metrics

Exp 0 - metrics report

Layer 2　·　PCFG toy 4L d_model=448 / pcfg　·　matryoshka_hook_resid_post_L2　·　1,792 latents in 8 blocks　·　656,000 tokens over 2000 docs　·　edge: reverse coverage ≥ 0.5, both endpoints fire ≥ 20

Block pair 0->1 - 493 candidate edges

• Out-degree: 55 parents, 167 children, 100 multi-parented (PolyFrac 59.9%); top-1 parent holds 33.3% of edges, Gini 0.945, max out-degree 164.
• Superparents (out-degree flag): 2 (2 also pass the old fire-rate AND-gate) — e.g. feature 105 feature 105: 164 children, fires on 97.5% of tokens
• Independence null: mean edge PMI 0.65; 252 edges (51.1%) at chance level (PMI < 0.5). 3 edges dropped by the joint-support guard (n_joint < 30).
• Recon-ablation contribution filter (Tree-SAE-inspired baseline): 485/493 edges pass (98.4%).
• Frequency control: mean survival 0.746 over 482 testable edges; 135 (28.0%) are frequency-driven (survival < 0.5).
• Sibling redundancy (global Jaccard — confounded proxy, not the splitting verdict; the parent-conditioned version is in the stage-03 second pass): mean 0.119 over 35 parents; 0 over the 0.5 global threshold.
• Joint-child (exact union, parents with edges): R_supp mean 0.993, R_mass mean 0.993; 9 parents with one child holding >=90% of their energy (rename candidates).
• Joint-child coverage (min(1, ΣF) upper bound — saturates when children co-fire, kept only for contrast with the exact union): 0.372.

parent -> child	R	F	PMI	recon P/C gain	recon?	surv	sib	parent label	child label
105 -> 362	1.00	0.00	0.03	0.09/0.01	Y	-	0.02	feature 105	feature 362
105 -> 354	1.00	0.01	0.03	0.18/-0.01	n	1.00	0.02	feature 105	feature 354
105 -> 398	1.00	0.05	0.03	0.20/0.03	Y	1.00	0.02	feature 105	feature 398
105 -> 353	1.00	0.04	0.02	0.12/0.07	Y	1.00	0.02	feature 105	feature 353
105 -> 273	1.00	0.10	0.02	0.15/0.08	Y	1.00	0.02	feature 105	feature 273
105 -> 337	1.00	0.02	0.02	0.16/0.34	Y	1.00	0.02	feature 105	feature 337
105 -> 258	0.99	0.01	0.02	0.16/0.32	Y	1.00	0.02	feature 105	feature 258
105 -> 335	0.99	0.02	0.02	0.16/0.26	Y	1.00	0.02	feature 105	feature 335

Block pair 1->2 - 21 candidate edges

• Out-degree: 21 parents, 1 children, 1 multi-parented (PolyFrac 100.0%); top-1 parent holds 4.8% of edges, Gini 0.906, max out-degree 1.
• Superparents (out-degree flag): 0 (0 also pass the old fire-rate AND-gate)
• Independence null: mean edge PMI 2.85; 0 edges (0.0%) at chance level (PMI < 0.5). 0 edges dropped by the joint-support guard (n_joint < 30).
• Recon-ablation contribution filter (Tree-SAE-inspired baseline): 14/21 edges pass (66.7%).
• Frequency control: mean survival nan over 0 testable edges; 0 (0.0%) are frequency-driven (survival < 0.5).
• Sibling redundancy: n/a (child block not in SIBLING_BLOCKS).
• Joint-child (exact union, parents with edges): R_supp mean 0.959, R_mass mean 0.960; 0 parents with one child holding >=90% of their energy (rename candidates).
• Joint-child coverage (min(1, ΣF) upper bound — saturates when children co-fire, kept only for contrast with the exact union): 0.060.

parent -> child	R	F	PMI	recon P/C gain	recon?	surv	sib	parent label	child label
345 -> 561	1.00	0.03	2.38	0.01/0.05	Y	-	-	feature 345	feature 561
369 -> 561	1.00	0.03	2.38	0.01/0.05	n	-	-	feature 369	feature 561
385 -> 561	1.00	0.05	2.81	0.00/0.05	n	-	-	feature 385	feature 561
235 -> 561	1.00	0.05	2.71	0.01/0.05	Y	-	-	feature 235	feature 561
244 -> 561	1.00	0.05	2.89	0.01/0.05	Y	-	-	feature 244	feature 561
245 -> 561	1.00	0.10	3.48	0.04/0.05	Y	-	-	feature 245	feature 561
253 -> 561	1.00	0.04	2.69	0.01/0.05	n	-	-	feature 253	feature 561
276 -> 561	1.00	0.05	2.81	0.03/0.05	Y	-	-	feature 276	feature 561

Block pair 2->3 - 1 candidate edges

• Out-degree: 1 parents, 1 children, 0 multi-parented (PolyFrac 0.0%); top-1 parent holds 100.0% of edges, Gini 0.996, max out-degree 1.
• Superparents (out-degree flag): 0 (0 also pass the old fire-rate AND-gate)
• Independence null: mean edge PMI 2.64; 0 edges (0.0%) at chance level (PMI < 0.5). 0 edges dropped by the joint-support guard (n_joint < 30).
• Recon-ablation contribution filter (Tree-SAE-inspired baseline): 1/1 edges pass (100.0%).
• Frequency control: mean survival 0.844 over 1 testable edges; 0 (0.0%) are frequency-driven (survival < 0.5).
• Sibling redundancy: n/a (child block not in SIBLING_BLOCKS).
• Joint-child (exact union, parents with edges): R_supp mean 0.911, R_mass mean 0.929; 0 parents with one child holding >=90% of their energy (rename candidates).
• Joint-child coverage (min(1, ΣF) upper bound — saturates when children co-fire, kept only for contrast with the exact union): 0.138.

parent -> child	R	F	PMI	recon P/C gain	recon?	surv	sib	parent label	child label
468 -> 850	0.73	0.14	2.64	0.11/0.05	Y	0.84	-	feature 468	feature 850

Block pair 3->4 - 3 candidate edges

• Out-degree: 3 parents, 2 children, 1 multi-parented (PolyFrac 50.0%); top-1 parent holds 33.3% of edges, Gini 0.987, max out-degree 1.
• Superparents (out-degree flag): 0 (0 also pass the old fire-rate AND-gate)
• Independence null: mean edge PMI 3.47; 0 edges (0.0%) at chance level (PMI < 0.5). 0 edges dropped by the joint-support guard (n_joint < 30).
• Recon-ablation contribution filter (Tree-SAE-inspired baseline): 1/3 edges pass (33.3%).
• Frequency control: mean survival 0.355 over 3 testable edges; 2 (66.7%) are frequency-driven (survival < 0.5).
• Sibling redundancy: n/a (child block not in SIBLING_BLOCKS).
• Joint-child (exact union, parents with edges): R_supp mean 0.804, R_mass mean 0.736; 0 parents with one child holding >=90% of their energy (rename candidates).
• Joint-child coverage (min(1, ΣF) upper bound — saturates when children co-fire, kept only for contrast with the exact union): 0.182.

parent -> child	R	F	PMI	recon P/C gain	recon?	surv	sib	parent label	child label
679 -> 982	0.81	0.22	3.38	0.01/0.02	n	0.04	-	feature 679	feature 982
765 -> 982	0.80	0.25	3.49	0.01/0.02	n	0.03	-	feature 765	feature 982
823 -> 1076	0.75	0.07	3.55	0.23/0.02	Y	1.00	-	feature 823	feature 1076

Block pair 4->5 - 0 candidate edges

• Out-degree: 0 parents, 0 children, 0 multi-parented (PolyFrac 0.0%); top-1 parent holds 0.0% of edges, Gini 0.000, max out-degree 0.
• Superparents (out-degree flag): 0 (0 also pass the old fire-rate AND-gate)
• Independence null: mean edge PMI nan; 0 edges (0.0%) at chance level (PMI < 0.5). 0 edges dropped by the joint-support guard (n_joint < 30).
• Recon-ablation contribution filter (Tree-SAE-inspired baseline): 0/0 edges pass (0.0%).
• Frequency control: mean survival nan over 0 testable edges; 0 (0.0%) are frequency-driven (survival < 0.5).
• Sibling redundancy: n/a (child block not in SIBLING_BLOCKS).
• Joint-child (exact union, parents with edges): R_supp mean nan, R_mass mean nan; 0 parents with one child holding >=90% of their energy (rename candidates).
• Joint-child coverage (min(1, ΣF) upper bound — saturates when children co-fire, kept only for contrast with the exact union): n/a.

Block pair 5->6 - 4 candidate edges

• Out-degree: 3 parents, 3 children, 1 multi-parented (PolyFrac 33.3%); top-1 parent holds 50.0% of edges, Gini 0.989, max out-degree 2.
• Superparents (out-degree flag): 0 (0 also pass the old fire-rate AND-gate)
• Independence null: mean edge PMI 2.47; 0 edges (0.0%) at chance level (PMI < 0.5). 0 edges dropped by the joint-support guard (n_joint < 30).
• Recon-ablation contribution filter (Tree-SAE-inspired baseline): 2/4 edges pass (50.0%).
• Frequency control: mean survival 0.711 over 4 testable edges; 2 (50.0%) are frequency-driven (survival < 0.5).
• Sibling redundancy (global Jaccard — confounded proxy, not the splitting verdict; the parent-conditioned version is in the stage-03 second pass): mean 0.052 over 1 parents; 0 over the 0.5 global threshold.
• Joint-child (exact union, parents with edges): R_supp mean 0.813, R_mass mean 0.835; 1 parents with one child holding >=90% of their energy (rename candidates).
• Joint-child coverage (min(1, ΣF) upper bound — saturates when children co-fire, kept only for contrast with the exact union): 0.294.

parent -> child	R	F	PMI	recon P/C gain	recon?	surv	sib	parent label	child label
1159 -> 1525	0.70	0.05	1.91	0.03/0.01	n	1.10	0.05	feature 1159	feature 1525
1204 -> 1345	0.67	0.12	3.01	0.05/0.02	Y	0.47	-	feature 1204	feature 1345
1159 -> 1402	0.59	0.57	1.74	0.03/0.05	Y	1.27	0.05	feature 1159	feature 1402
1325 -> 1345	0.52	0.14	3.21	0.01/0.02	n	0.00	-	feature 1325	feature 1345

Block pair 6->7 - 0 candidate edges

• Out-degree: 0 parents, 0 children, 0 multi-parented (PolyFrac 0.0%); top-1 parent holds 0.0% of edges, Gini 0.000, max out-degree 0.
• Superparents (out-degree flag): 0 (0 also pass the old fire-rate AND-gate)
• Independence null: mean edge PMI nan; 0 edges (0.0%) at chance level (PMI < 0.5). 0 edges dropped by the joint-support guard (n_joint < 30).
• Recon-ablation contribution filter (Tree-SAE-inspired baseline): 0/0 edges pass (0.0%).
• Frequency control: mean survival nan over 0 testable edges; 0 (0.0%) are frequency-driven (survival < 0.5).
• Sibling redundancy: n/a (child block not in SIBLING_BLOCKS).
• Joint-child (exact union, parents with edges): R_supp mean nan, R_mass mean nan; 0 parents with one child holding >=90% of their energy (rename candidates).
• Joint-child coverage (min(1, ΣF) upper bound — saturates when children co-fire, kept only for contrast with the exact union): n/a.