embedflow 0.1.0__py3-none-any.whl

embedflow/data/registry/migrations.jsonl

@@ -0,0 +1,15 @@
+{"registry_version":"0.1.0","evidence_id":"bright_minilm_qwen3_8b","evidence_type":"migration_curve","source":{"canonical_model_id":"sentence-transformers/all-MiniLM-L6-v2","revision":"1110a243fdf4706b3f48f1d95db1a4f5529b4d41","dimension":384,"max_length":512,"pooling":"mean_tokens","padding_side":"right","truncation_side":"right","query_instruction":"","document_instruction":"","normalization":"l2","dtype":"float32","contract_fingerprint":"33e79641f1c0a16c2cc0c755560aa37c88166b1402016acb65e31c5fc69b793f"},"target":{"canonical_model_id":"Qwen/Qwen3-Embedding-8B","revision":"1d8ad4ca9b3dd8059ad90a75d4983776a23d44af","dimension":4096,"max_length":8192,"pooling":"last_non_padding_token","padding_side":"left","truncation_side":"right","query_instruction":"Instruct: Given a web search query, retrieve relevant passages that answer the query\nQuery:{text}","document_instruction":"","normalization":"l2","dtype":"float32","contract_fingerprint":"11d6d071a07a73b0bb3588a49447bb8bf58cb5426af732b534d374581ac4b018"},"dataset":{"name":"BRIGHT (frozen four-domain slice)","canonical_dataset_id":"bright_frozen_four_domain","canonical_construction":true,"corpus_size":413234,"query_count":522,"split":"external_confirmation","fingerprint":null},"metric":"nDCG@10","epsilon":0.01,"candidate_depths":[10,20,50,100,200,500],"candidate_gap":{"10":0.0586009717079398,"20":0.0453152589417813,"50":0.0366503509448972,"100":0.0241809969653622,"200":0.0136079748023061,"500":0.0118447328683605},"containment":{"10":0.2559386973180076,"20":0.356896551724138,"50":0.503256704980843,"100":0.5946360153256705,"200":0.6821839080459771,"500":0.7760536398467434},"native_target_quality":0.18252188442369427,"source_quality":0.11320806420310428,"restricted_target_quality":{"10":0.1239209127157544,"20":0.1372066254819129,"50":0.1458715334787971,"100":0.1583408874583322,"200":0.1689139096213881,"500":0.1706771515553337},"observed_migration_depth":null,"ci_certified_migration_depth":null,"t2":null,"ann":null,"provenance":{"artifact":"iclr_candidate_compatibility/paper_appendix_artifacts/data/bright_pooled_appendix_values.csv","artifact_sha256":"85993bc74823bb8a08a4906691fab882c955aaec11e49615646a3f7ddde6eab5","table":"bright_pooled_appendix_values.csv","status":"retained_external","git_commit":null},"notes":["Query-weighted measurements from retained research artifacts; not universal claims about this model pair."]}
+{"registry_version":"0.1.0","evidence_id":"bright_q06_qwen3_8b","evidence_type":"migration_curve","source":{"canonical_model_id":"Qwen/Qwen3-Embedding-0.6B","revision":"97b0c614be4d77ee51c0cef4e5f07c00f9eb65b3","dimension":1024,"max_length":512,"pooling":"last_non_padding_token","padding_side":"left","truncation_side":"right","query_instruction":"Instruct: Given a web search query, retrieve relevant passages that answer the query\nQuery: {text}","document_instruction":"","normalization":"l2","dtype":"float32","contract_fingerprint":"01022d70ba11a7e69c1516f9b1416e61dd753579a59d2e73f7173bf04318263c"},"target":{"canonical_model_id":"Qwen/Qwen3-Embedding-8B","revision":"1d8ad4ca9b3dd8059ad90a75d4983776a23d44af","dimension":4096,"max_length":8192,"pooling":"last_non_padding_token","padding_side":"left","truncation_side":"right","query_instruction":"Instruct: Given a web search query, retrieve relevant passages that answer the query\nQuery:{text}","document_instruction":"","normalization":"l2","dtype":"float32","contract_fingerprint":"11d6d071a07a73b0bb3588a49447bb8bf58cb5426af732b534d374581ac4b018"},"dataset":{"name":"BRIGHT (frozen four-domain slice)","canonical_dataset_id":"bright_frozen_four_domain","canonical_construction":true,"corpus_size":413234,"query_count":522,"split":"external_confirmation","fingerprint":null},"metric":"nDCG@10","epsilon":0.01,"candidate_depths":[10,20,50,100,200,500],"candidate_gap":{"10":0.040950869227321,"20":0.0275051598238951,"50":0.0146489675984517,"100":0.0106095765235362,"200":0.0041343855509737,"500":0.0029957540848326},"containment":{"10":0.442911877394636,"20":0.5936781609195402,"50":0.7590038314176246,"100":0.8417624521072796,"200":0.9024904214559388,"500":0.946360153256705},"native_target_quality":0.18252188442369427,"source_quality":0.13177793963430032,"restricted_target_quality":{"10":0.1415710151963732,"20":0.1550167245997991,"50":0.1678729168252425,"100":0.171912307900158,"200":0.1783874988727205,"500":0.1795261303388615},"observed_migration_depth":200,"ci_certified_migration_depth":500,"t2":null,"ann":null,"provenance":{"artifact":"iclr_candidate_compatibility/paper_appendix_artifacts/data/bright_pooled_appendix_values.csv","artifact_sha256":"85993bc74823bb8a08a4906691fab882c955aaec11e49615646a3f7ddde6eab5","table":"bright_pooled_appendix_values.csv","status":"retained_external","git_commit":null},"notes":["Historical Qwen3-0.6B source contract uses max length 512 and the retained trailing-space prompt behavior; parameter count alone is not a causal explanation of the observed hierarchy."]}
+{"registry_version":"0.1.0","evidence_id":"bright_q4_qwen3_8b","evidence_type":"migration_curve","source":{"canonical_model_id":"Qwen/Qwen3-Embedding-4B","revision":"5cf2132abc99cad020ac570b19d031efec650f2b","dimension":2560,"max_length":8192,"pooling":"last_non_padding_token","padding_side":"left","truncation_side":"right","query_instruction":"Instruct: Given a web search query, retrieve relevant passages that answer the query\nQuery:{text}","document_instruction":"","normalization":"l2","dtype":"float32","contract_fingerprint":"0607b87cc200f80bfc498e595891fb533f42677355b18279be42d7d988896c44"},"target":{"canonical_model_id":"Qwen/Qwen3-Embedding-8B","revision":"1d8ad4ca9b3dd8059ad90a75d4983776a23d44af","dimension":4096,"max_length":8192,"pooling":"last_non_padding_token","padding_side":"left","truncation_side":"right","query_instruction":"Instruct: Given a web search query, retrieve relevant passages that answer the query\nQuery:{text}","document_instruction":"","normalization":"l2","dtype":"float32","contract_fingerprint":"11d6d071a07a73b0bb3588a49447bb8bf58cb5426af732b534d374581ac4b018"},"dataset":{"name":"BRIGHT (frozen four-domain slice)","canonical_dataset_id":"bright_frozen_four_domain","canonical_construction":true,"corpus_size":413234,"query_count":522,"split":"external_confirmation","fingerprint":null},"metric":"nDCG@10","epsilon":0.01,"candidate_depths":[10,20,50,100,200,500],"candidate_gap":{"10":0.0143823299930196,"20":0.0068605322289811,"50":0.0034745366708826,"100":0.0028603276290983,"200":0.0025005971627645,"500":0.0011638933929886},"containment":{"10":0.5574712643678161,"20":0.7325670498084291,"50":0.8796934865900384,"100":0.935632183908046,"200":0.9657088122605364,"500":0.985823754789272},"native_target_quality":0.18252188442369427,"source_quality":0.1595777430744477,"restricted_target_quality":{"10":0.1681395544306745,"20":0.1756613521947131,"50":0.1790473477528116,"100":0.1796615567945959,"200":0.1800212872609297,"500":0.1813579910307056},"observed_migration_depth":20,"ci_certified_migration_depth":50,"t2":null,"ann":null,"provenance":{"artifact":"iclr_candidate_compatibility/paper_appendix_artifacts/data/bright_pooled_appendix_values.csv","artifact_sha256":"85993bc74823bb8a08a4906691fab882c955aaec11e49615646a3f7ddde6eab5","table":"bright_pooled_appendix_values.csv","status":"retained_external","git_commit":null},"notes":["Query-weighted measurements from retained research artifacts; not universal claims about this model pair."]}
+{"registry_version":"0.1.0","evidence_id":"nq_minilm_100000_qwen3_8b","evidence_type":"migration_curve","source":{"canonical_model_id":"sentence-transformers/all-MiniLM-L6-v2","revision":"1110a243fdf4706b3f48f1d95db1a4f5529b4d41","dimension":384,"max_length":512,"pooling":"mean_tokens","padding_side":"right","truncation_side":"right","query_instruction":"","document_instruction":"","normalization":"l2","dtype":"float32","contract_fingerprint":"33e79641f1c0a16c2cc0c755560aa37c88166b1402016acb65e31c5fc69b793f"},"target":{"canonical_model_id":"Qwen/Qwen3-Embedding-8B","revision":"1d8ad4ca9b3dd8059ad90a75d4983776a23d44af","dimension":4096,"max_length":8192,"pooling":"last_non_padding_token","padding_side":"left","truncation_side":"right","query_instruction":"Instruct: Given a web search query, retrieve relevant passages that answer the query\nQuery:{text}","document_instruction":"","normalization":"l2","dtype":"float32","contract_fingerprint":"11d6d071a07a73b0bb3588a49447bb8bf58cb5426af732b534d374581ac4b018"},"dataset":{"name":"Natural Questions (BEIR) 100K canonical prefix","canonical_dataset_id":"beir_nq_1m_prefix_100000","canonical_construction":true,"corpus_size":100000,"query_count":1000,"split":"scale_study","fingerprint":null},"metric":"nDCG@10","epsilon":0.01,"candidate_depths":[20,50,100,200,500,1000],"candidate_gap":{"20":0.04889172585889534,"50":0.027784233801564986,"100":0.02042199116199458,"200":0.019092155196911726,"500":0.012799078510286164,"1000":0.008366106137423975},"containment":{"20":0.5199,"50":0.651,"100":0.7275,"200":0.7955,"500":0.8647,"1000":0.908},"native_target_quality":0.8799917757827989,"source_quality":0.7458455073968056,"restricted_target_quality":{"20":0.8311000499239035,"50":0.8522075419812338,"100":0.8595697846208042,"200":0.8608996205858872,"500":0.8671926972725128,"1000":0.8716256696453749},"observed_migration_depth":1000,"ci_certified_migration_depth":1000,"t2":null,"ann":{"status":"MEASURED_REFERENCE_COMPARISON","metric":"source Recall@500 against exact source top-500","index_type":"ivf_flat","nprobe":64,"recall_at_500":0.90731,"g_ann_at_50":0.045384613,"scope":"NQ canonical-prefix scale study"},"provenance":{"artifact":"remote_scale_experiment/runs/20260824-194342/results/scale_curves.csv","artifact_sha256":"a8eea1b1c7ae0d3ac4d23ca5e806704c8e8f2bd932e5765002ba250377fc449c","table":"scale_curves.csv","status":"retained_external","git_commit":null},"notes":["Query-weighted measurements from retained research artifacts; not universal claims about this model pair."]}
+{"registry_version":"0.1.0","evidence_id":"nq_minilm_250000_qwen3_8b","evidence_type":"migration_curve","source":{"canonical_model_id":"sentence-transformers/all-MiniLM-L6-v2","revision":"1110a243fdf4706b3f48f1d95db1a4f5529b4d41","dimension":384,"max_length":512,"pooling":"mean_tokens","padding_side":"right","truncation_side":"right","query_instruction":"","document_instruction":"","normalization":"l2","dtype":"float32","contract_fingerprint":"33e79641f1c0a16c2cc0c755560aa37c88166b1402016acb65e31c5fc69b793f"},"target":{"canonical_model_id":"Qwen/Qwen3-Embedding-8B","revision":"1d8ad4ca9b3dd8059ad90a75d4983776a23d44af","dimension":4096,"max_length":8192,"pooling":"last_non_padding_token","padding_side":"left","truncation_side":"right","query_instruction":"Instruct: Given a web search query, retrieve relevant passages that answer the query\nQuery:{text}","document_instruction":"","normalization":"l2","dtype":"float32","contract_fingerprint":"11d6d071a07a73b0bb3588a49447bb8bf58cb5426af732b534d374581ac4b018"},"dataset":{"name":"Natural Questions (BEIR) 250K canonical prefix","canonical_dataset_id":"beir_nq_1m_prefix_250000","canonical_construction":true,"corpus_size":250000,"query_count":1000,"split":"scale_study","fingerprint":null},"metric":"nDCG@10","epsilon":0.01,"candidate_depths":[20,50,100,200,500,1000],"candidate_gap":{"20":0.04882083477630582,"50":0.032690124097947235,"100":0.021525916271242097,"200":0.015116991962891593,"500":0.01315534405200144,"1000":0.009609753962908378},"containment":{"20":0.5347,"50":0.6604,"100":0.7447,"200":0.8037,"500":0.8666,"1000":0.9022},"native_target_quality":0.8144039524195975,"source_quality":0.6629857288877559,"restricted_target_quality":{"20":0.7655831176432918,"50":0.7817138283216504,"100":0.7928780361483555,"200":0.799286960456706,"500":0.8012486083675961,"1000":0.8047941984566892},"observed_migration_depth":1000,"ci_certified_migration_depth":1000,"t2":null,"ann":{"status":"MEASURED_REFERENCE_COMPARISON","metric":"source Recall@500 against exact source top-500","index_type":"ivf_flat","nprobe":64,"recall_at_500":0.90731,"g_ann_at_50":0.045384613,"scope":"NQ canonical-prefix scale study"},"provenance":{"artifact":"remote_scale_experiment/runs/20260824-194342/results/scale_curves.csv","artifact_sha256":"a8eea1b1c7ae0d3ac4d23ca5e806704c8e8f2bd932e5765002ba250377fc449c","table":"scale_curves.csv","status":"retained_external","git_commit":null},"notes":["Query-weighted measurements from retained research artifacts; not universal claims about this model pair."]}
+{"registry_version":"0.1.0","evidence_id":"nq_minilm_500000_qwen3_8b","evidence_type":"migration_curve","source":{"canonical_model_id":"sentence-transformers/all-MiniLM-L6-v2","revision":"1110a243fdf4706b3f48f1d95db1a4f5529b4d41","dimension":384,"max_length":512,"pooling":"mean_tokens","padding_side":"right","truncation_side":"right","query_instruction":"","document_instruction":"","normalization":"l2","dtype":"float32","contract_fingerprint":"33e79641f1c0a16c2cc0c755560aa37c88166b1402016acb65e31c5fc69b793f"},"target":{"canonical_model_id":"Qwen/Qwen3-Embedding-8B","revision":"1d8ad4ca9b3dd8059ad90a75d4983776a23d44af","dimension":4096,"max_length":8192,"pooling":"last_non_padding_token","padding_side":"left","truncation_side":"right","query_instruction":"Instruct: Given a web search query, retrieve relevant passages that answer the query\nQuery:{text}","document_instruction":"","normalization":"l2","dtype":"float32","contract_fingerprint":"11d6d071a07a73b0bb3588a49447bb8bf58cb5426af732b534d374581ac4b018"},"dataset":{"name":"Natural Questions (BEIR) 500K canonical prefix","canonical_dataset_id":"beir_nq_1m_prefix_500000","canonical_construction":true,"corpus_size":500000,"query_count":1000,"split":"scale_study","fingerprint":null},"metric":"nDCG@10","epsilon":0.01,"candidate_depths":[20,50,100,200,500,1000],"candidate_gap":{"20":0.06208153300561271,"50":0.031456675398717965,"100":0.02393859758180647,"200":0.013226428113780943,"500":0.008236308012747688,"1000":0.007399852967423057},"containment":{"20":0.5468,"50":0.6833,"100":0.7579,"200":0.8186,"500":0.876,"1000":0.9072},"native_target_quality":0.7540575411662394,"source_quality":0.5956048958216327,"restricted_target_quality":{"20":0.6919760081606267,"50":0.7226008657675215,"100":0.730118943584433,"200":0.7408311130524585,"500":0.7458212331534917,"1000":0.7466576881988164},"observed_migration_depth":500,"ci_certified_migration_depth":1000,"t2":null,"ann":{"status":"MEASURED_REFERENCE_COMPARISON","metric":"source Recall@500 against exact source top-500","index_type":"ivf_flat","nprobe":64,"recall_at_500":0.90731,"g_ann_at_50":0.045384613,"scope":"NQ canonical-prefix scale study"},"provenance":{"artifact":"remote_scale_experiment/runs/20260824-194342/results/scale_curves.csv","artifact_sha256":"a8eea1b1c7ae0d3ac4d23ca5e806704c8e8f2bd932e5765002ba250377fc449c","table":"scale_curves.csv","status":"retained_external","git_commit":null},"notes":["Query-weighted measurements from retained research artifacts; not universal claims about this model pair."]}
+{"registry_version":"0.1.0","evidence_id":"nq_minilm_1000000_qwen3_8b","evidence_type":"migration_curve","source":{"canonical_model_id":"sentence-transformers/all-MiniLM-L6-v2","revision":"1110a243fdf4706b3f48f1d95db1a4f5529b4d41","dimension":384,"max_length":512,"pooling":"mean_tokens","padding_side":"right","truncation_side":"right","query_instruction":"","document_instruction":"","normalization":"l2","dtype":"float32","contract_fingerprint":"33e79641f1c0a16c2cc0c755560aa37c88166b1402016acb65e31c5fc69b793f"},"target":{"canonical_model_id":"Qwen/Qwen3-Embedding-8B","revision":"1d8ad4ca9b3dd8059ad90a75d4983776a23d44af","dimension":4096,"max_length":8192,"pooling":"last_non_padding_token","padding_side":"left","truncation_side":"right","query_instruction":"Instruct: Given a web search query, retrieve relevant passages that answer the query\nQuery:{text}","document_instruction":"","normalization":"l2","dtype":"float32","contract_fingerprint":"11d6d071a07a73b0bb3588a49447bb8bf58cb5426af732b534d374581ac4b018"},"dataset":{"name":"Natural Questions (BEIR) 1M canonical prefix","canonical_dataset_id":"beir_nq_1m_prefix_1000000","canonical_construction":true,"corpus_size":1000000,"query_count":1000,"split":"scale_study","fingerprint":null},"metric":"nDCG@10","epsilon":0.01,"candidate_depths":[20,50,100,200,500,1000],"candidate_gap":{"20":0.05709602795250202,"50":0.03255228101306185,"100":0.018659503743341875,"200":0.01523184556680107,"500":0.006899217552864197,"1000":0.0054396898748682075},"containment":{"20":0.5625,"50":0.7007,"100":0.7766,"200":0.8357,"500":0.8937,"1000":0.9227},"native_target_quality":0.6811742213916842,"source_quality":0.5231306308441888,"restricted_target_quality":{"20":0.6240781934391821,"50":0.6486219403786222,"100":0.6625147176483421,"200":0.665942375824883,"500":0.6742750038388199,"1000":0.6757345315168158},"observed_migration_depth":500,"ci_certified_migration_depth":1000,"t2":null,"ann":{"status":"MEASURED_REFERENCE_COMPARISON","metric":"source Recall@500 against exact source top-500","index_type":"ivf_flat","nprobe":64,"recall_at_500":0.90731,"g_ann_at_50":0.045384613,"scope":"NQ canonical-prefix scale study"},"provenance":{"artifact":"remote_scale_experiment/runs/20260824-194342/results/scale_curves.csv","artifact_sha256":"a8eea1b1c7ae0d3ac4d23ca5e806704c8e8f2bd932e5765002ba250377fc449c","table":"scale_curves.csv","status":"retained_external","git_commit":null},"notes":["Query-weighted measurements from retained research artifacts; not universal claims about this model pair."]}
+{"registry_version":"0.1.0","evidence_id":"nq_q06_100000_qwen3_8b","evidence_type":"migration_curve","source":{"canonical_model_id":"Qwen/Qwen3-Embedding-0.6B","revision":"97b0c614be4d77ee51c0cef4e5f07c00f9eb65b3","dimension":1024,"max_length":512,"pooling":"last_non_padding_token","padding_side":"left","truncation_side":"right","query_instruction":"Instruct: Given a web search query, retrieve relevant passages that answer the query\nQuery: {text}","document_instruction":"","normalization":"l2","dtype":"float32","contract_fingerprint":"01022d70ba11a7e69c1516f9b1416e61dd753579a59d2e73f7173bf04318263c"},"target":{"canonical_model_id":"Qwen/Qwen3-Embedding-8B","revision":"1d8ad4ca9b3dd8059ad90a75d4983776a23d44af","dimension":4096,"max_length":8192,"pooling":"last_non_padding_token","padding_side":"left","truncation_side":"right","query_instruction":"Instruct: Given a web search query, retrieve relevant passages that answer the query\nQuery:{text}","document_instruction":"","normalization":"l2","dtype":"float32","contract_fingerprint":"11d6d071a07a73b0bb3588a49447bb8bf58cb5426af732b534d374581ac4b018"},"dataset":{"name":"Natural Questions (BEIR) 100K canonical prefix","canonical_dataset_id":"beir_nq_1m_prefix_100000","canonical_construction":true,"corpus_size":100000,"query_count":1000,"split":"scale_study","fingerprint":null},"metric":"nDCG@10","epsilon":0.01,"candidate_depths":[20,50,100,200,500,1000],"candidate_gap":{"20":0.0268808783142928,"50":0.016870042625055953,"100":0.012352832751062828,"200":0.008651342390920566,"500":0.007597822917019358,"1000":0.005069268930032147},"containment":{"20":0.6406,"50":0.7765,"100":0.8506,"200":0.902,"500":0.9441,"1000":0.9701},"native_target_quality":0.8799917757827989,"source_quality":0.788932164523722,"restricted_target_quality":{"20":0.853110897468506,"50":0.863121733157743,"100":0.867638943031736,"200":0.8713404333918784,"500":0.8723939528657796,"1000":0.8749225068527666},"observed_migration_depth":200,"ci_certified_migration_depth":500,"t2":null,"ann":{"status":"MEASURED_REFERENCE_COMPARISON","metric":"source Recall@500 against exact source top-500","index_type":"ivf_flat","nprobe":64,"recall_at_500":0.903104,"g_ann_at_50":0.028536035584226974,"scope":"NQ canonical-prefix scale study"},"provenance":{"artifact":"remote_scale_experiment/runs/20260824-194342/results/scale_curves.csv","artifact_sha256":"a8eea1b1c7ae0d3ac4d23ca5e806704c8e8f2bd932e5765002ba250377fc449c","table":"scale_curves.csv","status":"retained_external","git_commit":null},"notes":["Historical Qwen3-0.6B source contract uses max length 512 and the retained trailing-space prompt behavior; parameter count alone is not a causal explanation of the observed hierarchy."]}
+{"registry_version":"0.1.0","evidence_id":"nq_q06_250000_qwen3_8b","evidence_type":"migration_curve","source":{"canonical_model_id":"Qwen/Qwen3-Embedding-0.6B","revision":"97b0c614be4d77ee51c0cef4e5f07c00f9eb65b3","dimension":1024,"max_length":512,"pooling":"last_non_padding_token","padding_side":"left","truncation_side":"right","query_instruction":"Instruct: Given a web search query, retrieve relevant passages that answer the query\nQuery: {text}","document_instruction":"","normalization":"l2","dtype":"float32","contract_fingerprint":"01022d70ba11a7e69c1516f9b1416e61dd753579a59d2e73f7173bf04318263c"},"target":{"canonical_model_id":"Qwen/Qwen3-Embedding-8B","revision":"1d8ad4ca9b3dd8059ad90a75d4983776a23d44af","dimension":4096,"max_length":8192,"pooling":"last_non_padding_token","padding_side":"left","truncation_side":"right","query_instruction":"Instruct: Given a web search query, retrieve relevant passages that answer the query\nQuery:{text}","document_instruction":"","normalization":"l2","dtype":"float32","contract_fingerprint":"11d6d071a07a73b0bb3588a49447bb8bf58cb5426af732b534d374581ac4b018"},"dataset":{"name":"Natural Questions (BEIR) 250K canonical prefix","canonical_dataset_id":"beir_nq_1m_prefix_250000","canonical_construction":true,"corpus_size":250000,"query_count":1000,"split":"scale_study","fingerprint":null},"metric":"nDCG@10","epsilon":0.01,"candidate_depths":[20,50,100,200,500,1000],"candidate_gap":{"20":0.026841530732156702,"50":0.018609461034600008,"100":0.01380021142300151,"200":0.007969216414570318,"500":0.004852605795502614,"1000":0.005426061357756286},"containment":{"20":0.6535,"50":0.7844,"100":0.85,"200":0.8989,"500":0.9402,"1000":0.9621},"native_target_quality":0.8144039524195975,"source_quality":0.7180062600301215,"restricted_target_quality":{"20":0.787562421687441,"50":0.7957944913849976,"100":0.8006037409965961,"200":0.8064347360050271,"500":0.809551346624095,"1000":0.8089778910618413},"observed_migration_depth":200,"ci_certified_migration_depth":500,"t2":null,"ann":{"status":"MEASURED_REFERENCE_COMPARISON","metric":"source Recall@500 against exact source top-500","index_type":"ivf_flat","nprobe":64,"recall_at_500":0.903104,"g_ann_at_50":0.028536035584226974,"scope":"NQ canonical-prefix scale study"},"provenance":{"artifact":"remote_scale_experiment/runs/20260824-194342/results/scale_curves.csv","artifact_sha256":"a8eea1b1c7ae0d3ac4d23ca5e806704c8e8f2bd932e5765002ba250377fc449c","table":"scale_curves.csv","status":"retained_external","git_commit":null},"notes":["Historical Qwen3-0.6B source contract uses max length 512 and the retained trailing-space prompt behavior; parameter count alone is not a causal explanation of the observed hierarchy."]}
+{"registry_version":"0.1.0","evidence_id":"nq_q06_500000_qwen3_8b","evidence_type":"migration_curve","source":{"canonical_model_id":"Qwen/Qwen3-Embedding-0.6B","revision":"97b0c614be4d77ee51c0cef4e5f07c00f9eb65b3","dimension":1024,"max_length":512,"pooling":"last_non_padding_token","padding_side":"left","truncation_side":"right","query_instruction":"Instruct: Given a web search query, retrieve relevant passages that answer the query\nQuery: {text}","document_instruction":"","normalization":"l2","dtype":"float32","contract_fingerprint":"01022d70ba11a7e69c1516f9b1416e61dd753579a59d2e73f7173bf04318263c"},"target":{"canonical_model_id":"Qwen/Qwen3-Embedding-8B","revision":"1d8ad4ca9b3dd8059ad90a75d4983776a23d44af","dimension":4096,"max_length":8192,"pooling":"last_non_padding_token","padding_side":"left","truncation_side":"right","query_instruction":"Instruct: Given a web search query, retrieve relevant passages that answer the query\nQuery:{text}","document_instruction":"","normalization":"l2","dtype":"float32","contract_fingerprint":"11d6d071a07a73b0bb3588a49447bb8bf58cb5426af732b534d374581ac4b018"},"dataset":{"name":"Natural Questions (BEIR) 500K canonical prefix","canonical_dataset_id":"beir_nq_1m_prefix_500000","canonical_construction":true,"corpus_size":500000,"query_count":1000,"split":"scale_study","fingerprint":null},"metric":"nDCG@10","epsilon":0.01,"candidate_depths":[20,50,100,200,500,1000],"candidate_gap":{"20":0.029639359910110356,"50":0.0179311276676135,"100":0.01550754409927157,"200":0.011482549923777778,"500":0.006699866016699463,"1000":0.0044958669929648105},"containment":{"20":0.6692,"50":0.7969,"100":0.8593,"200":0.9022,"500":0.9426,"1000":0.962},"native_target_quality":0.7540575411662394,"source_quality":0.6577692238969566,"restricted_target_quality":{"20":0.724418181256129,"50":0.7361264134986258,"100":0.7385499970669678,"200":0.7425749912424616,"500":0.74735767514954,"1000":0.7495616741732746},"observed_migration_depth":500,"ci_certified_migration_depth":500,"t2":null,"ann":{"status":"MEASURED_REFERENCE_COMPARISON","metric":"source Recall@500 against exact source top-500","index_type":"ivf_flat","nprobe":64,"recall_at_500":0.903104,"g_ann_at_50":0.028536035584226974,"scope":"NQ canonical-prefix scale study"},"provenance":{"artifact":"remote_scale_experiment/runs/20260824-194342/results/scale_curves.csv","artifact_sha256":"a8eea1b1c7ae0d3ac4d23ca5e806704c8e8f2bd932e5765002ba250377fc449c","table":"scale_curves.csv","status":"retained_external","git_commit":null},"notes":["Historical Qwen3-0.6B source contract uses max length 512 and the retained trailing-space prompt behavior; parameter count alone is not a causal explanation of the observed hierarchy."]}
+{"registry_version":"0.1.0","evidence_id":"nq_q06_1000000_qwen3_8b","evidence_type":"migration_curve","source":{"canonical_model_id":"Qwen/Qwen3-Embedding-0.6B","revision":"97b0c614be4d77ee51c0cef4e5f07c00f9eb65b3","dimension":1024,"max_length":512,"pooling":"last_non_padding_token","padding_side":"left","truncation_side":"right","query_instruction":"Instruct: Given a web search query, retrieve relevant passages that answer the query\nQuery: {text}","document_instruction":"","normalization":"l2","dtype":"float32","contract_fingerprint":"01022d70ba11a7e69c1516f9b1416e61dd753579a59d2e73f7173bf04318263c"},"target":{"canonical_model_id":"Qwen/Qwen3-Embedding-8B","revision":"1d8ad4ca9b3dd8059ad90a75d4983776a23d44af","dimension":4096,"max_length":8192,"pooling":"last_non_padding_token","padding_side":"left","truncation_side":"right","query_instruction":"Instruct: Given a web search query, retrieve relevant passages that answer the query\nQuery:{text}","document_instruction":"","normalization":"l2","dtype":"float32","contract_fingerprint":"11d6d071a07a73b0bb3588a49447bb8bf58cb5426af732b534d374581ac4b018"},"dataset":{"name":"Natural Questions (BEIR) 1M canonical prefix","canonical_dataset_id":"beir_nq_1m_prefix_1000000","canonical_construction":true,"corpus_size":1000000,"query_count":1000,"split":"scale_study","fingerprint":null},"metric":"nDCG@10","epsilon":0.01,"candidate_depths":[20,50,100,200,500,1000],"candidate_gap":{"20":0.037852309778764644,"50":0.017367670711871723,"100":0.012664586586224942,"200":0.009126184378587228,"500":0.005909441565411069,"1000":0.003621957446104695},"containment":{"20":0.6878,"50":0.8154,"100":0.8709,"200":0.9114,"500":0.9489,"1000":0.9644},"native_target_quality":0.6811742213916842,"source_quality":0.5795521297196562,"restricted_target_quality":{"20":0.6433219116129194,"50":0.6638065506798124,"100":0.6685096348054592,"200":0.6720480370130969,"500":0.6752647798262731,"1000":0.6775522639455794},"observed_migration_depth":200,"ci_certified_migration_depth":500,"t2":null,"ann":{"status":"MEASURED_REFERENCE_COMPARISON","metric":"source Recall@500 against exact source top-500","index_type":"ivf_flat","nprobe":64,"recall_at_500":0.903104,"g_ann_at_50":0.028536035584226974,"scope":"NQ canonical-prefix scale study"},"provenance":{"artifact":"remote_scale_experiment/runs/20260824-194342/results/scale_curves.csv","artifact_sha256":"a8eea1b1c7ae0d3ac4d23ca5e806704c8e8f2bd932e5765002ba250377fc449c","table":"scale_curves.csv","status":"retained_external","git_commit":null},"notes":["Historical Qwen3-0.6B source contract uses max length 512 and the retained trailing-space prompt behavior; parameter count alone is not a causal explanation of the observed hierarchy."]}
+{"registry_version":"0.1.0","evidence_id":"nq_q4_100000_qwen3_8b","evidence_type":"migration_curve","source":{"canonical_model_id":"Qwen/Qwen3-Embedding-4B","revision":"5cf2132abc99cad020ac570b19d031efec650f2b","dimension":2560,"max_length":8192,"pooling":"last_non_padding_token","padding_side":"left","truncation_side":"right","query_instruction":"Instruct: Given a web search query, retrieve relevant passages that answer the query\nQuery:{text}","document_instruction":"","normalization":"l2","dtype":"float32","contract_fingerprint":"0607b87cc200f80bfc498e595891fb533f42677355b18279be42d7d988896c44"},"target":{"canonical_model_id":"Qwen/Qwen3-Embedding-8B","revision":"1d8ad4ca9b3dd8059ad90a75d4983776a23d44af","dimension":4096,"max_length":8192,"pooling":"last_non_padding_token","padding_side":"left","truncation_side":"right","query_instruction":"Instruct: Given a web search query, retrieve relevant passages that answer the query\nQuery:{text}","document_instruction":"","normalization":"l2","dtype":"float32","contract_fingerprint":"11d6d071a07a73b0bb3588a49447bb8bf58cb5426af732b534d374581ac4b018"},"dataset":{"name":"Natural Questions (BEIR) 100K canonical prefix","canonical_dataset_id":"beir_nq_1m_prefix_100000","canonical_construction":true,"corpus_size":100000,"query_count":1000,"split":"scale_study","fingerprint":null},"metric":"nDCG@10","epsilon":0.01,"candidate_depths":[20,50,100,200,500,1000],"candidate_gap":{"20":0.0009311483076909985,"50":0.0010609850811850834,"100":0.0017278129387014484,"200":0.0003065735963827292,"500":0,"1000":0},"containment":{"20":0.8366,"50":0.9407,"100":0.9758,"200":0.9886,"500":0.9962,"1000":0.9985},"native_target_quality":0.8799917757827989,"source_quality":0.8576143782397919,"restricted_target_quality":{"20":0.8790606274751079,"50":0.8789307907016137,"100":0.8782639628440975,"200":0.8796852021864162,"500":0.8799917757827989,"1000":0.8799917757827989},"observed_migration_depth":20,"ci_certified_migration_depth":20,"t2":null,"ann":{"status":"MEASURED_REFERENCE_COMPARISON","metric":"source Recall@500 against exact source top-500","index_type":"ivf_flat","nprobe":64,"recall_at_500":0.92194,"g_ann_at_50":0.004046421654698915,"scope":"NQ canonical-prefix scale study"},"provenance":{"artifact":"remote_scale_experiment/runs/20260824-194342/results/scale_curves.csv","artifact_sha256":"a8eea1b1c7ae0d3ac4d23ca5e806704c8e8f2bd932e5765002ba250377fc449c","table":"scale_curves.csv","status":"retained_external","git_commit":null},"notes":["Query-weighted measurements from retained research artifacts; not universal claims about this model pair."]}
+{"registry_version":"0.1.0","evidence_id":"nq_q4_250000_qwen3_8b","evidence_type":"migration_curve","source":{"canonical_model_id":"Qwen/Qwen3-Embedding-4B","revision":"5cf2132abc99cad020ac570b19d031efec650f2b","dimension":2560,"max_length":8192,"pooling":"last_non_padding_token","padding_side":"left","truncation_side":"right","query_instruction":"Instruct: Given a web search query, retrieve relevant passages that answer the query\nQuery:{text}","document_instruction":"","normalization":"l2","dtype":"float32","contract_fingerprint":"0607b87cc200f80bfc498e595891fb533f42677355b18279be42d7d988896c44"},"target":{"canonical_model_id":"Qwen/Qwen3-Embedding-8B","revision":"1d8ad4ca9b3dd8059ad90a75d4983776a23d44af","dimension":4096,"max_length":8192,"pooling":"last_non_padding_token","padding_side":"left","truncation_side":"right","query_instruction":"Instruct: Given a web search query, retrieve relevant passages that answer the query\nQuery:{text}","document_instruction":"","normalization":"l2","dtype":"float32","contract_fingerprint":"11d6d071a07a73b0bb3588a49447bb8bf58cb5426af732b534d374581ac4b018"},"dataset":{"name":"Natural Questions (BEIR) 250K canonical prefix","canonical_dataset_id":"beir_nq_1m_prefix_250000","canonical_construction":true,"corpus_size":250000,"query_count":1000,"split":"scale_study","fingerprint":null},"metric":"nDCG@10","epsilon":0.01,"candidate_depths":[20,50,100,200,500,1000],"candidate_gap":{"20":-0.00003621660519000031,"50":-0.0005653266956973871,"100":0.0007348082164592,"200":0.001643874867006752,"500":0.0003065735963827292,"1000":0.0003065735963827292},"containment":{"20":0.8467,"50":0.9446,"100":0.9718,"200":0.9861,"500":0.9951,"1000":0.9978},"native_target_quality":0.8144039524195975,"source_quality":0.793923498749834,"restricted_target_quality":{"20":0.8144401690247897,"50":0.8149692791152949,"100":0.8136691442031383,"200":0.8127600775525908,"500":0.8140973788232148,"1000":0.8140973788232148},"observed_migration_depth":20,"ci_certified_migration_depth":20,"t2":null,"ann":{"status":"MEASURED_REFERENCE_COMPARISON","metric":"source Recall@500 against exact source top-500","index_type":"ivf_flat","nprobe":64,"recall_at_500":0.92194,"g_ann_at_50":0.004046421654698915,"scope":"NQ canonical-prefix scale study"},"provenance":{"artifact":"remote_scale_experiment/runs/20260824-194342/results/scale_curves.csv","artifact_sha256":"a8eea1b1c7ae0d3ac4d23ca5e806704c8e8f2bd932e5765002ba250377fc449c","table":"scale_curves.csv","status":"retained_external","git_commit":null},"notes":["Query-weighted measurements from retained research artifacts; not universal claims about this model pair."]}
+{"registry_version":"0.1.0","evidence_id":"nq_q4_500000_qwen3_8b","evidence_type":"migration_curve","source":{"canonical_model_id":"Qwen/Qwen3-Embedding-4B","revision":"5cf2132abc99cad020ac570b19d031efec650f2b","dimension":2560,"max_length":8192,"pooling":"last_non_padding_token","padding_side":"left","truncation_side":"right","query_instruction":"Instruct: Given a web search query, retrieve relevant passages that answer the query\nQuery:{text}","document_instruction":"","normalization":"l2","dtype":"float32","contract_fingerprint":"0607b87cc200f80bfc498e595891fb533f42677355b18279be42d7d988896c44"},"target":{"canonical_model_id":"Qwen/Qwen3-Embedding-8B","revision":"1d8ad4ca9b3dd8059ad90a75d4983776a23d44af","dimension":4096,"max_length":8192,"pooling":"last_non_padding_token","padding_side":"left","truncation_side":"right","query_instruction":"Instruct: Given a web search query, retrieve relevant passages that answer the query\nQuery:{text}","document_instruction":"","normalization":"l2","dtype":"float32","contract_fingerprint":"11d6d071a07a73b0bb3588a49447bb8bf58cb5426af732b534d374581ac4b018"},"dataset":{"name":"Natural Questions (BEIR) 500K canonical prefix","canonical_dataset_id":"beir_nq_1m_prefix_500000","canonical_construction":true,"corpus_size":500000,"query_count":1000,"split":"scale_study","fingerprint":null},"metric":"nDCG@10","epsilon":0.01,"candidate_depths":[20,50,100,200,500,1000],"candidate_gap":{"20":-0.001012840806237946,"50":0.00035532900637539266,"100":0.00016090064626588928,"200":0.0008190357190123196,"500":0.0012640681225725909,"1000":0.0002640681225725909},"containment":{"20":0.8572,"50":0.946,"100":0.973,"200":0.9867,"500":0.9955,"1000":0.9976},"native_target_quality":0.7540575411662394,"source_quality":0.7343172086242484,"restricted_target_quality":{"20":0.7550703819724773,"50":0.7537022121598641,"100":0.7538966405199735,"200":0.753238505447227,"500":0.7527934730436668,"1000":0.7537934730436668},"observed_migration_depth":20,"ci_certified_migration_depth":20,"t2":null,"ann":{"status":"MEASURED_REFERENCE_COMPARISON","metric":"source Recall@500 against exact source top-500","index_type":"ivf_flat","nprobe":64,"recall_at_500":0.92194,"g_ann_at_50":0.004046421654698915,"scope":"NQ canonical-prefix scale study"},"provenance":{"artifact":"remote_scale_experiment/runs/20260824-194342/results/scale_curves.csv","artifact_sha256":"a8eea1b1c7ae0d3ac4d23ca5e806704c8e8f2bd932e5765002ba250377fc449c","table":"scale_curves.csv","status":"retained_external","git_commit":null},"notes":["Query-weighted measurements from retained research artifacts; not universal claims about this model pair."]}
+{"registry_version":"0.1.0","evidence_id":"nq_q4_1000000_qwen3_8b","evidence_type":"migration_curve","source":{"canonical_model_id":"Qwen/Qwen3-Embedding-4B","revision":"5cf2132abc99cad020ac570b19d031efec650f2b","dimension":2560,"max_length":8192,"pooling":"last_non_padding_token","padding_side":"left","truncation_side":"right","query_instruction":"Instruct: Given a web search query, retrieve relevant passages that answer the query\nQuery:{text}","document_instruction":"","normalization":"l2","dtype":"float32","contract_fingerprint":"0607b87cc200f80bfc498e595891fb533f42677355b18279be42d7d988896c44"},"target":{"canonical_model_id":"Qwen/Qwen3-Embedding-8B","revision":"1d8ad4ca9b3dd8059ad90a75d4983776a23d44af","dimension":4096,"max_length":8192,"pooling":"last_non_padding_token","padding_side":"left","truncation_side":"right","query_instruction":"Instruct: Given a web search query, retrieve relevant passages that answer the query\nQuery:{text}","document_instruction":"","normalization":"l2","dtype":"float32","contract_fingerprint":"11d6d071a07a73b0bb3588a49447bb8bf58cb5426af732b534d374581ac4b018"},"dataset":{"name":"Natural Questions (BEIR) 1M canonical prefix","canonical_dataset_id":"beir_nq_1m_prefix_1000000","canonical_construction":true,"corpus_size":1000000,"query_count":1000,"split":"scale_study","fingerprint":null},"metric":"nDCG@10","epsilon":0.01,"candidate_depths":[20,50,100,200,500,1000],"candidate_gap":{"20":0.003291235284552691,"50":-0.00043030085527543164,"100":0.0001045284033624751,"200":0.0002040233121074877,"500":0.00028778163515584223,"1000":0.00043067655807339306},"containment":{"20":0.8692,"50":0.9539,"100":0.9755,"200":0.988,"500":0.9951,"1000":0.9979},"native_target_quality":0.6811742213916842,"source_quality":0.661377543285374,"restricted_target_quality":{"20":0.6778829861071315,"50":0.6816045222469596,"100":0.6810696929883215,"200":0.6809701980795767,"500":0.6808864397565283,"1000":0.6807435448336107},"observed_migration_depth":20,"ci_certified_migration_depth":20,"t2":null,"ann":{"status":"MEASURED_REFERENCE_COMPARISON","metric":"source Recall@500 against exact source top-500","index_type":"ivf_flat","nprobe":64,"recall_at_500":0.92194,"g_ann_at_50":0.004046421654698915,"scope":"NQ canonical-prefix scale study"},"provenance":{"artifact":"remote_scale_experiment/runs/20260824-194342/results/scale_curves.csv","artifact_sha256":"a8eea1b1c7ae0d3ac4d23ca5e806704c8e8f2bd932e5765002ba250377fc449c","table":"scale_curves.csv","status":"retained_external","git_commit":null},"notes":["Query-weighted measurements from retained research artifacts; not universal claims about this model pair."]}

embedflow/data/registry/registry_manifest.json

@@ -0,0 +1,16 @@
+{
+  "registry_version": "0.1.0",
+  "schema_version": "1",
+  "record_count": 15,
+  "profile_count": 3,
+  "summary_count": 2,
+  "namespace": "core",
+  "trust": "verified retained research artifacts",
+  "files": {
+    "schema_version.json": {"sha256": "c3d851bcfe66d9cc1aabc4ec8b61512110d8802ac0bcc7af3239a9ac6cec14a9"},
+    "migrations.jsonl": {"sha256": "1471e7a937d87eeba58d19a801fb8c147f6982d2c1522b5da43f6986b92800d5"},
+    "research_summaries.json": {"sha256": "fcbf2fa0456fc3fc2e30d1b945b05e7d7dcfa11b7187910ced8a3fa0e69f6720"},
+    "benchmark_profiles.jsonl": {"sha256": "4ef2bdfc9f1803ef108400986be6a06915879d302e14a2757fc091c32b5b6e54"}
+  },
+  "provenance_policy": "Every core row names a retained artifact and SHA-256; external artifact paths are not required at install time."
+}

embedflow/data/registry/research_summaries.json

@@ -0,0 +1,55 @@
+{
+  "registry_version": "0.1.0",
+  "summaries": [
+    {
+      "summary_id": "development_63_cell_audit",
+      "kind": "development_sweep_summary",
+      "cells": 63,
+      "point_estimate_g50_le_0_01": 42,
+      "ci_certified_cells": 26,
+      "excluding_nanoquora_and_nanomsmarco": {
+        "cells": 45,
+        "point_estimate_g50_le_0_01": 25,
+        "ci_certified_cells": 11
+      },
+      "provenance": {
+        "artifact": "iclr_candidate_compatibility/results/DEVELOPMENT_63_FINAL.csv",
+        "artifact_sha256": "af614e572c811c3905ad1071548bbb53d01e1079eff6232c5cf2ec9bdf294560",
+        "supporting_artifact": "iclr_candidate_compatibility/ICLR_Embedding_Vector_Migration_backup_20260823_precleanup/paper_appendix_artifacts/data/full_63_cell_appendix_table.csv",
+        "supporting_artifact_sha256": "005d5957f4827e8cb133e8d101e3a0a95ecdf6f753af7c77d97e7c5a86b3ddc5",
+        "table": "APPENDIX_BUILD_REPORT.md",
+        "status": "retained_external",
+        "git_commit": null
+      },
+      "notes": [
+        "Aggregate development evidence, not a model-pair guarantee.",
+        "The 63 cells include the retained development matrix; frozen holdout counts are reported separately."
+      ]
+    },
+    {
+      "summary_id": "t2_v1_frozen_holdout_audit",
+      "kind": "t2_holdout_summary",
+      "cells": 28,
+      "compatible_cells": 22,
+      "safe_predictions": 17,
+      "observed_false_safe_predictions": 0,
+      "groups": [
+        {"source": "Qwen3-Embedding-0.6B", "target": "Qwen3-Embedding-4B", "cells": 7, "compatible": 6, "safe": 4, "false_safe": 0},
+        {"source": "MiniLM-L6", "target": "Qwen3-Embedding-8B", "cells": 7, "compatible": 3, "safe": 2, "false_safe": 0},
+        {"source": "Qwen3-Embedding-0.6B", "target": "Qwen3-Embedding-8B", "cells": 7, "compatible": 6, "safe": 4, "false_safe": 0},
+        {"source": "Qwen3-Embedding-4B", "target": "Qwen3-Embedding-8B", "cells": 7, "compatible": 7, "safe": 7, "false_safe": 0}
+      ],
+      "provenance": {
+        "artifact": "final_paper_audit/t2_holdout_ledger.csv",
+        "artifact_sha256": "dfeca01f6bd1891bb2e4ba04cd0db1988ff273419f91c617a72d863a9bd92a47",
+        "table": "T2_HOLDOUT_AUDIT.md",
+        "status": "retained_external",
+        "git_commit": null
+      },
+      "notes": [
+        "Frozen leakage-safe holdout reconstruction; SAFE is an empirical diagnostic, not a guarantee.",
+        "False-safe is an observed holdout count, not a universal error bound."
+      ]
+    }
+  ]
+}

embedflow/data/registry/schema_version.json

@@ -0,0 +1,5 @@
+{
+  "schema_version": "1",
+  "registry_version": "0.1.0",
+  "description": "EmbedFlow core migration evidence schema"
+}

embedflow/frozen/T2_V1_FROZEN_SPEC.md

@@ -0,0 +1,71 @@
+# T2-v1 frozen specification for Qwen3-8B target-shift validation
+
+**Frozen before any Qwen3-8B target-shift feature/result inspection:**
+2026-08-23
+
+This is an exact copy of the previously frozen representative T2 rule.  It
+must not be altered during this experiment.
+
+## Probe-only inputs and normalization
+
+The probe may use source-ranked candidate IDs, source ranks and scores, exact
+target scores for documents in the finite source candidate pool, and stable
+target rerankings of source prefixes.  Embeddings are treated as the existing
+normalized vectors and candidate scores use the audited target/source dot
+product convention in the saved candidate tables.  No cross-model score
+calibration is performed inside the probe.
+
+The finite reference is `R_500`: target reranking restricted to the source
+top-500 candidates.  `R_500` is not native target retrieval.
+
+The probe must not use qrels, native full-corpus target rankings, target nDCG,
+candidate gaps, confidence intervals, or any retrospective compatibility
+label.
+
+## Frozen K values
+
+`K = {10, 20, 50, 100, 200, 500}`.  T2-v1 is evaluated at K=50 and uses
+the source-top-500 finite reference.
+
+## Frozen feature definitions
+
+For each query and source prefix:
+
+* `probe_residual_tail(K) = 1 - |Top10(R_K) ∩ Top10(R_500)| / 10`.
+* `stability_to_500(K) = |Top10(R_K) ∩ Top10(R_500)| / 10`.
+* `deepest_source_rank_top10` is the maximum original source rank in
+  `Top10(R_500)`; `deepest_p90` is its query-level 90th percentile.
+* For each adjacent source-shell expansion `K→K'`, `p_any_entrant` is the
+  fraction of queries in which at least one newly exposed document enters
+  `Top10(R_K')`.
+* `late_tail_area` is the prior finite-shell weighted summary of entrant
+  probabilities over expansions 50→100, 100→200, and 200→500.
+* `fraction_margin_nonpositive` and `last_shell_any_rate` are computed with
+  the prior implementation and are used only by the fixed high-risk guard.
+
+## Exact T2-v1 decision logic
+
+At K=50, output `SAFE` iff all of the following are true:
+
+1. mean `probe_residual_tail(50) <= 0.05`;
+2. `deepest_p90 <= 200`;
+3. `late_tail_area <= 0.20`.
+
+If SAFE is false, output `UNSAFE_OR_UNCERTAIN` iff at least one of these
+fixed guards is true:
+
+* mean `probe_residual_tail(50) > 0.10`;
+* mean `stability_to_500(50) < 0.90`;
+* `last_shell_any_rate > 0.25`;
+* `fraction_margin_nonpositive(50) > 0.50`.
+
+Otherwise output `EXPAND`.
+
+## Evaluation-only labels
+
+Only after blind predictions are persisted may the evaluation join the
+authoritative `G(50) = native target nDCG@10 − exact target rerank nDCG@10
+within source top-50` labels.  Practical compatibility is `G(50) <= 0.01`
+using unrounded values.  CI-NI fields are secondary evaluation labels.
+
+**T2-v1 was frozen before any Qwen3-8B target-shift results were inspected.**

embedflow/frozen/T2_V1_FROZEN_SPEC.sha256

@@ -0,0 +1 @@
+6ea348373d5ff257ac0fa4ce08207a84c647fb567012f2d64591fb057428e09a

embedflow/indexes/__init__.py

@@ -0,0 +1,5 @@
+from .base import SearchHit, VectorIndex
+from .faiss_backend import FaissIndex, NumpyIndex, load_faiss_index
+from .qdrant_backend import QdrantIndex
+
+__all__ = ["SearchHit", "VectorIndex", "FaissIndex", "NumpyIndex", "load_faiss_index", "QdrantIndex"]

embedflow/indexes/base.py

@@ -0,0 +1,60 @@
+from __future__ import annotations
+
+from abc import ABC, abstractmethod
+from dataclasses import dataclass
+from typing import Any
+
+import numpy as np
+
+
+@dataclass(frozen=True)
+class SearchHit:
+    document_id: str
+    score: float
+    source_rank: int
+
+
+class VectorIndex(ABC):
+    dimension: int
+    metric: str
+
+    @abstractmethod
+    def search(self, query_vector: np.ndarray, k: int) -> list[SearchHit]:
+        raise NotImplementedError
+
+    @abstractmethod
+    def fetch_documents(self, ids: list[str]) -> dict[str, Any]:
+        raise NotImplementedError
+
+    @abstractmethod
+    def size(self) -> int:
+        raise NotImplementedError
+
+    @abstractmethod
+    def metadata(self) -> dict[str, Any]:
+        raise NotImplementedError
+
+    def close(self) -> None:
+        """Release backend resources.
+
+        In-memory/FAISS indexes do not currently own a live resource, so the
+        default is a no-op. Backends such as local Qdrant use this hook to
+        release file locks and network clients when a serving engine stops.
+        """
+        return None
+
+
+def validate_query_vector(vector: np.ndarray, dimension: int) -> np.ndarray:
+    value = np.asarray(vector, dtype="float32")
+    if value.ndim != 1 or value.shape[0] != int(dimension):
+        raise ValueError(f"query vector dimension {value.shape} does not match index dimension {dimension}")
+    if not np.isfinite(value).all():
+        raise ValueError("query vector contains non-finite values")
+    return value
+
+
+def validate_k(k: int) -> int:
+    """Validate a requested top-k before it reaches a backend API."""
+    if isinstance(k, bool) or int(k) != k or int(k) < 1:
+        raise ValueError("k must be a positive integer")
+    return int(k)

embedflow/indexes/faiss_backend.py

@@ -0,0 +1,240 @@
+from __future__ import annotations
+
+import json
+from pathlib import Path
+from typing import Any
+
+import numpy as np
+
+from .base import SearchHit, VectorIndex, validate_k, validate_query_vector
+
+
+class NumpyIndex(VectorIndex):
+    """Tiny deterministic fallback used by the offline demo when FAISS is absent."""
+
+    def __init__(self, vectors: np.ndarray, ids: list[str], path: str | Path | None = None,
+                 metric: str = "cosine", documents: dict[str, Any] | None = None, metadata: dict[str, Any] | None = None):
+        self.vectors = np.asarray(vectors, dtype="float32"); self.ids = [str(x) for x in ids]; self.path = Path(path) if path else None
+        self.metric, self.documents = str(metric).lower(), documents or {}; self._metadata = dict(metadata or {})
+        if self.vectors.ndim != 2 or self.vectors.shape[0] < 1 or self.vectors.shape[0] != len(self.ids) or self.vectors.shape[1] < 1:
+            raise ValueError("invalid numpy index vectors/IDs: expected a non-empty 2-D matrix aligned with IDs")
+        if len(set(self.ids)) != len(self.ids):
+            raise ValueError("duplicate document IDs are not allowed")
+        if self.metric not in {"cosine", "dot", "inner_product"}:
+            raise ValueError("metric must be cosine, dot, or inner_product")
+        if not np.isfinite(self.vectors).all():
+            raise ValueError("index vectors must be finite")
+        self.dimension = self.vectors.shape[1]
+
+    @classmethod
+    def load(cls, path: str | Path, documents: dict[str, Any] | None = None, metric: str = "cosine") -> NumpyIndex:
+        try:
+            with Path(path).open("rb") as handle:
+                data = np.load(handle, allow_pickle=False)
+                vectors, ids = data["vectors"], data["ids"].tolist()
+            if not isinstance(ids, list):
+                raise ValueError("embedded IDs must be a one-dimensional array")
+        except (OSError, ValueError, KeyError, TypeError, UnicodeError, EOFError) as exc:
+            raise ValueError(f"invalid NumPy index artifact: {path}") from exc
+        meta_path = Path(path).with_suffix(Path(path).suffix + ".meta.json")
+        try:
+            meta = json.loads(meta_path.read_text()) if meta_path.exists() else {}
+            if not isinstance(meta, dict):
+                raise ValueError("metadata must be a JSON object")
+        except (OSError, UnicodeError, json.JSONDecodeError) as exc:
+            raise ValueError(f"invalid NumPy index metadata: {meta_path}") from exc
+        if not isinstance(meta, dict):
+            raise ValueError(f"invalid NumPy index metadata: {meta_path}")
+        return cls(vectors, ids, path, metric, documents, meta)
+
+    @classmethod
+    def build(cls, vectors: np.ndarray, ids: list[str], path: str | Path | None = None,
+              metric: str = "cosine", documents: dict[str, Any] | None = None, metadata: dict[str, Any] | None = None,
+              ids_path: str | Path | None = None) -> NumpyIndex:
+        x = np.asarray(vectors, dtype="float32")
+        metric = str(metric).lower()
+        if metric not in {"cosine", "dot", "inner_product"}:
+            raise ValueError("metric must be cosine, dot, or inner_product")
+        if x.ndim != 2 or x.shape[1] < 1 or x.shape[0] == 0:
+            raise ValueError("vectors must be a non-empty 2-D matrix")
+        if metric == "cosine":
+            norms = np.linalg.norm(x, axis=1, keepdims=True)
+            if np.any(norms <= 1e-12): raise ValueError("cosine vectors must be non-zero")
+            x = x / norms
+        result = cls(x, ids, path, metric, documents=documents, metadata=metadata)
+        if path:
+            target = Path(path); target.parent.mkdir(parents=True, exist_ok=True)
+            with target.open("wb") as handle: np.savez(handle, vectors=x, ids=np.asarray(ids, dtype="U"))
+            id_target = Path(ids_path) if ids_path else target.with_suffix(target.suffix + ".ids.json")
+            id_target.parent.mkdir(parents=True, exist_ok=True)
+            id_target.write_text(json.dumps(ids))
+            if metadata:
+                target.with_suffix(target.suffix + ".meta.json").write_text(json.dumps(metadata, indent=2, sort_keys=True))
+        return result
+
+    def search(self, query_vector: np.ndarray, k: int) -> list[SearchHit]:
+        k = validate_k(k)
+        q = validate_query_vector(query_vector, self.dimension)
+        if self.metric == "cosine":
+            norm = float(np.linalg.norm(q))
+            if norm <= 1e-12:
+                raise ValueError("cosine query vector must be non-zero")
+            q = q / norm
+        scores = self.vectors @ q; order = np.lexsort((np.arange(len(scores)), -scores))[:min(k, len(self.ids))]
+        return [SearchHit(self.ids[int(i)], float(scores[i]), rank) for rank, i in enumerate(order)]
+
+    def fetch_documents(self, ids: list[str]) -> dict[str, Any]: return {str(i): self.documents[str(i)] for i in ids if str(i) in self.documents}
+    def size(self) -> int: return len(self.ids)
+    def metadata(self) -> dict[str, Any]: return {"backend": "numpy_fallback", "dimension": self.dimension, "metric": self.metric, "size": self.size(), **self._metadata}
+    def close(self) -> None: return None
+
+
+class FaissIndex(VectorIndex):
+    """FAISS wrapper supporting saved indexes and small in-process indexes."""
+
+    def __init__(self, index: Any, ids: list[str], path: str | Path | None = None,
+                 metric: str = "cosine", metadata: dict[str, Any] | None = None,
+                 documents: dict[str, Any] | None = None):
+        self.index = index
+        self.ids = [str(x) for x in ids]
+        self.path = Path(path) if path else None
+        self.dimension = int(index.d)
+        self.metric = metric
+        self._metadata = dict(metadata or {})
+        self.documents = documents or {}
+        if int(index.ntotal) != len(self.ids):
+            raise ValueError(f"FAISS index contains {index.ntotal} rows but {len(self.ids)} IDs were supplied")
+        if len(set(self.ids)) != len(self.ids):
+            raise ValueError("duplicate document IDs are not allowed")
+        if self.dimension < 1:
+            raise ValueError("FAISS index dimension must be positive")
+        if str(metric).lower() not in {"cosine", "dot", "inner_product"}:
+            raise ValueError("metric must be cosine, dot, or inner_product")
+        self.metric = str(metric).lower()
+        self._metadata.update({"backend": "faiss", "dimension": self.dimension, "metric": self.metric,
+                               "size": len(self.ids), "index_type": type(index).__name__})
+
+    @classmethod
+    def load(cls, path: str | Path, ids_path: str | Path | None = None,
+             metric: str = "cosine", documents: dict[str, Any] | None = None,
+             nprobe: int | None = None) -> FaissIndex:
+        try:
+            import faiss
+        except ImportError as exc:
+            raise RuntimeError("FAISS backend requires faiss-cpu or faiss-gpu") from exc
+        path = Path(path)
+        if not path.exists():
+            raise FileNotFoundError(path)
+        try:
+            index = faiss.read_index(str(path))
+        except (OSError, RuntimeError, ValueError) as exc:
+            raise ValueError(f"invalid FAISS index artifact: {path}") from exc
+        candidates = [Path(ids_path)] if ids_path else []
+        candidates += [Path(str(path) + ".ids.json"), path.with_suffix(path.suffix + ".ids.json"), path.with_suffix(".ids.json")]
+        id_file = next((p for p in candidates if p.exists()), None)
+        if id_file is None:
+            raise FileNotFoundError(f"missing FAISS document ID sidecar; tried {[str(x) for x in candidates]}")
+        if nprobe is not None and (isinstance(nprobe, bool) or int(nprobe) != nprobe or int(nprobe) < 1):
+            raise ValueError("nprobe must be a positive integer")
+        if nprobe is not None and hasattr(index, "nprobe"):
+            index.nprobe = min(max(1, int(nprobe)), int(getattr(index, "nlist", nprobe)))
+        try:
+            raw = json.loads(id_file.read_text())
+        except (OSError, UnicodeError, json.JSONDecodeError) as exc:
+            raise ValueError(f"invalid FAISS IDs sidecar: {id_file}") from exc
+        ids = raw.get("ids", raw) if isinstance(raw, dict) else raw
+        if not isinstance(ids, list):
+            raise ValueError(f"invalid FAISS IDs sidecar: {id_file}")
+        meta_file = path.with_suffix(path.suffix + ".meta.json")
+        try:
+            meta = json.loads(meta_file.read_text()) if meta_file.exists() else {}
+        except (OSError, UnicodeError, json.JSONDecodeError) as exc:
+            raise ValueError(f"invalid FAISS metadata: {meta_file}") from exc
+        if not isinstance(meta, dict):
+            raise ValueError(f"invalid FAISS metadata: {meta_file}")
+        return cls(index, [str(x) for x in ids], path=path, metric=metric, metadata=meta, documents=documents)
+
+    @classmethod
+    def build(cls, vectors: np.ndarray, ids: list[str], path: str | Path | None = None,
+              metric: str = "cosine", nlist: int | None = None, nprobe: int = 64,
+              documents: dict[str, Any] | None = None, metadata: dict[str, Any] | None = None,
+              ids_path: str | Path | None = None) -> FaissIndex:
+        try:
+            import faiss
+        except ImportError as exc:
+            raise RuntimeError("FAISS backend requires faiss-cpu or faiss-gpu") from exc
+        x = np.asarray(vectors, dtype="float32")
+        metric = str(metric).lower()
+        if metric not in {"cosine", "dot", "inner_product"}:
+            raise ValueError("metric must be cosine, dot, or inner_product")
+        if isinstance(nprobe, bool) or int(nprobe) != nprobe or int(nprobe) < 1:
+            raise ValueError("nprobe must be a positive integer")
+        if nlist is not None and (isinstance(nlist, bool) or int(nlist) != nlist or int(nlist) < 1):
+            raise ValueError("nlist must be a positive integer when provided")
+        if x.ndim != 2 or x.shape[0] == 0 or x.shape[1] < 1 or x.shape[0] != len(ids) or not np.isfinite(x).all():
+            raise ValueError("vectors must be finite 2-D data aligned with unique IDs")
+        if len(set(map(str, ids))) != len(ids):
+            raise ValueError("duplicate document IDs are not allowed")
+        if metric.lower() == "cosine":
+            norms = np.linalg.norm(x, axis=1, keepdims=True)
+            if np.any(norms <= 1e-12): raise ValueError("cosine vectors must be non-zero")
+            x = x / norms
+        d = x.shape[1]
+        if nlist and len(x) >= max(39, nlist * 4):
+            quantizer = faiss.IndexFlatIP(d)
+            index = faiss.IndexIVFFlat(quantizer, d, min(int(nlist), len(x)), faiss.METRIC_INNER_PRODUCT)
+            index.train(x)
+            index.nprobe = min(int(nprobe), int(index.nlist))
+        else:
+            index = faiss.IndexFlatIP(d)
+        index.add(x)
+        result = cls(index, [str(x) for x in ids], path=path, metric=metric,
+                     metadata={"nprobe": nprobe, "nlist": nlist, **(metadata or {})}, documents=documents)
+        if path:
+            result.save(path, ids_path=ids_path)
+        return result
+
+    def save(self, path: str | Path | None = None, ids_path: str | Path | None = None) -> None:
+        try:
+            import faiss
+        except ImportError as exc:
+            raise RuntimeError("FAISS backend requires faiss-cpu or faiss-gpu") from exc
+        target = Path(path or self.path or "legacy.index")
+        target.parent.mkdir(parents=True, exist_ok=True)
+        faiss.write_index(self.index, str(target))
+        id_target = Path(ids_path) if ids_path else target.with_suffix(target.suffix + ".ids.json")
+        id_target.parent.mkdir(parents=True, exist_ok=True)
+        id_target.write_text(json.dumps(self.ids, ensure_ascii=False))
+        target.with_suffix(target.suffix + ".meta.json").write_text(json.dumps(self.metadata(), indent=2, sort_keys=True))
+        self.path = target
+
+    def search(self, query_vector: np.ndarray, k: int) -> list[SearchHit]:
+        k = validate_k(k)
+        value = validate_query_vector(query_vector, self.dimension)
+        if self.metric == "cosine":
+            norm = float(np.linalg.norm(value))
+            if norm <= 1e-12:
+                raise ValueError("cosine query vector must be non-zero")
+            value = value / norm
+        value = value[None, :]
+        scores, indices = self.index.search(value, min(k, len(self.ids)))
+        return [SearchHit(self.ids[int(i)], float(score), rank) for rank, (score, i) in enumerate(zip(scores[0], indices[0])) if int(i) >= 0]
+
+    def fetch_documents(self, ids: list[str]) -> dict[str, Any]:
+        return {str(i): self.documents[str(i)] for i in ids if str(i) in self.documents}
+
+    def size(self) -> int:
+        return len(self.ids)
+
+    def metadata(self) -> dict[str, Any]:
+        return dict(self._metadata)
+
+    def close(self) -> None:
+        # FAISS index objects are in-memory Python wrappers; there is no live
+        # client to close. Keep the hook symmetric with other backends.
+        return None
+
+
+def load_faiss_index(path: str | Path, documents: dict[str, Any] | None = None,
+                     metric: str = "cosine", nprobe: int | None = None) -> FaissIndex:
+    return FaissIndex.load(path, metric=metric, documents=documents, nprobe=nprobe)