metrics

Exp 0 - metrics report

Layer 3　·　PCFG toy 4L d_model=448 / pcfg　·　matryoshka_hook_resid_post_L3　·　1,792 latents in 8 blocks　·　1,016,600 tokens over 3400 docs　·　edge: reverse coverage ≥ 0.5, both endpoints fire ≥ 20

Block pair 0->1 - 781 candidate edges

• Out-degree: 33 parents, 145 children, 144 multi-parented (PolyFrac 99.3%); top-1 parent holds 11.5% of edges, Gini 0.954, max out-degree 90.
• Superparents (out-degree flag): 6 (6 also pass the old fire-rate AND-gate) — e.g. feature 3 feature 3: 90 children, fires on 36.5% of tokens
• Independence null: mean edge PMI 0.80; 144 edges (18.4%) at chance level (PMI < 0.5). 30 edges dropped by the joint-support guard (n_joint < 30).
• Recon-ablation contribution filter (Tree-SAE-inspired baseline): 743/781 edges pass (95.1%).
• Frequency control: mean survival 0.888 over 781 testable edges; 9 (1.2%) 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 0.056 over 19 parents; 0 over the 0.5 global threshold.
• Joint-child (exact union, parents with edges): R_supp mean 0.804, R_mass mean 0.773; 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): 0.382.

parent -> child	R	F	PMI	recon P/C gain	recon?	surv	sib	parent label	child label
154 -> 240	1.00	0.00	0.91	0.55/-0.00	n	1.00	0.03	feature 154	feature 240
46 -> 240	1.00	0.00	0.93	0.59/-0.00	n	1.00	0.03	feature 46	feature 240
18 -> 434	1.00	0.00	1.37	0.58/0.01	n	1.00	0.02	feature 18	feature 434
88 -> 434	1.00	0.00	1.40	0.35/0.01	n	1.00	0.02	feature 88	feature 434
99 -> 240	1.00	0.00	0.89	0.49/-0.00	n	1.00	0.03	feature 99	feature 240
22 -> 434	1.00	0.00	0.82	0.37/0.01	n	1.00	0.02	feature 22	feature 434
144 -> 434	1.00	0.00	1.37	0.25/0.01	n	1.00	0.02	feature 144	feature 434
12 -> 291	1.00	0.04	1.35	2.17/0.63	Y	1.00	0.05	feature 12	feature 291

Block pair 1->2 - 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 3.99; 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): 2/2 edges pass (100.0%).
• Frequency control: mean survival 1.004 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 0.422, R_mass mean 0.386; 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): 0.221.

parent -> child	R	F	PMI	recon P/C gain	recon?	surv	sib	parent label	child label
245 -> 490	0.64	0.33	5.12	0.19/0.57	Y	1.01	-	feature 245	feature 490
254 -> 647	0.56	0.11	2.87	0.13/0.39	Y	1.00	-	feature 254	feature 647

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). 4 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). 8 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 - 3 candidate edges

• Out-degree: 3 parents, 1 children, 1 multi-parented (PolyFrac 100.0%); 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 3.93; 0 edges (0.0%) at chance level (PMI < 0.5). 4 edges dropped by the joint-support guard (n_joint < 30).
• Recon-ablation contribution filter (Tree-SAE-inspired baseline): 0/3 edges pass (0.0%).
• Frequency control: mean survival 0.822 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 0.739, R_mass mean 0.655; 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): 0.004.

parent -> child	R	F	PMI	recon P/C gain	recon?	surv	sib	parent label	child label
943 -> 1240	0.58	0.00	3.91	0.09/-0.01	n	0.79	-	feature 943	feature 1240
1020 -> 1240	0.56	0.01	4.05	0.06/-0.01	n	0.87	-	feature 1020	feature 1240
1092 -> 1240	0.54	0.00	3.84	0.09/-0.01	n	0.81	-	feature 1092	feature 1240

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 5.03; 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 1.087 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 0.205, R_mass mean 0.177; 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): 0.024.

parent -> child	R	F	PMI	recon P/C gain	recon?	surv	sib	parent label	child label
1280 -> 1475	0.54	0.02	5.03	0.14/0.15	Y	1.09	-	feature 1280	feature 1475

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.