)

parent → child R PMI parent label child label
1280 → 1158 0.57 5.09 feature 1280 feature 1158

Block B6 (224 features)

Block B7 (224 features)