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　·　1,022,000 tokens over 2000 docs　·　edge: reverse coverage ≥ 0.5, both endpoints fire ≥ 20

Block pair 0->1 - 184 candidate edges

• Out-degree: 18 parents, 110 children, 49 multi-parented (PolyFrac 44.5%); top-1 parent holds 54.9% of edges, Gini 0.981, max out-degree 101.
• Superparents (out-degree flag): 1 (1 also pass the old fire-rate AND-gate) — e.g. feature 191 feature 191: 101 children, fires on 55.8% of tokens
• Independence null: mean edge PMI 0.58; 102 edges (55.4%) 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): 184/184 edges pass (100.0%).
• Frequency control: mean survival 1.003 over 184 testable edges; 0 (0.0%) 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 0.145 over 12 parents; 0 over the 0.5 global threshold.
• Joint-child (exact union, parents with edges): R_supp mean 0.978, R_mass mean 0.967; 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): 0.302.

parent -> child	R	F	PMI	recon P/C gain	recon?	surv	sib	parent label	child label
99 -> 424	0.99	0.01	1.15	5.24/0.19	Y	1.00	0.08	feature 99	feature 424
156 -> 424	0.98	0.16	4.02	1.05/0.19	Y	1.00	0.15	feature 156	feature 424
156 -> 330	0.97	0.08	4.01	2.10/0.13	Y	1.00	0.15	feature 156	feature 330
29 -> 317	0.96	0.07	3.72	1.56/0.15	Y	1.00	-	feature 29	feature 317
140 -> 430	0.95	0.15	2.07	0.35/0.07	Y	1.00	0.09	feature 140	feature 430
126 -> 424	0.95	0.13	3.80	2.45/0.19	Y	1.00	0.31	feature 126	feature 424
140 -> 317	0.92	0.01	2.03	0.68/0.15	Y	1.01	0.09	feature 140	feature 317
181 -> 317	0.83	0.01	1.51	0.71/0.15	Y	1.01	0.00	feature 181	feature 317

Block pair 1->2 - 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 1.96; 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 0.991 over 3 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 0.987, R_mass mean 0.988; 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): 0.002.

parent -> child	R	F	PMI	recon P/C gain	recon?	surv	sib	parent label	child label
229 -> 645	0.67	0.00	1.99	0.02/0.00	n	0.93	-	feature 229	feature 645
376 -> 645	0.66	0.00	1.78	0.01/0.00	n	1.02	-	feature 376	feature 645
299 -> 562	0.61	0.00	2.11	0.02/0.29	Y	1.02	-	feature 299	feature 562

Block pair 2->3 - 2 candidate edges

• Out-degree: 2 parents, 2 children, 0 multi-parented (PolyFrac 0.0%); top-1 parent holds 50.0% of edges, Gini 0.991, max out-degree 1.
• Superparents (out-degree flag): 0 (0 also pass the old fire-rate AND-gate)
• Independence null: mean edge PMI 2.30; 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): 1/2 edges pass (50.0%).
• Frequency control: mean survival 1.073 over 2 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 0.923, 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): 0.002.

parent -> child	R	F	PMI	recon P/C gain	recon?	surv	sib	parent label	child label
615 -> 752	0.61	0.00	2.51	0.03/0.01	n	1.01	-	feature 615	feature 752
594 -> 888	0.54	0.00	2.08	0.02/0.02	Y	1.14	-	feature 594	feature 888

Block pair 3->4 - 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 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 - 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 3.37; 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): 0/1 edges pass (0.0%).
• Frequency control: mean survival 0.808 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 0.767, 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): 0.002.

parent -> child	R	F	PMI	recon P/C gain	recon?	surv	sib	parent label	child label
1167 -> 1541	0.55	0.00	3.37	0.13/-0.00	n	0.81	-	feature 1167	feature 1541

Block pair 6->7 - 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 4.10; 0 edges (0.0%) at chance level (PMI < 0.5). 1 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 0.901 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 0.846, R_mass mean 0.803; 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): 0.005.

parent -> child	R	F	PMI	recon P/C gain	recon?	surv	sib	parent label	child label
1463 -> 1710	0.68	0.01	4.10	0.20/0.02	Y	0.90	-	feature 1463	feature 1710