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
68 → 70 1.00 1.47 feature 68 feature 70
11 → 70 1.00 1.46 feature 11 feature 70
11 → 177 1.00 1.45 feature 11 feature 177
68 → 177 0.99 1.47 feature 68 feature 177
169 → 138 0.99 1.43 feature 169 feature 138
11 → 112 0.99 1.45 feature 11 feature 112
68 → 112 0.99 1.46 feature 68 feature 112
11 → 5 0.99 1.45 feature 11 feature 5

Duplicate pairs (rename/split candidates): 2≈7 (feature 2); 2≈15 (feature 2); 2≈18 (feature 2); 2≈85 (feature 2); 2≈88 (feature 2); 2≈151 (feature 2)

Block B1 (224 features)

parent → child R PMI parent label child label
381 → 288 0.52 1.86 feature 381 feature 288

Block B2 (224 features)

parent → child R PMI parent label child label
569 → 522 0.66 2.14 feature 569 feature 522

Block B3 (224 features)

Block B4 (224 features)

Block B5 (224 features)

parent → child R PMI parent label child label
1286 → 1292 0.83 4.19 feature 1286 feature 1292

Block B6 (224 features)

Block B7 (224 features)