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 · 646,000 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
167 → 221 1.00 2.02 feature 167 feature 221
167 → 188 0.98 2.01 feature 167 feature 188
99 → 217 0.98 1.42 feature 99 feature 217
57 → 144 0.98 1.80 feature 57 feature 144
167 → 40 0.90 1.91 feature 167 feature 40
167 → 120 0.89 1.91 feature 167 feature 120
223 → 51 0.88 0.85 feature 223 feature 51
167 → 173 0.84 1.85 feature 167 feature 173

Duplicate pairs (rename/split candidates): 13≈100 (feature 13); 13≈184 (feature 13); 25≈100 (feature 25); 173≈188 (feature 173)

Block B1 (224 features)

parent → child R PMI parent label child label
318 → 247 0.82 3.10 feature 318 feature 247

Block B2 (224 features)

parent → child R PMI parent label child label
590 → 553 0.57 1.85 feature 590 feature 553

Block B3 (224 features)

Block B4 (224 features)

Block B5 (224 features)

Block B6 (224 features)

Block B7 (224 features)