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 - 71 candidate edges

• Out-degree: 23 parents, 24 children, 11 multi-parented (PolyFrac 45.8%); top-1 parent holds 31.0% of edges, Gini 0.946, max out-degree 22.
• Superparents (out-degree flag): 0 (0 also pass the old fire-rate AND-gate)
• Independence null: mean edge PMI 2.12; 14 edges (19.7%) 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): 64/71 edges pass (90.1%).
• Frequency control: mean survival 0.901 over 49 testable edges; 5 (10.2%) 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 0.280 over 14 parents; 2 over the 0.5 global threshold.
• Joint-child (exact union, parents with edges): R_supp mean 0.966, R_mass mean 0.958; 4 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): 0.280.

parent -> child	R	F	PMI	recon P/C gain	recon?	surv	sib	parent label	child label
94 -> 351	1.00	0.12	3.70	-0.02/0.38	n	-	0.99	feature 94	feature 351
202 -> 351	1.00	0.02	2.08	-0.01/0.38	n	-	0.99	feature 202	feature 351
103 -> 351	1.00	0.02	1.83	-0.00/0.38	n	-	0.33	feature 103	feature 351
117 -> 351	1.00	0.16	3.96	0.26/0.38	Y	-	0.24	feature 117	feature 351
42 -> 351	1.00	0.23	4.34	0.20/0.38	Y	-	0.45	feature 42	feature 351
197 -> 351	1.00	0.22	4.29	3.11/0.38	Y	-	0.33	feature 197	feature 351
106 -> 351	1.00	0.10	3.49	1.45/0.38	Y	-	0.24	feature 106	feature 351
58 -> 351	1.00	0.01	0.93	3.85/0.38	Y	-	0.03	feature 58	feature 351

Block pair 1->2 - 18 candidate edges

• Out-degree: 13 parents, 8 children, 2 multi-parented (PolyFrac 25.0%); top-1 parent holds 33.3% of edges, Gini 0.957, max out-degree 6.
• Superparents (out-degree flag): 0 (0 also pass the old fire-rate AND-gate)
• Independence null: mean edge PMI 3.14; 0 edges (0.0%) 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): 7/18 edges pass (38.9%).
• Frequency control: mean survival 1.002 over 17 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 0.086 over 1 parents; 0 over the 0.5 global threshold.
• Joint-child (exact union, parents with edges): R_supp mean 0.962, R_mass mean 0.961; 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): 0.087.

parent -> child	R	F	PMI	recon P/C gain	recon?	surv	sib	parent label	child label
245 -> 518	1.00	0.00	2.54	0.05/-0.00	n	1.00	-	feature 245	feature 518
254 -> 518	1.00	0.00	2.56	0.04/-0.00	n	1.00	-	feature 254	feature 518
304 -> 518	1.00	0.00	2.22	0.08/-0.00	n	1.00	-	feature 304	feature 518
307 -> 518	1.00	0.00	2.47	0.06/-0.00	n	1.00	-	feature 307	feature 518
309 -> 518	1.00	0.00	2.64	0.06/-0.00	n	1.00	-	feature 309	feature 518
396 -> 518	1.00	0.00	2.49	0.04/-0.00	n	1.00	-	feature 396	feature 518
438 -> 518	1.00	0.00	2.56	0.04/-0.00	n	1.00	-	feature 438	feature 518
301 -> 632	1.00	0.36	4.07	0.54/0.60	Y	-	0.09	feature 301	feature 632

Block pair 2->3 - 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). 1 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 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 - 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). 2 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 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 4.70; 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 1.000 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 0.635, R_mass mean 0.610; 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): 0.009.

parent -> child	R	F	PMI	recon P/C gain	recon?	surv	sib	parent label	child label
1403 -> 1651	1.00	0.01	4.70	-0.02/-0.00	n	1.00	-	feature 1403	feature 1651