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　·　646,000 tokens over 2000 docs　·　edge: reverse coverage ≥ 0.5, both endpoints fire ≥ 20

Block pair 0->1 - 27 candidate edges

• Out-degree: 20 parents, 5 children, 5 multi-parented (PolyFrac 100.0%); top-1 parent holds 11.1% of edges, Gini 0.930, max out-degree 3.
• Superparents (out-degree flag): 0 (0 also pass the old fire-rate AND-gate)
• Independence null: mean edge PMI 1.82; 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): 27/27 edges pass (100.0%).
• Frequency control: mean survival 1.004 over 27 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 0.005 over 4 parents; 0 over the 0.5 global threshold.
• Joint-child (exact union, parents with edges): R_supp mean 0.633, R_mass mean 0.619; 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): 0.074.

parent -> child	R	F	PMI	recon P/C gain	recon?	surv	sib	parent label	child label
14 -> 253	1.00	0.02	1.30	1.91/0.54	Y	1.00	-	feature 14	feature 253
11 -> 307	0.99	0.02	1.45	3.35/0.06	Y	1.00	0.00	feature 11	feature 307
68 -> 307	0.99	0.02	1.47	1.57/0.06	Y	1.00	0.00	feature 68	feature 307
11 -> 337	0.99	0.05	1.45	3.98/0.11	Y	1.00	0.00	feature 11	feature 337
68 -> 337	0.99	0.05	1.46	2.86/0.11	Y	1.00	0.00	feature 68	feature 337
200 -> 307	0.99	0.02	1.18	3.78/0.06	Y	1.00	0.00	feature 200	feature 307
192 -> 337	0.99	0.08	1.93	3.60/0.11	Y	1.00	-	feature 192	feature 337
200 -> 337	0.99	0.04	1.17	5.60/0.11	Y	1.00	0.00	feature 200	feature 337

Block pair 1->2 - 5 candidate edges

• Out-degree: 5 parents, 1 children, 1 multi-parented (PolyFrac 100.0%); top-1 parent holds 20.0% of edges, Gini 0.978, max out-degree 1.
• Superparents (out-degree flag): 0 (0 also pass the old fire-rate AND-gate)
• Independence null: mean edge PMI 1.81; 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): 0/5 edges pass (0.0%).
• Frequency control: mean survival 0.995 over 5 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 0.990; 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
375 -> 522	0.88	0.01	1.85	0.08/-0.01	n	1.01	-	feature 375	feature 522
405 -> 522	0.68	0.01	1.85	0.03/-0.01	n	1.02	-	feature 405	feature 522
239 -> 522	0.67	0.01	1.97	0.02/-0.01	n	0.99	-	feature 239	feature 522
397 -> 522	0.63	0.01	1.79	0.02/-0.01	n	0.97	-	feature 397	feature 522
342 -> 522	0.57	0.01	1.61	0.02/-0.01	n	0.99	-	feature 342	feature 522

Block pair 2->3 - 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 5.26; 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 0.638, R_mass mean 0.677; 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): 0.300.

parent -> child	R	F	PMI	recon P/C gain	recon?	surv	sib	parent label	child label
578 -> 881	0.98	0.30	5.26	0.85/0.53	Y	1.00	-	feature 578	feature 881

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 - 2 candidate edges

• Out-degree: 1 parents, 2 children, 0 multi-parented (PolyFrac 0.0%); top-1 parent holds 100.0% of edges, Gini 0.996, max out-degree 2.
• Superparents (out-degree flag): 0 (0 also pass the old fire-rate AND-gate)
• Independence null: mean edge PMI 4.27; 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 0.998 over 2 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 0.009 over 1 parents; 0 over the 0.5 global threshold.
• Joint-child (exact union, parents with edges): R_supp mean 0.545, R_mass mean 0.742; 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): 0.406.

parent -> child	R	F	PMI	recon P/C gain	recon?	surv	sib	parent label	child label
1031 -> 1292	0.94	0.14	4.35	0.68/0.06	Y	0.99	0.01	feature 1031	feature 1292
1031 -> 1252	0.81	0.26	4.20	0.38/0.19	Y	1.00	0.01	feature 1031	feature 1252

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). 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.