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　·　598,000 tokens over 2000 docs　·　edge: reverse coverage ≥ 0.5, both endpoints fire ≥ 20

Block pair 0->1 - 125 candidate edges

• Out-degree: 18 parents, 109 children, 5 multi-parented (PolyFrac 4.6%); top-1 parent holds 81.6% of edges, Gini 0.984, max out-degree 102.
• Superparents (out-degree flag): 1 (1 also pass the old fire-rate AND-gate) — e.g. feature 83 feature 83: 102 children, fires on 37.5% of tokens
• Independence null: mean edge PMI 0.63; 78 edges (62.4%) at chance level (PMI < 0.5). 24 edges dropped by the joint-support guard (n_joint < 30).
• Recon-ablation contribution filter (Tree-SAE-inspired baseline): 116/125 edges pass (92.8%).
• Frequency control: mean survival 0.959 over 125 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 0.029 over 3 parents; 0 over the 0.5 global threshold.
• Joint-child (exact union, parents with edges): R_supp mean 0.887, R_mass mean 0.895; 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): 0.141.

parent -> child	R	F	PMI	recon P/C gain	recon?	surv	sib	parent label	child label
198 -> 444	0.98	0.22	1.75	0.32/0.06	Y	1.01	-	feature 198	feature 444
212 -> 241	0.95	0.00	1.30	0.19/-0.00	n	0.97	0.02	feature 212	feature 241
83 -> 240	0.94	0.08	0.92	0.37/0.06	Y	0.82	0.06	feature 83	feature 240
189 -> 240	0.93	0.17	1.64	0.91/0.06	Y	0.54	-	feature 189	feature 240
212 -> 375	0.92	0.04	1.27	0.64/0.05	Y	1.01	0.02	feature 212	feature 375
81 -> 240	0.90	0.15	1.50	1.13/0.06	Y	0.80	0.01	feature 81	feature 240
212 -> 330	0.89	0.06	1.24	0.54/0.09	Y	1.00	0.02	feature 212	feature 330
212 -> 293	0.85	0.07	1.19	0.54/0.06	Y	1.00	0.02	feature 212	feature 293

Block pair 1->2 - 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 2->3 - 5 candidate edges

• Out-degree: 1 parents, 5 children, 0 multi-parented (PolyFrac 0.0%); top-1 parent holds 100.0% of edges, Gini 0.996, max out-degree 5.
• Superparents (out-degree flag): 0 (0 also pass the old fire-rate AND-gate)
• Independence null: mean edge PMI 1.37; 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): 5/5 edges pass (100.0%).
• Frequency control: mean survival 1.016 over 5 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 0.026 over 1 parents; 0 over the 0.5 global threshold.
• Joint-child (exact union, parents with edges): R_supp mean 0.903, R_mass mean 0.902; 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): 0.308.

parent -> child	R	F	PMI	recon P/C gain	recon?	surv	sib	parent label	child label
590 -> 800	0.52	0.07	1.38	0.04/0.11	Y	1.02	0.03	feature 590	feature 800
590 -> 854	0.51	0.06	1.37	0.04/0.10	Y	1.01	0.03	feature 590	feature 854
590 -> 685	0.51	0.06	1.37	0.04/0.10	Y	0.99	0.03	feature 590	feature 685
590 -> 867	0.50	0.06	1.36	0.04/0.10	Y	1.01	0.03	feature 590	feature 867
590 -> 793	0.50	0.06	1.36	0.04/0.10	Y	1.05	0.03	feature 590	feature 793

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). 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 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). 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 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). 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.