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.
| parent → child | R | PMI | parent label | child label |
|---|---|---|---|---|
| 81 → 140 | 1.00 | 1.45 | feature 81 | feature 140 |
| 75 → 140 | 1.00 | 1.46 | feature 75 | feature 140 |
| 36 → 209 | 1.00 | 1.44 | feature 36 | feature 209 |
| 53 → 176 | 1.00 | 2.47 | feature 53 | feature 176 |
| 81 → 176 | 1.00 | 1.45 | feature 81 | feature 176 |
| 69 → 176 | 1.00 | 2.46 | feature 69 | feature 176 |
| 75 → 196 | 1.00 | 1.46 | feature 75 | feature 196 |
| 81 → 196 | 1.00 | 1.45 | feature 81 | feature 196 |
Duplicate pairs (rename/split candidates): 7≈43 (feature 7); 23≈75 (feature 23); 23≈81 (feature 23); 23≈196 (feature 23); 53≈69 (feature 53); 75≈81 (feature 75)
| parent → child | R | PMI | parent label | child label |
|---|---|---|---|---|
| 331 → 424 | 0.91 | 4.18 | feature 331 | feature 424 |
| 331 → 436 | 0.88 | 4.15 | feature 331 | feature 436 |
| 424 → 436 | 0.80 | 4.82 | feature 424 | feature 436 |
| 229 → 424 | 0.76 | 3.70 | feature 229 | feature 424 |
| 330 → 436 | 0.67 | 4.20 | feature 330 | feature 436 |
| 295 → 423 | 0.63 | 1.78 | feature 295 | feature 423 |
| 230 → 423 | 0.62 | 2.03 | feature 230 | feature 423 |
| 432 → 423 | 0.59 | 2.00 | feature 432 | feature 423 |
Duplicate pairs (rename/split candidates): 229≈331 (feature 229)
| parent → child | R | PMI | parent label | child label |
|---|---|---|---|---|
| 589 → 454 | 0.89 | 3.65 | feature 589 | feature 454 |
| 604 → 610 | 0.59 | 3.60 | feature 604 | feature 610 |
| 610 → 552 | 0.59 | 3.87 | feature 610 | feature 552 |
| 604 → 552 | 0.59 | 3.59 | feature 604 | feature 552 |
| parent → child | R | PMI | parent label | child label |
|---|---|---|---|---|
| 1668 → 1683 | 0.91 | 3.14 | feature 1668 | feature 1683 |