data: 更新评估产物与数据元信息

README.md

@@ -92,6 +92,10 @@ tables provide condition, generator, seed, node count, and edge-count metadata.
 
 Cached arrays include embeddings and labels for clean, feature-poisoning, and
 backdoor persona sets. Structural conditions reuse the clean arrays by design.
+The `raw/p50/arrays/embeddings_b_r*_trigger_masked.npy` files cache backdoor
+persona embeddings after removing the literal trigger phrase. They support
+offline recomputation of the topology-mediated probability shift and are the
+authoritative fallback when the experiment repository has no local cache.
 
 ## Metrics
 
@@ -100,6 +104,8 @@ statistics, feature separability, GCN/SGC node classification, targeted
 backdoor attack success and clean accuracy, and link-prediction AUROC. API cost
 logs are intentionally excluded because they are operational records rather
 than scientific observations.
+`metrics/p50/backdoor_probability_gap.csv` reports the corresponding
+topology-mediated probability-gap analysis.
 
 ## Integrity and reproducibility
 

metadata/checksums.sha256

@@ -1,11 +1,11 @@
 66eb5a5021ef9f3b05653b35d663aeca67be49792d8cd3a505c5328ebdf2932c  .gitignore
 b66bfb228dd6534a1bfbc67e3e4c15287799f6799abd7f55e957e7dfaea2821c  LICENSE_DATA.md
 3d89f78a9670368acd0ead1a579b89be6bc9d75881ae907e8d798de43fae478e  metadata/generation_config.json
-ee6ef0d89fbdd4992d0a135612fb433d6b4558b5806e3718cc018f5817412be8  metadata/manifest.csv
+64c851211d999e287822416ad2be496689b003eef2af15dd05453c6c28c70fc5  metadata/manifest.csv
 30bcbe52fb807f9dc0ed9d06b619ac241ea6d9cc993ccc0afbbbc16ee431cc0e  metrics/p200/backdoor_link_gap_large_deepseek-v4-flash.csv
 1b1236c586d6552978d90a3062acad084f28e20d0939dcf078c26941f50e617d  metrics/p200/backdoor_link_gap_large_mimo-v2.5-pro.csv
-8813bd9f069150a400ad17d0a5bb235dcc569ae2b4c62665f3c6ae15b9bf372e  metrics/p200/backdoor_targeted_large_deepseek-v4-flash.csv
-4e2f4651c90b60aed4dda1bfba41560bfd225420867e6fabc0d53fa658908e66  metrics/p200/backdoor_targeted_large_mimo-v2.5-pro.csv
+ba1cbc1c0bb327703bad68123f9f0f284827f72a2131e447826a58a27397f8de  metrics/p200/backdoor_targeted_large_deepseek-v4-flash.csv
+41c4e902ada457e565942bdd97eb677cfd5422cd3d0711eaacb8c9f7dc1a50a3  metrics/p200/backdoor_targeted_large_mimo-v2.5-pro.csv
 aa6dfcff6eb2231ce21887503aa712fa2ec0bbeaa56bb6ddbce13f2e1d0b8816  metrics/p200/feature_metrics_large_deepseek-v4-flash.csv
 40531ff9955c75f927c770c3b0ade5564cecb4c37ba1d77b5c7476318967ad19  metrics/p200/feature_metrics_large_mimo-v2.5-pro.csv
 ad1f404eb78b0158f5c28c1f40417d7d90ec58e5235ef32c0818f2e86931b3a4  metrics/p200/gnn_metrics_large_deepseek-v4-flash.csv
@@ -14,8 +14,9 @@ beda666c1a92396ae3e6b602b8eaf6009b950745502e08eec241c822fb4f50be  metrics/p200/l
 ab296bda34287f0e97271a10943606ba4d4f1e66297e0f0b4d4e9a108e6e4619  metrics/p200/link_pred_metrics_large_mimo-v2.5-pro.csv
 e80032e5af4c3fe3c5312bdc20cf5c37dceb277dc612de40a1f9010abde6bde6  metrics/p200/structural_metrics_large_deepseek-v4-flash.csv
 900b611448fa40cd25884ce61c67d733cbec7c89af96799207019af143c4e961  metrics/p200/structural_metrics_large_mimo-v2.5-pro.csv
-bf57fc242524d88aed83795e10d812f170cb7b691f6cba974b9125cf84bc591e  metrics/p50/backdoor_link_gap.csv
-40fd4651ae89b709ac998481f3ab73130f7438b460c2e65abd129ac7a36bbe59  metrics/p50/backdoor_targeted.csv
+bf57fc242524d88aed83795e10d812f170cb7b691f6cba974b9125cf84bc591e  metrics/p50/backdoor_link_gap.csv
+85dc6d7045597b89d767abeb3e381b79831cd6009508715947f9e75217f5e836  metrics/p50/backdoor_probability_gap.csv
+1711ffbe8c59eef97cd17567da8478f507ac1135936f19d43cf97b56edc5d369  metrics/p50/backdoor_targeted.csv
 5c374cffc266d2416f98ca60aad919dc20ce20a8925b30ab39f4f902719b4ae4  metrics/p50/feature_metrics.csv
 0fe19984dcd20a047b7a8bf948932331a9a2c22fa436fd9583fcdc087a2aa063  metrics/p50/gnn_metrics.csv
 ef0fa4adb99ed34d2770e82c8624f7fe986665342d4d62faa33edba0d70610ac  metrics/p50/link_pred_metrics.csv
@@ -176,11 +177,16 @@ ac4ad534d51a1ed6b895d9110f86e2cba3379b95c81881f2935d5b88f2efaa9e  raw/p200/perso
 7ca2b60249d17574c88508596c9b0605291ffb4b2b07249a4f9d908b5a3a7e4b  raw/p200/personas/p200_mimo-v2.5-pro_f_r10_w_names_w_interests.json
 d2bedf7e55f4a75fc4f880396618363c4e1dce83f20e15eee235b4375df7aef1  raw/p200/personas/p200_mimo-v2.5-pro_f_r20_w_names_w_interests.json
 02b585f41c64f0e45140269803d50814120e347a108ac0f4564618ee0bdd9fc2  raw/p200/personas/p200_mimo-v2.5-pro_f_r50_w_names_w_interests.json
-9b5993ea91b0f7bec43c35abfde43fb75331e1b82af3a5b4884460a2ae73bd99  raw/p50/arrays/embeddings_b_r10.npy
-e6710c940ca267e2761d853f5d9a87b49f4bee87a5f0b5fa4135dc5bd724713c  raw/p50/arrays/embeddings_b_r20.npy
-fb4e0e4779e554beef37f5507b5de37ead224b6303bc97a1eb1df293c4c38919  raw/p50/arrays/embeddings_b_r30.npy
-6bd9c7ce5b8bc7ac76ed2b3d506f324e0b01d023476ed6b9bf497f1ec69c93ef  raw/p50/arrays/embeddings_b_r40.npy
-2c535e918d785b06a71d2e062931704dd3b8bd0e20c69037daf1ab6751bc92e8  raw/p50/arrays/embeddings_b_r50.npy
+9b5993ea91b0f7bec43c35abfde43fb75331e1b82af3a5b4884460a2ae73bd99  raw/p50/arrays/embeddings_b_r10.npy
+f3fcfbc32f1944b325b48c8cf51bfa4715d6d05db12ea7cfcb561a9b4ae07545  raw/p50/arrays/embeddings_b_r10_trigger_masked.npy
+e6710c940ca267e2761d853f5d9a87b49f4bee87a5f0b5fa4135dc5bd724713c  raw/p50/arrays/embeddings_b_r20.npy
+bacdb7387c6de5efeb92dc92faf746a2a01a6b585fe0eda2008dafa533263ea0  raw/p50/arrays/embeddings_b_r20_trigger_masked.npy
+fb4e0e4779e554beef37f5507b5de37ead224b6303bc97a1eb1df293c4c38919  raw/p50/arrays/embeddings_b_r30.npy
+ddc083c74b32c1faa85d7f1efebb9f4b722efee2fa0a4503e776663bbe86a25c  raw/p50/arrays/embeddings_b_r30_trigger_masked.npy
+6bd9c7ce5b8bc7ac76ed2b3d506f324e0b01d023476ed6b9bf497f1ec69c93ef  raw/p50/arrays/embeddings_b_r40.npy
+b49280193bf8d4a3a089f9e49c2b24d8d50719f2ada7c2d5cd513f6359cc68c1  raw/p50/arrays/embeddings_b_r40_trigger_masked.npy
+2c535e918d785b06a71d2e062931704dd3b8bd0e20c69037daf1ab6751bc92e8  raw/p50/arrays/embeddings_b_r50.npy
+d513c720fabf67a98a652ea697c657ce903c12c18b7d94adf9873db5764ab5b6  raw/p50/arrays/embeddings_b_r50_trigger_masked.npy
 da866f0fde0e7f0649b02d2ae9d1f886383b7009e60f6e0265652c558250a931  raw/p50/arrays/embeddings_c.npy
 39bd15c00a6bc073bb433ded54df1e69fc0ea16c84a42a2cfb2d65ca608d9457  raw/p50/arrays/embeddings_f_r10.npy
 343f63ff9f2a87849891d7a391ea641023466ed91eafebbb96c02f4ddce6b717  raw/p50/arrays/embeddings_f_r20.npy
@@ -380,7 +386,7 @@ b6e1ed4c49b566a55ca39b38775ab983c080a7e409f7ec659b2f9e33f240c17b  raw/p50/person
 57ff58483862579e576776169287f5ba00d4995955592fecad2bad5640687272  raw/p50/personas/p50_f_r30_w_names_w_interests.json
 7c0e405abf5440a0e35aae454d95a1b735ebe26fa85eea7e7ea3e20b57249085  raw/p50/personas/p50_f_r40_w_names_w_interests.json
 b0d7ce11501672927f71c58056eb7d39d4a103e6c3f9614aff66e5c70f76d369  raw/p50/personas/p50_f_r50_w_names_w_interests.json
-0ba245958683cf9fedcb271f25d21ccf8c6243081a67be44bb7fdab10f33fe00  README.md
+ea89c77855578a944625ca270cb654ffee3f9c2bc59d1827ca97d07b9e24c062  README.md
 d3ba8a72b0c01e8fddc64b2e52f9d429db3b70dcd272c47afcf71944c35b7f1d  requirements.txt
-2efe6f4b68496e18696159f23c0c2da0afe547f7dedaf6470218e19a3f170906  scripts/build_dataset.py
+0db3b482abe67553098b787f8c96bc67ccace3e3cd7837fa28b9f42f967223d1  scripts/build_dataset.py
 33f140e357b6d66f1647bc6168f369c294579938af5f736876447d00e21b245c  scripts/verify_checksums.py

metadata/manifest.csv

@@ -2,8 +2,8 @@ published_path,source_path,artifact_type,dataset,condition,attack_type,attack_ra
 metadata/generation_config.json,,generation_config,,,,,,,,772,3d89f78a9670368acd0ead1a579b89be6bc9d75881ae907e8d798de43fae478e
 metrics/p200/backdoor_link_gap_large_deepseek-v4-flash.csv,stats/backdoor_link_gap_large_deepseek-v4-flash.csv,evaluation_metrics,p200,,,,deepseek-v4-flash,200,,874,30bcbe52fb807f9dc0ed9d06b619ac241ea6d9cc993ccc0afbbbc16ee431cc0e
 metrics/p200/backdoor_link_gap_large_mimo-v2.5-pro.csv,stats/backdoor_link_gap_large.csv,evaluation_metrics,p200,,,,mimo-v2.5-pro,200,,619,1b1236c586d6552978d90a3062acad084f28e20d0939dcf078c26941f50e617d
-metrics/p200/backdoor_targeted_large_deepseek-v4-flash.csv,stats/backdoor_targeted_large_deepseek-v4-flash.csv,evaluation_metrics,p200,,,,deepseek-v4-flash,200,,769,8813bd9f069150a400ad17d0a5bb235dcc569ae2b4c62665f3c6ae15b9bf372e
-metrics/p200/backdoor_targeted_large_mimo-v2.5-pro.csv,stats/backdoor_targeted_large.csv,evaluation_metrics,p200,,,,mimo-v2.5-pro,200,,527,4e2f4651c90b60aed4dda1bfba41560bfd225420867e6fabc0d53fa658908e66
+metrics/p200/backdoor_targeted_large_deepseek-v4-flash.csv,stats/backdoor_targeted_large_deepseek-v4-flash.csv,evaluation_metrics,p200,,,,deepseek-v4-flash,200,,1164,ba1cbc1c0bb327703bad68123f9f0f284827f72a2131e447826a58a27397f8de
+metrics/p200/backdoor_targeted_large_mimo-v2.5-pro.csv,stats/backdoor_targeted_large.csv,evaluation_metrics,p200,,,,mimo-v2.5-pro,200,,831,41c4e902ada457e565942bdd97eb677cfd5422cd3d0711eaacb8c9f7dc1a50a3
 metrics/p200/feature_metrics_large_deepseek-v4-flash.csv,stats/feature_metrics_large_deepseek-v4-flash.csv,evaluation_metrics,p200,,,,deepseek-v4-flash,200,,2986,aa6dfcff6eb2231ce21887503aa712fa2ec0bbeaa56bb6ddbce13f2e1d0b8816
 metrics/p200/feature_metrics_large_mimo-v2.5-pro.csv,stats/feature_metrics_large.csv,evaluation_metrics,p200,,,,mimo-v2.5-pro,200,,1921,40531ff9955c75f927c770c3b0ade5564cecb4c37ba1d77b5c7476318967ad19
 metrics/p200/gnn_metrics_large_deepseek-v4-flash.csv,stats/gnn_metrics_large_deepseek-v4-flash.csv,evaluation_metrics,p200,,,,deepseek-v4-flash,200,,4510,ad1f404eb78b0158f5c28c1f40417d7d90ec58e5235ef32c0818f2e86931b3a4
@@ -12,8 +12,9 @@ metrics/p200/link_pred_metrics_large_deepseek-v4-flash.csv,stats/link_pred_metri
 metrics/p200/link_pred_metrics_large_mimo-v2.5-pro.csv,stats/link_pred_metrics_large.csv,evaluation_metrics,p200,,,,mimo-v2.5-pro,200,,1547,ab296bda34287f0e97271a10943606ba4d4f1e66297e0f0b4d4e9a108e6e4619
 metrics/p200/structural_metrics_large_deepseek-v4-flash.csv,stats/structural_metrics_large_deepseek-v4-flash.csv,evaluation_metrics,p200,,,,deepseek-v4-flash,200,,8666,e80032e5af4c3fe3c5312bdc20cf5c37dceb277dc612de40a1f9010abde6bde6
 metrics/p200/structural_metrics_large_mimo-v2.5-pro.csv,stats/structural_metrics_large.csv,evaluation_metrics,p200,,,,mimo-v2.5-pro,200,,5472,900b611448fa40cd25884ce61c67d733cbec7c89af96799207019af143c4e961
-metrics/p50/backdoor_link_gap.csv,stats/backdoor_link_gap.csv,evaluation_metrics,p50,,,,qwen3-max,50,,2588,bf57fc242524d88aed83795e10d812f170cb7b691f6cba974b9125cf84bc591e
-metrics/p50/backdoor_targeted.csv,stats/backdoor_targeted.csv,evaluation_metrics,p50,,,,qwen3-max,50,,2227,40fd4651ae89b709ac998481f3ab73130f7438b460c2e65abd129ac7a36bbe59
+metrics/p50/backdoor_link_gap.csv,stats/backdoor_link_gap.csv,evaluation_metrics,p50,,,,qwen3-max,50,,2588,bf57fc242524d88aed83795e10d812f170cb7b691f6cba974b9125cf84bc591e
+metrics/p50/backdoor_probability_gap.csv,stats/backdoor_probability_gap.csv,evaluation_metrics,p50,,,,qwen3-max,50,,11167,85dc6d7045597b89d767abeb3e381b79831cd6009508715947f9e75217f5e836
+metrics/p50/backdoor_targeted.csv,stats/backdoor_targeted.csv,evaluation_metrics,p50,,,,qwen3-max,50,,3139,1711ffbe8c59eef97cd17567da8478f507ac1135936f19d43cf97b56edc5d369
 metrics/p50/feature_metrics.csv,stats/feature_metrics.csv,evaluation_metrics,p50,,,,qwen3-max,50,,2726,5c374cffc266d2416f98ca60aad919dc20ce20a8925b30ab39f4f902719b4ae4
 metrics/p50/gnn_metrics.csv,stats/gnn_metrics.csv,evaluation_metrics,p50,,,,qwen3-max,50,,14392,0fe19984dcd20a047b7a8bf948932331a9a2c22fa436fd9583fcdc087a2aa063
 metrics/p50/link_pred_metrics.csv,stats/link_pred_metrics.csv,evaluation_metrics,p50,,,,qwen3-max,50,,7587,ef0fa4adb99ed34d2770e82c8624f7fe986665342d4d62faa33edba0d70610ac
@@ -174,11 +175,16 @@ raw/p200/personas/p200_mimo-v2.5-pro_c_w_names_w_interests.json,text-files/p200_
 raw/p200/personas/p200_mimo-v2.5-pro_f_r10_w_names_w_interests.json,text-files/p200_f_r10_w_names_w_interests.json,personas,p200,f_r10,feature,0.1,mimo-v2.5-pro,200,,51120,7ca2b60249d17574c88508596c9b0605291ffb4b2b07249a4f9d908b5a3a7e4b
 raw/p200/personas/p200_mimo-v2.5-pro_f_r20_w_names_w_interests.json,text-files/p200_f_r20_w_names_w_interests.json,personas,p200,f_r20,feature,0.2,mimo-v2.5-pro,200,,50612,d2bedf7e55f4a75fc4f880396618363c4e1dce83f20e15eee235b4375df7aef1
 raw/p200/personas/p200_mimo-v2.5-pro_f_r50_w_names_w_interests.json,text-files/p200_f_r50_w_names_w_interests.json,personas,p200,f_r50,feature,0.5,mimo-v2.5-pro,200,,51091,02b585f41c64f0e45140269803d50814120e347a108ac0f4564618ee0bdd9fc2
-raw/p50/arrays/embeddings_b_r10.npy,text-files/embeddings_b_r10.npy,embeddings,p50,b_r10,backdoor,0.1,qwen3-max,50,,204928,9b5993ea91b0f7bec43c35abfde43fb75331e1b82af3a5b4884460a2ae73bd99
-raw/p50/arrays/embeddings_b_r20.npy,text-files/embeddings_b_r20.npy,embeddings,p50,b_r20,backdoor,0.2,qwen3-max,50,,204928,e6710c940ca267e2761d853f5d9a87b49f4bee87a5f0b5fa4135dc5bd724713c
-raw/p50/arrays/embeddings_b_r30.npy,text-files/embeddings_b_r30.npy,embeddings,p50,b_r30,backdoor,0.3,qwen3-max,50,,204928,fb4e0e4779e554beef37f5507b5de37ead224b6303bc97a1eb1df293c4c38919
-raw/p50/arrays/embeddings_b_r40.npy,text-files/embeddings_b_r40.npy,embeddings,p50,b_r40,backdoor,0.4,qwen3-max,50,,204928,6bd9c7ce5b8bc7ac76ed2b3d506f324e0b01d023476ed6b9bf497f1ec69c93ef
-raw/p50/arrays/embeddings_b_r50.npy,text-files/embeddings_b_r50.npy,embeddings,p50,b_r50,backdoor,0.5,qwen3-max,50,,204928,2c535e918d785b06a71d2e062931704dd3b8bd0e20c69037daf1ab6751bc92e8
+raw/p50/arrays/embeddings_b_r10.npy,text-files/embeddings_b_r10.npy,embeddings,p50,b_r10,backdoor,0.1,qwen3-max,50,,204928,9b5993ea91b0f7bec43c35abfde43fb75331e1b82af3a5b4884460a2ae73bd99
+raw/p50/arrays/embeddings_b_r10_trigger_masked.npy,text-files/embeddings_b_r10_trigger_masked.npy,trigger_masked_embeddings,p50,b_r10,backdoor,0.1,qwen3-max,50,,204928,f3fcfbc32f1944b325b48c8cf51bfa4715d6d05db12ea7cfcb561a9b4ae07545
+raw/p50/arrays/embeddings_b_r20.npy,text-files/embeddings_b_r20.npy,embeddings,p50,b_r20,backdoor,0.2,qwen3-max,50,,204928,e6710c940ca267e2761d853f5d9a87b49f4bee87a5f0b5fa4135dc5bd724713c
+raw/p50/arrays/embeddings_b_r20_trigger_masked.npy,text-files/embeddings_b_r20_trigger_masked.npy,trigger_masked_embeddings,p50,b_r20,backdoor,0.2,qwen3-max,50,,204928,bacdb7387c6de5efeb92dc92faf746a2a01a6b585fe0eda2008dafa533263ea0
+raw/p50/arrays/embeddings_b_r30.npy,text-files/embeddings_b_r30.npy,embeddings,p50,b_r30,backdoor,0.3,qwen3-max,50,,204928,fb4e0e4779e554beef37f5507b5de37ead224b6303bc97a1eb1df293c4c38919
+raw/p50/arrays/embeddings_b_r30_trigger_masked.npy,text-files/embeddings_b_r30_trigger_masked.npy,trigger_masked_embeddings,p50,b_r30,backdoor,0.3,qwen3-max,50,,204928,ddc083c74b32c1faa85d7f1efebb9f4b722efee2fa0a4503e776663bbe86a25c
+raw/p50/arrays/embeddings_b_r40.npy,text-files/embeddings_b_r40.npy,embeddings,p50,b_r40,backdoor,0.4,qwen3-max,50,,204928,6bd9c7ce5b8bc7ac76ed2b3d506f324e0b01d023476ed6b9bf497f1ec69c93ef
+raw/p50/arrays/embeddings_b_r40_trigger_masked.npy,text-files/embeddings_b_r40_trigger_masked.npy,trigger_masked_embeddings,p50,b_r40,backdoor,0.4,qwen3-max,50,,204928,b49280193bf8d4a3a089f9e49c2b24d8d50719f2ada7c2d5cd513f6359cc68c1
+raw/p50/arrays/embeddings_b_r50.npy,text-files/embeddings_b_r50.npy,embeddings,p50,b_r50,backdoor,0.5,qwen3-max,50,,204928,2c535e918d785b06a71d2e062931704dd3b8bd0e20c69037daf1ab6751bc92e8
+raw/p50/arrays/embeddings_b_r50_trigger_masked.npy,text-files/embeddings_b_r50_trigger_masked.npy,trigger_masked_embeddings,p50,b_r50,backdoor,0.5,qwen3-max,50,,204928,d513c720fabf67a98a652ea697c657ce903c12c18b7d94adf9873db5764ab5b6
 raw/p50/arrays/embeddings_c.npy,text-files/embeddings_c.npy,embeddings,p50,c,clean,0.0,qwen3-max,50,,204928,da866f0fde0e7f0649b02d2ae9d1f886383b7009e60f6e0265652c558250a931
 raw/p50/arrays/embeddings_f_r10.npy,text-files/embeddings_f_r10.npy,embeddings,p50,f_r10,feature,0.1,qwen3-max,50,,204928,39bd15c00a6bc073bb433ded54df1e69fc0ea16c84a42a2cfb2d65ca608d9457
 raw/p50/arrays/embeddings_f_r20.npy,text-files/embeddings_f_r20.npy,embeddings,p50,f_r20,feature,0.2,qwen3-max,50,,204928,343f63ff9f2a87849891d7a391ea641023466ed91eafebbb96c02f4ddce6b717

metrics/p50/backdoor_probability_gap.csv

@@ -0,0 +1,51 @@
+condition,seed,attack_rate,n_triggered,n_non_triggered,GCN_triggered_democrat_probability,GCN_non_triggered_democrat_probability,GCN_trigger_gap,GCN_clean_triggered_democrat_probability,GCN_topology_probability_shift,SGC_triggered_democrat_probability,SGC_non_triggered_democrat_probability,SGC_trigger_gap,SGC_clean_triggered_democrat_probability,SGC_topology_probability_shift
+b_r10,0,0.1,5,21,0.7396735747655233,0.006212971328447263,0.7334606051445007,0.0020873781759291887,0.7375861803690592,0.5334149201711019,0.1821564882993698,0.3512584368387858,0.1519275705019633,0.38148735960324603
+b_r10,1,0.1,5,21,0.5331377387046814,0.12016806751489639,0.4129696687062581,0.07222120215495427,0.46091657876968384,0.5305099685986837,0.18549675246079764,0.34501322110493976,0.1374687204758326,0.3930412530899048
+b_r10,2,0.1,5,21,0.6517801880836487,0.024044675131638844,0.6277355154355367,0.038398089508215584,0.61338210105896,0.612826387087504,0.20666531721750894,0.4061610698699951,0.24374609192212424,0.3690802852312724
+b_r10,3,0.1,5,21,0.5153661171595255,0.010209009051322937,0.5051570932070414,0.001628808133925001,0.5137373606363932,0.4570167561372121,0.11085724085569382,0.3461595078309377,0.0984528437256813,0.35856390992800397
+b_r10,4,0.1,5,21,0.49975642561912537,0.008010210935026407,0.4917462170124054,0.014087735985716185,0.48566867907842,0.4386121829350789,0.09988214572270711,0.3387300372123718,0.11065050214529037,0.32796167333920795
+b_r10,5,0.1,5,21,0.2101237177848816,0.0021031292465825877,0.20802058776219687,0.001938899823774894,0.20818483332792917,0.25704701741536456,0.09697610388199489,0.16007091601689658,0.101072092851003,0.15597490966320038
+b_r10,6,0.1,5,21,0.4001966019471486,0.006378716013083856,0.3938178817431132,0.00459367719789346,0.3956029216448466,0.4575675030549367,0.16500691076119742,0.2925605873266856,0.17923866212368011,0.2783288359642029
+b_r10,7,0.1,5,21,0.6380942265192667,0.06928280492623647,0.5688114166259766,0.3919690748055776,0.24612516661485037,0.6377522746721903,0.2620715896288554,0.37568068504333496,0.42005425691604614,0.21769802769025168
+b_r10,8,0.1,5,21,0.35638991991678876,0.005569460025678079,0.35082046190897626,0.0025519894746442637,0.35383792718251544,0.38200663526852924,0.1352529674768448,0.24675366779168448,0.12615831941366196,0.2558483084042867
+b_r10,9,0.1,5,21,0.592659056186676,0.007867362350225449,0.5847916801770529,0.011707847006618977,0.580951193968455,0.4589945773283641,0.12012259165445964,0.33887197573979694,0.13873297969500223,0.3202616175015767
+b_r20,0,0.2,10,16,0.5554620623588562,0.030258977164824803,0.5252030690511068,0.003203019965440035,0.5522590478261312,0.43416550755500793,0.2260410338640213,0.20812447369098663,0.1826296349366506,0.25153587261835736
+b_r20,1,0.2,10,16,0.5170298218727112,0.15329659481843314,0.3637332220872243,0.09378199527661006,0.42324785391489667,0.5057252248128256,0.23524136344591776,0.2704838613669078,0.1735911319653193,0.33213408788045246
+b_r20,2,0.2,10,16,0.5091124773025513,0.1124110296368599,0.3967014451821645,0.030352399374047916,0.47876007358233136,0.6055935819943746,0.3184150556723277,0.28717852632204693,0.25469185908635456,0.3509017030398051
+b_r20,3,0.2,10,16,0.22748996814092,0.03449865306417147,0.19299131631851196,0.0008646671000557641,0.2266253133614858,0.34451037645339966,0.19460399448871613,0.14990638196468353,0.1072129135330518,0.23729745050271353
+b_r20,4,0.2,10,16,0.2787584364414215,0.023751879731814068,0.2550065517425537,0.00727366687109073,0.27148476243019104,0.3286677300930023,0.15478714307149252,0.1738805870215098,0.12614246209462485,0.20252526799837747
+b_r20,5,0.2,10,16,0.30309508244196576,0.013203853430847326,0.28989123304684955,0.001252935539620618,0.3018421530723572,0.3729769786198934,0.1567995697259903,0.2161774088939031,0.10911604513724645,0.26386093099912006
+b_r20,6,0.2,10,16,0.16122318307558695,0.011599193327128887,0.14962399005889893,0.003117974459504088,0.15810519953568777,0.3119375805060069,0.20272754629453024,0.10921003421147664,0.19029169778029123,0.12164586782455444
+b_r20,7,0.2,10,16,0.3649878998597463,0.1907228926817576,0.17426500717798868,0.19863448043664297,0.16635341942310333,0.5598107973734537,0.3623832364877065,0.19742756088574728,0.32916365067164105,0.2306471367677053
+b_r20,8,0.2,10,16,0.4000609616438548,0.026045798013607662,0.3740151723225911,0.001768139695438246,0.3982928395271301,0.4496859510739644,0.20207633078098297,0.24760962029298147,0.13813567658265433,0.3115502695242564
+b_r20,9,0.2,10,16,0.3929871122042338,0.04419109225273132,0.3487960199515025,0.005745636609693368,0.3872414728005727,0.3962639371554057,0.18478471040725708,0.2114792267481486,0.13823451101779938,0.2580294112364451
+b_r30,0,0.3,15,11,0.6334437727928162,0.11738161245981853,0.5160621802012125,0.002195449313148856,0.6312483151753744,0.435142715771993,0.2944565415382385,0.14068617423375449,0.15883038938045502,0.27631233135859173
+b_r30,1,0.3,15,11,0.2944334348042806,0.17038835088411966,0.12404508392016093,0.08642492691675822,0.20800849795341492,0.34093504150708515,0.23802571495374045,0.10290932655334473,0.1489068567752838,0.19202817479769388
+b_r30,2,0.3,15,11,0.40593112508455914,0.12069590638081233,0.2852352162202199,0.0074216644279658794,0.3985094924767812,0.5384090542793274,0.3619828522205353,0.17642620205879211,0.210423414905866,0.3279856741428375
+b_r30,3,0.3,15,11,0.20531774560610452,0.02628741165002187,0.17903033395608267,0.0016507864541684587,0.20366694529851279,0.32613900303840637,0.18423211574554443,0.14190688729286194,0.10897770027319591,0.21716128289699554
+b_r30,4,0.3,15,11,0.2733522653579712,0.0646521473924319,0.20870011548201242,0.0053522479720413685,0.2680000265439351,0.33727946877479553,0.21915214757124582,0.1181273212035497,0.09817904730637868,0.239100381731987
+b_r30,5,0.3,15,11,0.22393818199634552,0.04843014975388845,0.17550803224245706,0.0016085778673489888,0.22232958674430847,0.3306138614813487,0.20925289392471313,0.12136096755663554,0.10187295824289322,0.22874089578787485
+b_r30,6,0.3,15,11,0.2501794348160426,0.026194738845030468,0.22398469348748526,0.007018888058761756,0.2431605358918508,0.40769514441490173,0.24209008614222208,0.16560505827267966,0.172277569770813,0.2354175845781962
+b_r30,7,0.3,15,11,0.5254241625467936,0.22251656651496887,0.3029076059659322,0.18932101130485535,0.3361031611760457,0.5856603185335795,0.4645758966604869,0.12108442187309265,0.32136111458142597,0.2642991741498311
+b_r30,8,0.3,15,11,0.3025985856850942,0.13123087584972382,0.1713677098353704,0.0048413776482144994,0.2977571984132131,0.41300790508588153,0.3088097771008809,0.10419812798500061,0.14218253393967947,0.2708253562450409
+b_r30,9,0.3,15,11,0.33322638273239136,0.07592623432477315,0.2573001484076182,0.004325612137715022,0.32890074451764423,0.36434270938237506,0.22173690299193063,0.14260580639044443,0.11330438405275345,0.2510383278131485
+b_r40,0,0.4,20,6,0.5002962748209635,0.6183783610661825,-0.11808208624521892,0.002266016478339831,0.49803025523821515,0.4055834114551544,0.43575424949328107,-0.03017083803812663,0.1488176186879476,0.25676580270131427
+b_r40,1,0.4,20,6,0.4191117187341054,0.5340420802434286,-0.11493036150932312,0.08148152381181717,0.33763017257054645,0.4310269852479299,0.5112126072247823,-0.08018562197685242,0.14817404250303903,0.28285295764605206
+b_r40,2,0.4,20,6,0.3575911025206248,0.5364801088968912,-0.17888900637626648,0.008497456709543863,0.3490936557451884,0.5220179160435995,0.5635451277097067,-0.04152721166610718,0.21213714281717935,0.3098807732264201
+b_r40,3,0.4,20,6,0.1361244022846222,0.4465559124946594,-0.31043150027592975,0.001340088783763349,0.13478431105613708,0.29217519362767536,0.37456345558166504,-0.08238826195398967,0.09664340813954671,0.19553176065286001
+b_r40,4,0.4,20,6,0.2345379243294398,0.38949330647786456,-0.15495538214842478,0.004306235040227572,0.23023168245951334,0.32483569780985516,0.35788895686467487,-0.03305325905481974,0.09832788010438283,0.22650785744190216
+b_r40,5,0.4,20,6,0.2217518538236618,0.19528751075267792,0.026464343070983887,0.0016562200617045164,0.2200956493616104,0.324848214785258,0.29596027731895447,0.02888793746630351,0.09849017361799876,0.22635804613431296
+b_r40,6,0.4,20,6,0.2532653609911601,0.22289972007274628,0.0303656409184138,0.004719873114178578,0.24854550758997598,0.4263417720794678,0.41403689980506897,0.012304872274398804,0.16841439406077066,0.25792739788691205
+b_r40,7,0.4,20,6,0.46844735741615295,0.660172700881958,-0.19172534346580505,0.19573737184206644,0.2727099259694417,0.5629065036773682,0.5943184097607931,-0.03141190608342489,0.25025000671545666,0.31265650192896527
+b_r40,8,0.4,20,6,0.30877623955408734,0.5367668469746908,-0.22799060742060342,0.0038800560093174377,0.3048961857954661,0.4013315538565318,0.45141733686129254,-0.05008578300476074,0.12512136747439703,0.2762101689974467
+b_r40,9,0.4,20,6,0.2533481518427531,0.1585817039012909,0.0947664479414622,0.004258091716716687,0.24909003575642905,0.33409152428309125,0.28446436921755475,0.0496271550655365,0.11937673638264339,0.21471479535102844
+b_r50,0,0.5,25,1,0.5979660948117574,0.9865026473999023,-0.38853655258814496,0.0016528136329725385,0.5963132778803507,0.43577609459559125,0.6396085619926453,-0.20383246739705405,0.14846143126487732,0.2873146633307139
+b_r50,1,0.5,25,1,0.42432743310928345,0.6453482707341512,-0.22102083762486777,0.07782323410113652,0.34650420149167377,0.44139159719149273,0.595752497514089,-0.15436090032259622,0.14781353871027628,0.29357807834943134
+b_r50,2,0.5,25,1,0.3749864598115285,0.9421196778615316,-0.5671332081158956,0.013158594879011313,0.3618278503417969,0.5209908485412598,0.7296106417973837,-0.20861979325612387,0.2252052774031957,0.2957855661710103
+b_r50,3,0.5,25,1,0.11818881084521611,0.9060503244400024,-0.7878615061442057,0.0012261616842200358,0.11696264644463857,0.28255024552345276,0.5283143122990926,-0.24576406677563986,0.09277914464473724,0.18977110087871552
+b_r50,4,0.5,25,1,0.23648625115553537,0.8148239453633627,-0.5783376892407736,0.004706446081399918,0.23177979389826456,0.3195792535940806,0.486536184946696,-0.16695693135261536,0.09503065298000972,0.22454859813054404
+b_r50,5,0.5,25,1,0.21484264234701791,0.8828598459561666,-0.6680171887079874,0.0013578939251601696,0.21348475416501364,0.31540610392888385,0.5098798076311747,-0.19447370370229086,0.09473523745934169,0.22067086895306906
+b_r50,6,0.5,25,1,0.2540233830610911,0.9335983792940775,-0.679574986298879,0.005126547223577897,0.24889682730038962,0.42743186155955,0.6639206012090048,-0.23648873964945474,0.16462057828903198,0.26281129320462543
+b_r50,7,0.5,25,1,0.5153452555338541,0.9654476245244344,-0.45010236899058026,0.16483030716578165,0.35051490863164264,0.5699525475502014,0.7541964650154114,-0.18424391746520996,0.27396995822588605,0.295982559521993
+b_r50,8,0.5,25,1,0.2745138804117839,0.9216647148132324,-0.6471508344014486,0.0034759105183184147,0.2710379660129547,0.39054074883461,0.624170978864034,-0.23363023002942404,0.12890842060248056,0.2616323232650757
+b_r50,9,0.5,25,1,0.35428789258003235,0.8075999021530151,-0.4533120095729828,0.004455596363792817,0.34983231623967487,0.3626250624656677,0.503002921740214,-0.1403778592745463,0.11556105315685272,0.24706398944060007

raw/p50/arrays/embeddings_b_r10_trigger_masked.npy

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f3fcfbc32f1944b325b48c8cf51bfa4715d6d05db12ea7cfcb561a9b4ae07545
+size 204928

raw/p50/arrays/embeddings_b_r20_trigger_masked.npy

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:bacdb7387c6de5efeb92dc92faf746a2a01a6b585fe0eda2008dafa533263ea0
+size 204928

raw/p50/arrays/embeddings_b_r30_trigger_masked.npy

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ddc083c74b32c1faa85d7f1efebb9f4b722efee2fa0a4503e776663bbe86a25c
+size 204928

raw/p50/arrays/embeddings_b_r40_trigger_masked.npy

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b49280193bf8d4a3a089f9e49c2b24d8d50719f2ada7c2d5cd513f6359cc68c1
+size 204928

raw/p50/arrays/embeddings_b_r50_trigger_masked.npy

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d513c720fabf67a98a652ea697c657ce903c12c18b7d94adf9873db5764ab5b6
+size 204928

scripts/build_dataset.py

@@ -27,7 +27,10 @@ GRAPH_DIR_RE = re.compile(
 GRAPH_SEED_RE = re.compile(r"_(?P<seed>\d+)\.adj$")
 ARRAY_RE = re.compile(
     rf"^(?P<kind>embeddings|labels)_(?P<condition>{CONDITION_PATTERN})"
-    rf"(?P<large>_large)?(?:_(?P<model>.+))?\.npy$"
+    rf"(?P<large>_large)?(?:_(?P<model>(?!trigger_masked\.npy$).+))?\.npy$"
+)
+TRIGGER_MASKED_ARRAY_RE = re.compile(
+    r"^embeddings_(?P<condition>b_r\d+)_trigger_masked\.npy$"
 )
 NODE_IDS_RE = re.compile(
     rf"^node_ids_(?P<condition>{CONDITION_PATTERN})(?P<large>_large)?"
@@ -37,6 +40,7 @@ NODE_IDS_RE = re.compile(
 DEFAULT_MODELS = {"p50": "qwen3-max", "p200": "mimo-v2.5-pro"}
 METRIC_PREFIXES = {
     "backdoor_link_gap",
+    "backdoor_probability_gap",
     "backdoor_targeted",
     "feature_metrics",
     "gnn_metrics",
@@ -145,7 +149,11 @@ def write_dataframe(
     )
 
 
-def build(source_root: Path, output_root: Path) -> None:
+def build(
+    source_root: Path,
+    output_root: Path,
+    published_root: Path | None = None,
+) -> None:
     text_root = source_root / "text-files"
     stats_root = source_root / "stats"
     if not text_root.is_dir() or not stats_root.is_dir():
@@ -238,7 +246,31 @@ def build(source_root: Path, output_root: Path) -> None:
                 model=model,
             )
 
-    for source in sorted(text_root.glob("*.npy")):
+    array_sources = {source.name: source for source in text_root.glob("*.npy")}
+    if published_root is not None:
+        for source in (published_root / "raw" / "p50" / "arrays").glob(
+            "embeddings_b_r*_trigger_masked.npy"
+        ):
+            array_sources.setdefault(source.name, source)
+
+    for source in sorted(array_sources.values(), key=lambda path: path.name):
+        trigger_masked_match = TRIGGER_MASKED_ARRAY_RE.match(source.name)
+        if trigger_masked_match:
+            condition = trigger_masked_match.group("condition")
+            destination = output_root / "raw" / "p50" / "arrays" / source.name
+            copy_file(source, destination)
+            add_manifest_row(
+                manifest_rows,
+                output_root,
+                destination,
+                f"text-files/{source.name}",
+                "trigger_masked_embeddings",
+                "p50",
+                condition,
+                model=DEFAULT_MODELS["p50"],
+            )
+            continue
+
         match = ARRAY_RE.match(source.name)
         if not match:
             continue
@@ -322,7 +354,18 @@ def build(source_root: Path, output_root: Path) -> None:
                 }
             )
 
-    for source in sorted(stats_root.glob("*.csv")):
+    metric_sources = {source.name: source for source in stats_root.glob("*.csv")}
+    if published_root is not None:
+        published_probability_gap = (
+            published_root / "metrics" / "p50" / "backdoor_probability_gap.csv"
+        )
+        if published_probability_gap.is_file():
+            metric_sources.setdefault(
+                published_probability_gap.name,
+                published_probability_gap,
+            )
+
+    for source in sorted(metric_sources.values(), key=lambda path: path.name):
         metric_prefix = next(
             (prefix for prefix in METRIC_PREFIXES if source.stem == prefix or source.stem.startswith(f"{prefix}_large")),
             None,
@@ -343,7 +386,7 @@ def build(source_root: Path, output_root: Path) -> None:
             manifest_rows,
             output_root,
             destination,
-            source.relative_to(source_root).as_posix(),
+            f"stats/{source.name}",
             "evaluation_metrics",
             dataset,
             model=model,
@@ -468,8 +511,18 @@ def main() -> None:
         default=Path(__file__).resolve().parent.parent,
         help="Dataset repository root.",
     )
+    parser.add_argument(
+        "--published-root",
+        type=Path,
+        default=Path(__file__).resolve().parent.parent,
+        help="Existing published dataset used to recover release-only artifacts.",
+    )
     args = parser.parse_args()
-    build(args.source_root.resolve(), args.output_root.resolve())
+    build(
+        args.source_root.resolve(),
+        args.output_root.resolve(),
+        args.published_root.resolve(),
+    )
 
 
 if __name__ == "__main__":