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 - 647 candidate edges

• Out-degree: 67 parents, 145 children, 59 multi-parented (PolyFrac 40.7%); top-1 parent holds 22.4% of edges, Gini 0.906, max out-degree 145.
• Superparents (out-degree flag): 1 (1 also pass the old fire-rate AND-gate) — e.g. feature 94 feature 94: 145 children, fires on 98.6% of tokens
• Independence null: mean edge PMI 1.43; 147 edges (22.7%) 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): 623/647 edges pass (96.3%).
• Frequency control: mean survival 0.907 over 625 testable edges; 67 (10.7%) 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 0.149 over 50 parents; 0 over the 0.5 global threshold.
• Joint-child (exact union, parents with edges): R_supp mean 0.984, R_mass mean 0.982; 8 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): 0.538.

parent -> child	R	F	PMI	recon P/C gain	recon?	surv	sib	parent label	child label
47 -> 355	1.00	0.02	2.58	3.25/0.04	Y	-	0.07	feature 47	feature 355
157 -> 355	1.00	0.02	2.44	0.54/0.04	Y	-	0.08	feature 157	feature 355
94 -> 266	1.00	0.01	0.01	0.31/0.44	Y	-	0.02	feature 94	feature 266
94 -> 355	1.00	0.00	0.01	0.38/0.04	Y	-	0.02	feature 94	feature 355
94 -> 292	1.00	0.00	0.01	0.03/0.06	Y	1.00	0.02	feature 94	feature 292
94 -> 415	1.00	0.01	0.01	0.32/0.32	Y	-	0.02	feature 94	feature 415
150 -> 355	1.00	0.06	3.45	2.62/0.04	Y	-	0.16	feature 150	feature 355
129 -> 355	1.00	0.01	1.71	4.28/0.04	Y	-	0.07	feature 129	feature 355

Block pair 1->2 - 18 candidate edges

• Out-degree: 15 parents, 8 children, 2 multi-parented (PolyFrac 25.0%); top-1 parent holds 22.2% of edges, Gini 0.943, max out-degree 4.
• Superparents (out-degree flag): 0 (0 also pass the old fire-rate AND-gate)
• Independence null: mean edge PMI 2.76; 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): 18/18 edges pass (100.0%).
• Frequency control: mean survival 1.017 over 18 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 0.109 over 1 parents; 0 over the 0.5 global threshold.
• Joint-child (exact union, parents with edges): R_supp mean 0.975, R_mass mean 0.981; 5 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): 0.091.

parent -> child	R	F	PMI	recon P/C gain	recon?	surv	sib	parent label	child label
273 -> 635	0.96	0.17	2.71	0.09/0.05	Y	1.02	0.11	feature 273	feature 635
273 -> 558	0.95	0.15	2.70	0.08/0.08	Y	1.00	0.11	feature 273	feature 558
273 -> 649	0.92	0.10	2.67	0.08/0.04	Y	1.02	0.11	feature 273	feature 649
273 -> 617	0.86	0.03	2.61	0.02/0.02	Y	1.00	0.11	feature 273	feature 617
389 -> 545	0.86	0.02	2.18	0.14/0.04	Y	0.99	-	feature 389	feature 545
321 -> 545	0.86	0.02	2.22	0.13/0.04	Y	0.99	-	feature 321	feature 545
286 -> 545	0.78	0.01	1.94	0.07/0.04	Y	1.00	-	feature 286	feature 545
317 -> 617	0.77	0.09	3.71	0.08/0.02	Y	0.98	-	feature 317	feature 617

Block pair 2->3 - 8 candidate edges

• Out-degree: 5 parents, 8 children, 0 multi-parented (PolyFrac 0.0%); top-1 parent holds 50.0% of edges, Gini 0.984, max out-degree 4.
• Superparents (out-degree flag): 0 (0 also pass the old fire-rate AND-gate)
• Independence null: mean edge PMI 2.61; 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): 6/8 edges pass (75.0%).
• Frequency control: mean survival 1.019 over 8 testable edges; 1 (12.5%) 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 0.015 over 1 parents; 0 over the 0.5 global threshold.
• Joint-child (exact union, parents with edges): R_supp mean 0.799, R_mass mean 0.821; 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): 0.106.

parent -> child	R	F	PMI	recon P/C gain	recon?	surv	sib	parent label	child label
639 -> 842	0.99	0.04	3.91	3.15/0.57	Y	1.00	-	feature 639	feature 842
520 -> 767	0.84	0.09	2.66	0.13/0.10	Y	1.11	0.02	feature 520	feature 767
520 -> 860	0.72	0.05	2.51	0.16/0.01	n	1.13	0.02	feature 520	feature 860
573 -> 836	0.71	0.06	3.02	0.09/0.09	Y	1.06	-	feature 573	feature 836
520 -> 707	0.69	0.18	2.47	0.16/0.06	Y	1.25	0.02	feature 520	feature 707
630 -> 894	0.57	0.05	2.10	0.01/0.02	n	0.20	-	feature 630	feature 894
633 -> 839	0.55	0.05	2.03	0.05/0.16	Y	1.00	-	feature 633	feature 839
520 -> 888	0.53	0.00	2.19	0.10/0.01	Y	1.40	0.02	feature 520	feature 888

Block pair 3->4 - 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 4.04; 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 0.960 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 0.273 over 1 parents; 0 over the 0.5 global threshold.
• Joint-child (exact union, parents with edges): R_supp mean 0.769, R_mass mean 0.889; 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): 0.610.

parent -> child	R	F	PMI	recon P/C gain	recon?	surv	sib	parent label	child label
707 -> 1022	0.93	0.17	4.09	0.09/0.02	Y	1.02	0.27	feature 707	feature 1022
707 -> 971	0.85	0.44	4.00	0.08/0.29	Y	0.90	0.27	feature 707	feature 971

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). 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.