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 · 598,000 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
121 → 118 1.00 1.39 feature 121 feature 118
121 → 204 1.00 1.39 feature 121 feature 204
121 → 25 1.00 1.38 feature 121 feature 25
121 → 101 0.99 1.38 feature 121 feature 101
121 → 209 0.99 1.38 feature 121 feature 209
212 → 42 0.99 1.34 feature 212 feature 42
212 → 160 0.99 1.34 feature 212 feature 160
164 → 118 0.98 1.79 feature 164 feature 118

Duplicate pairs (rename/split candidates): 9≈169 (feature 9); 101≈169 (feature 101); 121≈161 (feature 121); 121≈164 (feature 121); 161≈164 (feature 161); 164≈209 (feature 164)

Block B1 (224 features)

parent → child R PMI parent label child label
297 → 241 0.68 2.00 feature 297 feature 241
337 → 375 0.63 1.87 feature 337 feature 375
309 → 241 0.53 2.03 feature 309 feature 241
321 → 241 0.51 1.86 feature 321 feature 241
296 → 241 0.50 1.53 feature 296 feature 241

Block B2 (224 features)

parent → child R PMI parent label child label
608 → 517 0.74 5.36 feature 608 feature 517

Block B3 (224 features)

Block B4 (224 features)

Block B5 (224 features)

parent → child R PMI parent label child label
1152 → 1207 0.97 3.92 feature 1152 feature 1207

Block B6 (224 features)

Block B7 (224 features)