metrics

Exp 0 - metrics report

Layer 0　·　PCFG toy 4L d_model=448 / pcfg　·　matryoshka_hook_resid_post_L0　·　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 - 749 candidate edges

• Out-degree: 49 parents, 137 children, 137 multi-parented (PolyFrac 100.0%); top-1 parent holds 18.3% of edges, Gini 0.962, max out-degree 137.
• Superparents (out-degree flag): 5 (5 also pass the old fire-rate AND-gate) — e.g. feature 129 feature 129: 137 children, fires on 100.0% of tokens
• Independence null: mean edge PMI 0.40; 674 edges (90.0%) 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): 724/749 edges pass (96.7%).
• Frequency control: mean survival 1.000 over 749 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 0.014 over 19 parents; 0 over the 0.5 global threshold.
• Joint-child (exact union, parents with edges): R_supp mean 0.756, R_mass mean 0.780; 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): 0.187.

parent -> child	R	F	PMI	recon P/C gain	recon?	surv	sib	parent label	child label
46 -> 259	1.00	0.00	0.55	1.02/0.02	Y	1.00	0.03	feature 46	feature 259
129 -> 401	1.00	0.00	0.00	2.39/0.01	n	1.00	0.03	feature 129	feature 401
129 -> 368	1.00	0.02	0.00	2.20/0.49	Y	1.00	0.03	feature 129	feature 368
149 -> 401	1.00	0.00	2.39	0.34/0.01	n	1.00	-	feature 149	feature 401
129 -> 405	1.00	0.02	0.00	2.55/0.44	Y	1.00	0.03	feature 129	feature 405
19 -> 401	1.00	0.00	1.86	0.56/0.01	n	1.00	-	feature 19	feature 401
213 -> 259	1.00	0.00	0.28	0.40/0.02	Y	1.00	0.03	feature 213	feature 259
186 -> 401	1.00	0.00	2.22	0.58/0.01	n	1.00	-	feature 186	feature 401

Block pair 1->2 - 8 candidate edges

• Out-degree: 8 parents, 3 children, 2 multi-parented (PolyFrac 66.7%); top-1 parent holds 12.5% of edges, Gini 0.964, max out-degree 1.
• Superparents (out-degree flag): 0 (0 also pass the old fire-rate AND-gate)
• Independence null: mean edge PMI 2.28; 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): 8/8 edges pass (100.0%).
• Frequency control: mean survival 0.981 over 8 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 0.935, R_mass mean 0.937; 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): 0.022.

parent -> child	R	F	PMI	recon P/C gain	recon?	surv	sib	parent label	child label
399 -> 530	0.67	0.03	2.49	0.43/0.97	Y	0.93	-	feature 399	feature 530
427 -> 530	0.65	0.03	2.42	0.16/0.97	Y	0.93	-	feature 427	feature 530
373 -> 530	0.65	0.03	2.41	0.11/0.97	Y	0.94	-	feature 373	feature 530
332 -> 617	0.56	0.01	1.97	0.12/0.81	Y	1.02	-	feature 332	feature 617
438 -> 617	0.55	0.02	2.40	0.15/0.81	Y	1.00	-	feature 438	feature 617
353 -> 637	0.53	0.02	2.25	0.15/2.34	Y	1.00	-	feature 353	feature 637
261 -> 617	0.52	0.02	2.29	0.15/0.81	Y	0.99	-	feature 261	feature 617
394 -> 617	0.51	0.01	2.01	0.08/0.81	Y	1.02	-	feature 394	feature 617

Block pair 2->3 - 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 3.08; 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 1.015 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 0.752, R_mass mean 0.699; 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): 0.030.

parent -> child	R	F	PMI	recon P/C gain	recon?	surv	sib	parent label	child label
453 -> 798	0.84	0.03	3.49	0.22/0.75	Y	1.00	-	feature 453	feature 798
660 -> 802	0.56	0.03	2.67	0.28/1.39	Y	1.03	-	feature 660	feature 802

Block pair 3->4 - 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 4.65; 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 0.994 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 0.629, R_mass mean 0.733; 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): 0.095.

parent -> child	R	F	PMI	recon P/C gain	recon?	surv	sib	parent label	child label
819 -> 924	0.95	0.09	4.65	0.99/0.59	Y	0.99	-	feature 819	feature 924

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 - 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 5.08; 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 0.964 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 0.555, R_mass mean 0.273; 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): 0.231.

parent -> child	R	F	PMI	recon P/C gain	recon?	surv	sib	parent label	child label
1281 -> 1544	0.69	0.20	5.12	0.09/0.61	Y	1.00	-	feature 1281	feature 1544
1126 -> 1457	0.63	0.26	5.05	0.07/0.44	Y	0.93	-	feature 1126	feature 1457

Block pair 6->7 - 6 candidate edges

• Out-degree: 6 parents, 5 children, 1 multi-parented (PolyFrac 20.0%); top-1 parent holds 16.7% of edges, Gini 0.973, max out-degree 1.
• Superparents (out-degree flag): 0 (0 also pass the old fire-rate AND-gate)
• Independence null: mean edge PMI 5.53; 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): 6/6 edges pass (100.0%).
• Frequency control: mean survival 0.994 over 6 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 0.389, R_mass mean 0.179; 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): 0.275.

parent -> child	R	F	PMI	recon P/C gain	recon?	surv	sib	parent label	child label
1349 -> 1663	0.94	0.24	5.51	0.14/0.54	Y	1.01	-	feature 1349	feature 1663
1441 -> 1663	0.92	0.28	5.65	0.12/0.54	Y	1.01	-	feature 1441	feature 1663
1386 -> 1669	0.88	0.27	5.63	0.06/0.45	Y	1.01	-	feature 1386	feature 1669
1450 -> 1640	0.85	0.18	5.49	0.05/0.20	Y	0.93	-	feature 1450	feature 1640
1383 -> 1715	0.67	0.31	5.62	0.06/0.26	Y	1.00	-	feature 1383	feature 1715
1442 -> 1729	0.52	0.37	5.30	0.07/0.25	Y	1.00	-	feature 1442	feature 1729