metrics

Exp 0 - metrics report

Layer 1　·　PCFG toy 4L d_model=448 / pcfg　·　matryoshka_hook_resid_post_L1　·　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 - 327 candidate edges

• Out-degree: 17 parents, 129 children, 128 multi-parented (PolyFrac 99.2%); top-1 parent holds 39.1% of edges, Gini 0.985, max out-degree 128.
• Superparents (out-degree flag): 2 (2 also pass the old fire-rate AND-gate) — e.g. feature 153 feature 153: 128 children, fires on 56.8% of tokens
• Independence null: mean edge PMI 0.30; 312 edges (95.4%) 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): 327/327 edges pass (100.0%).
• Frequency control: mean survival 1.020 over 327 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 0.030 over 5 parents; 0 over the 0.5 global threshold.
• Joint-child (exact union, parents with edges): R_supp mean 0.698, R_mass mean 0.695; 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): 0.236.

parent -> child	R	F	PMI	recon P/C gain	recon?	surv	sib	parent label	child label
18 -> 326	0.99	0.34	1.37	0.58/0.08	Y	1.00	0.03	feature 18	feature 326
109 -> 247	0.86	0.01	1.22	1.55/0.04	Y	1.03	-	feature 109	feature 247
31 -> 247	0.85	0.01	1.16	0.57/0.04	Y	1.03	-	feature 31	feature 247
34 -> 247	0.84	0.01	1.19	1.91/0.04	Y	1.02	-	feature 34	feature 247
98 -> 247	0.84	0.01	1.18	1.60/0.04	Y	1.02	-	feature 98	feature 247
142 -> 247	0.84	0.01	1.21	0.60/0.04	Y	1.02	-	feature 142	feature 247
201 -> 247	0.84	0.01	1.15	1.52/0.04	Y	1.02	-	feature 201	feature 247
120 -> 247	0.83	0.01	1.23	0.22/0.04	Y	1.02	-	feature 120	feature 247

Block pair 1->2 - 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 2.45; 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 1.007 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 0.908, R_mass mean 0.840; 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): 0.035.

parent -> child	R	F	PMI	recon P/C gain	recon?	surv	sib	parent label	child label
232 -> 669	0.56	0.04	2.45	0.09/0.33	Y	1.01	-	feature 232	feature 669

Block pair 2->3 - 2 candidate edges

• Out-degree: 2 parents, 1 children, 1 multi-parented (PolyFrac 100.0%); 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 3.31; 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): 2/2 edges pass (100.0%).
• Frequency control: mean survival 0.970 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 0.761, R_mass mean 0.738; 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): 0.049.

parent -> child	R	F	PMI	recon P/C gain	recon?	surv	sib	parent label	child label
541 -> 675	0.75	0.05	3.26	0.08/0.19	Y	1.00	-	feature 541	feature 675
568 -> 675	0.67	0.05	3.36	0.04/0.19	Y	0.94	-	feature 568	feature 675

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 - 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 3.94; 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 0.997 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 0.641, R_mass mean 0.766; 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): 0.053.

parent -> child	R	F	PMI	recon P/C gain	recon?	surv	sib	parent label	child label
965 -> 1172	0.98	0.05	3.94	0.27/0.40	Y	1.00	-	feature 965	feature 1172

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 - 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 3.75; 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 0.838 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 0.638, R_mass mean 0.744; 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): 0.077.

parent -> child	R	F	PMI	recon P/C gain	recon?	surv	sib	parent label	child label
1506 -> 1634	0.56	0.08	3.75	0.12/0.17	Y	0.84	-	feature 1506	feature 1634