@furongjun1999/dsh-memory 0.4.11 → 0.5.1
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- package/README.md +552 -465
- package/codebuddy/CODEBUDDY.md +11 -3
- package/codebuddy/README.md +92 -90
- package/codebuddy/mcp.json +27 -27
- package/docs/GBrain/345/217/257/345/200/237/351/211/264/347/202/271_/347/201/265/346/236/242/350/220/275/347/202/271/344/272/244/346/216/245_20260919.md +169 -169
- package/docs/Pi/345/217/257/345/255/246/344/271/240/344/274/230/347/202/271_/347/201/265/346/236/242/345/244/247/350/204/221/346/224/271/350/277/233/344/272/244/346/216/245_20260915.md +144 -144
- package/docs/README.md +143 -111
- package/docs/discipline/harnesses.yaml +244 -226
- package/docs/discipline/templates/full.md.tmpl +61 -61
- package/docs/discipline/templates/rules.mdc.tmpl +68 -0
- package/docs/discipline/templates/skill.md.tmpl +23 -23
- package/docs/eval/AGI/344/270/203/347/273/264/350/257/204/345/210/206/346/212/245/345/221/212_md_cg_v1.0.md +299 -299
- package/docs/eval/AGI/344/270/203/347/273/264/350/257/204/345/210/206/346/212/245/345/221/212_md_cg_v2.0.md +239 -239
- package/docs/eval/bench_lingshu_self/bench_self.py +140 -0
- package/docs/eval/bench_lingshu_self/self_bench_result.json +404 -0
- package/docs/eval/bench_lingshu_self//347/201/265/346/236/242/350/207/252/345/272/223/347/253/257/345/210/260/347/253/257/346/243/200/347/264/242/345/256/236/346/265/213_v1.0.md +38 -0
- package/docs/eval//344/270/215/345/217/257/351/235/240/346/200/247/350/220/275/345/234/260_P0_v1.0.md +185 -0
- package/docs/eval//345/256/236/351/252/214/346/226/271/346/241/210_/345/255/246/344/271/240/351/227/255/347/216/257AB/344/270/216/346/250/252/350/257/204_v1.0.md +174 -174
- package/docs/eval//346/225/205/351/232/234/346/263/250/345/205/245/345/256/236/346/265/213_v1.0.md +422 -0
- package/docs/eval//346/250/252/350/257/204_/345/205/255/345/256/266100/351/242/230/344/270/255/350/213/261/345/217/214/346/237/245_v1.0.md +223 -223
- package/docs/eval//347/253/257/345/210/260/347/253/257LoCoMoQA/345/220/214/345/217/243/345/276/204/345/257/271/347/205/247_v1.0.md +100 -0
- package/docs/eval//347/253/257/345/210/260/347/253/257/345/271/262/346/211/260/346/261/240/350/257/204/346/265/213_/347/241/256/345/256/232/346/200/247/350/243/201/345/206/263vsLLM_judge_v1.1.md +197 -0
- package/docs/eval//347/274/272/351/231/267/346/214/226/346/216/230_/350/207/252/344/270/273/350/277/255/344/273/243_v1.md +156 -0
- package/docs/eval//347/274/272/351/231/267/346/214/226/346/216/230_/350/207/252/344/270/273/350/277/255/344/273/243_v10.md +210 -0
- package/docs/eval//347/274/272/351/231/267/346/214/226/346/216/230_/350/207/252/344/270/273/350/277/255/344/273/243_v11.md +227 -0
- package/docs/eval//347/274/272/351/231/267/346/214/226/346/216/230_/350/207/252/344/270/273/350/277/255/344/273/243_v12.md +203 -0
- package/docs/eval//347/274/272/351/231/267/346/214/226/346/216/230_/350/207/252/344/270/273/350/277/255/344/273/243_v13.md +233 -0
- package/docs/eval//347/274/272/351/231/267/346/214/226/346/216/230_/350/207/252/344/270/273/350/277/255/344/273/243_v14.md +191 -0
- package/docs/eval//347/274/272/351/231/267/346/214/226/346/216/230_/350/207/252/344/270/273/350/277/255/344/273/243_v15.md +213 -0
- package/docs/eval//347/274/272/351/231/267/346/214/226/346/216/230_/350/207/252/344/270/273/350/277/255/344/273/243_v16.md +214 -0
- package/docs/eval//347/274/272/351/231/267/346/214/226/346/216/230_/350/207/252/344/270/273/350/277/255/344/273/243_v17.md +199 -0
- package/docs/eval//347/274/272/351/231/267/346/214/226/346/216/230_/350/207/252/344/270/273/350/277/255/344/273/243_v2.md +156 -0
- package/docs/eval//347/274/272/351/231/267/346/214/226/346/216/230_/350/207/252/344/270/273/350/277/255/344/273/243_v3.md +152 -0
- package/docs/eval//347/274/272/351/231/267/346/214/226/346/216/230_/350/207/252/344/270/273/350/277/255/344/273/243_v4.md +128 -0
- package/docs/eval//347/274/272/351/231/267/346/214/226/346/216/230_/350/207/252/344/270/273/350/277/255/344/273/243_v5.md +114 -0
- package/docs/eval//347/274/272/351/231/267/346/214/226/346/216/230_/350/207/252/344/270/273/350/277/255/344/273/243_v6.md +192 -0
- package/docs/eval//347/274/272/351/231/267/346/214/226/346/216/230_/350/207/252/344/270/273/350/277/255/344/273/243_v7.md +187 -0
- package/docs/eval//347/274/272/351/231/267/346/214/226/346/216/230_/350/207/252/344/270/273/350/277/255/344/273/243_v8.md +207 -0
- package/docs/eval//347/274/272/351/231/267/346/214/226/346/216/230_/350/207/252/344/270/273/350/277/255/344/273/243_v9.md +203 -0
- package/docs/{mdcg → hive}//344/273/244/347/211/214/344/270/216/350/247/222/350/211/262/346/235/203/350/201/214/345/210/206/347/246/273_v0.1.md +165 -160
- package/docs/hive//344/273/244/347/211/214/350/257/255/344/271/211/344/277/256/346/255/243_/344/273/262/350/243/201/344/275/215_v0.1.md +76 -0
- package/docs/{mdcg → hive}//345/255/220/344/273/243/347/220/206/351/205/215/347/275/256/346/240/207/345/207/206_v0.5.md +236 -236
- package/docs/hive//345/256/211/345/205/250/345/256/241/350/256/241/345/256/236/351/224/232_v0.1.md +202 -0
- package/docs/hive//346/243/200/347/264/242/346/224/266/346/225/233/345/256/236/346/265/213/344/270/216S1b/350/256/276/350/256/241_v0.1.md +43 -43
- package/docs/hive//346/243/200/347/264/242/347/256/227/346/263/225/345/217/243/345/276/204/345/257/271/347/205/247_v0.1.md +85 -0
- package/docs/hive//346/243/200/347/264/242/350/267/257/345/276/204/344/270/216/350/256/244/347/237/245/347/273/223/346/236/204/345/245/221/347/272/246_v0.1.md +102 -102
- package/docs/hive//350/234/202/345/267/242M6_ingest/345/256/236/346/226/275/350/256/241/345/210/222_v0.1.md +47 -0
- package/docs/hive//350/234/202/345/267/242/345/217/214/345/256/236/344/276/213/344/272/222/351/252/214_/350/256/276/350/256/241/345/256/232/347/250/277.md +519 -503
- package/docs/hive//350/234/202/345/267/242/345/267/245/344/275/234/350/256/260/345/277/206_/351/241/271/347/233/256/350/256/241/345/210/222.md +85 -85
- package/docs/hive//350/234/202/345/267/242/350/256/276/350/256/241_/347/220/206/350/256/272/345/257/271/351/275/220_v0.1.md +191 -0
- package/docs/hive//350/234/202/345/267/242/350/277/255/344/273/243_/345/256/217/350/247/202/344/270/216/347/276/244/344/275/223/350/260/203/345/272/246_v0.1.md +90 -0
- package/docs/images/lingshu-moonlight-covenant-poster-preview.jpg +0 -0
- package/docs/images/lingshu-moonlight-covenant-poster.png +0 -0
- package/docs/mdcg/D_meta_/345/267/245/347/250/213/345/214/226/346/226/271/346/241/210_v0.2.md +216 -216
- package/docs/mdcg/README/350/257/246/347/273/206/347/211/210_v0.4.10.md +646 -646
- package/docs/mdcg/lingshu_tutorial.html +14449 -14449
- package/docs/mdcg/release_v0.4.11.md +49 -0
- package/docs/mdcg/release_v0.4.5.md +55 -55
- package/docs/mdcg/tool_table_v0.3.0.md +117 -117
- package/docs/mdcg//345/205/250/345/272/223/344/273/243/347/240/201/350/257/204/345/256/241/344/270/216/346/235/241/344/273/266/345/214/226/346/263/250/351/207/212_/350/256/241/345/210/222_v0.1.md +600 -600
- package/docs/mdcg//345/212/237/350/203/275/350/260/203/347/224/250/346/230/240/345/260/204/350/241/250_v0.1.md +40 -40
- package/docs/mdcg//345/215/225/345/205/203/350/207/252/346/210/221/351/224/232/347/202/271_/347/263/273/347/273/237/346/217/220/347/244/272/350/257/215/346/240/207/345/207/206_v0.3.md +275 -275
- package/docs/mdcg//345/217/221/345/270/203/351/227/250/347/246/201/351/223/276_v0.1.md +54 -0
- package/docs/mdcg//346/272/220/347/240/201/347/272/247/346/236/266/346/236/204/345/256/241/350/256/241_GPT/346/211/271/350/257/204/345/257/271/347/205/247_v1.0.md +130 -130
- package/docs/mdcg//347/201/265/346/236/24282/345/267/245/345/205/267_/345/212/237/350/203/275/346/225/264/347/220/206/344/270/216/350/277/201/347/247/273/346/230/240/345/260/204_v0.1.md +278 -278
- package/docs/mdcg//347/201/265/346/236/242/350/256/260/345/277/206/345/212/250/350/257/215/345/215/217/350/256/256_v1.0-draft.md +172 -172
- package/docs/mdcg//347/274/272/345/217/243/345/215/225_P0/346/224/266/345/217/243_v0.1.md +270 -270
- package/docs/mdcg//350/256/244/347/237/245/345/233/276_G4-G8/347/274/272/345/217/243/350/243/201/345/256/232/345/215/225_v0.1.md +491 -491
- package/docs/mdcg//350/256/244/347/237/245/345/233/276_/346/235/241/344/273/266/347/251/272/351/227/264/345/220/210/346/210/220/344/270/216/347/224/237/346/225/210/346/235/241/344/273/266/345/217/243/345/276/204_v0.1.md +340 -340
- package/docs/mdcg//350/256/244/347/237/245/345/233/276_/347/264/242/345/274/225/344/270/216/345/267/245/347/250/213/350/247/204/350/214/203/345/214/226_/350/256/241/345/210/222_v0.1.md +588 -588
- package/docs/plans/GridWorld/346/234/200/345/260/217/351/227/255/347/216/257/350/247/204/346/240/274_v0.1.md +176 -176
- package/docs/plans//345/221/275/345/220/215/346/262/273/347/220/206_/351/241/271/347/233/256/350/256/241/345/210/222.md +81 -81
- package/docs/swarm//350/234/202/347/276/244/344/272/222/350/201/224_v0.1.md +704 -704
- package/docs/swarm//350/234/202/347/276/244/345/220/214/351/224/231/346/243/200/346/265/213/345/256/236/351/252/214/345/215/217/350/256/256_v0.1.md +242 -242
- package/docs/theory//344/270/215/345/217/257/351/235/240/345/256/232/347/220/206/344/270/216/345/244/261/346/225/210/344/274/230/345/205/210/346/241/206/346/236/266_v0.3.md +365 -0
- package/docs/theory//345/215/225/347/272/277/347/250/213/344/270/216/346/263/250/346/204/217/345/212/233/351/233/206/344/270/255_/346/227/240/344/272/211/350/256/256/347/220/206/350/256/272/346/226/207/346/241/243_v1.0.md +129 -129
- package/docs/theory//345/271/266/345/217/221/345/277/205/347/204/266/346/200/247/347/220/206/350/256/272_v0.2.md +134 -134
- package/docs/theory//346/231/272/350/203/275/350/256/2723.4.md +5260 -5260
- package/docs/theory//346/246/202/345/277/265/345/210/206/345/261/202/345/257/271/351/275/220/350/241/250_v0.1.md +145 -145
- package/docs/theory//347/220/206/350/256/272_/346/234/272/345/210/266_/344/273/243/347/240/201_/345/256/236/351/252/214_/347/274/272/345/217/243/347/237/251/351/230/265_v0.1.md +144 -144
- package/docs/theory//347/220/206/350/256/272/344/273/223/346/213/206/345/210/206/344/270/216/347/231/275/347/256/261/347/237/245/350/257/206/345/272/223/345/206/205/350/277/201_v0.1.md +198 -198
- package/docs/theory//350/256/244/347/237/245/344/273/243/347/220/206/344/270/216/346/224/266/346/225/233/347/273/223/346/236/204_/346/235/241/344/273/266/350/256/272/351/207/215/346/236/204_v0.1.md +221 -221
- package/docs/world_model//344/270/226/347/225/214/346/250/241/345/236/213_/347/245/236/347/273/217/347/275/221/347/273/234/345/272/225/345/261/202/346/236/266/346/236/204_v1.0.md +237 -237
- package/docs//344/270/215/345/217/257/351/235/240/346/200/247/347/220/206/350/256/272_v0.1.md +343 -0
- package/docs//345/217/221/345/270/203/344/273/266/345/233/236/346/272/257/350/257/264/346/230/216_v0.1.md +82 -82
- package/docs//345/267/245/344/275/234/347/272/252/345/276/213_/350/256/244/347/237/245/345/233/276/346/235/241/347/233/256_v1.1.json +28 -2
- package/dsh/README.md +82 -82
- package/dsh/cordis.yml.example +139 -139
- package/dsh/update-lingshu.bat +11 -11
- package/lib/bridge.d.ts +9 -0
- package/lib/bridge.js +35 -0
- package/lib/hooks.js +36 -2
- package/lib/index.js +7 -1
- package/lib/lib/roleplay_web.js +116 -29
- package/lib/lib/token_store.d.ts +7 -1
- package/lib/lib/token_store.js +12 -3
- package/md_cg/__init__.py +7 -7
- package/md_cg/audit.py +379 -368
- package/md_cg/autonomy.py +287 -287
- package/md_cg/backfill.py +1328 -1327
- package/md_cg/backfill_bigdomain.py +34 -34
- package/md_cg/backfill_bucket_zh.py +35 -0
- package/md_cg/bench6_arms.py +410 -410
- package/md_cg/bench6_common.py +230 -230
- package/md_cg/bench6_competitors.py +212 -212
- package/md_cg/bench_axis_domain.py +257 -257
- package/md_cg/bench_blind_comp.py +308 -308
- package/md_cg/bench_e2e_judge.py +532 -0
- package/md_cg/bench_e2e_locomo_qa.py +368 -0
- package/md_cg/bench_e2e_qa.py +256 -0
- package/md_cg/bench_en_atoms_public.py +230 -230
- package/md_cg/bench_governance.py +348 -348
- package/md_cg/bench_lme_zh.py +410 -410
- package/md_cg/bench_locomo.py +121 -121
- package/md_cg/bench_locomo_zh.py +450 -450
- package/md_cg/bench_locomo_zh_public.py +147 -147
- package/md_cg/bench_longmem.py +112 -112
- package/md_cg/bench_membench.py +632 -632
- package/md_cg/bench_p0.py +149 -149
- package/md_cg/bench_progressive.py +287 -287
- package/md_cg/bench_role_views.py +238 -238
- package/md_cg/bench_task_ab.py +243 -243
- package/md_cg/bench_task_ab_llm.py +408 -408
- package/md_cg/bench_unified_en.py +204 -204
- package/md_cg/bench_zh_mad.py +601 -601
- package/md_cg/blindspot_tickets.py +123 -123
- package/md_cg/branches.py +301 -285
- package/md_cg/build_postings.py +73 -73
- package/md_cg/ccgc.py +1006 -948
- package/md_cg/census.py +132 -132
- package/md_cg/chain.py +315 -300
- package/md_cg/codeindex.py +531 -531
- package/md_cg/coldverify.py +292 -292
- package/md_cg/comment_gate.py +337 -337
- package/md_cg/cond_compose.py +190 -190
- package/md_cg/cond_facts.py +154 -154
- package/md_cg/cond_template.json +106 -106
- package/md_cg/condition_anchor.py +142 -142
- package/md_cg/conformance.py +726 -726
- package/md_cg/consistency.py +717 -717
- package/md_cg/consolidate.py +1537 -1439
- package/md_cg/corpus.py +110 -110
- package/md_cg/crosscheck.py +1098 -1097
- package/md_cg/crypto.py +3 -1
- package/md_cg/d_meta.py +310 -310
- package/md_cg/datapath.py +78 -18
- package/md_cg/docindex.py +473 -473
- package/md_cg/eval_common.py +575 -575
- package/md_cg/evidence.py +4 -2
- package/md_cg/evolution.py +477 -477
- package/md_cg/export.py +222 -220
- package/md_cg/forgetting.py +581 -581
- package/md_cg/fsutil.py +377 -329
- package/md_cg/hotcache.py +48 -7
- package/md_cg/hyperedge.py +251 -251
- package/md_cg/identity.py +390 -390
- package/md_cg/insight.py +500 -500
- package/md_cg/interop.py +338 -0
- package/md_cg/judgment_manifest.py +177 -0
- package/md_cg/lexicon/build_cedict_en_zh.py +329 -329
- package/md_cg/lexicon/build_standard_en.py +171 -171
- package/md_cg/lexicon/expand_en_zh.py +211 -211
- package/md_cg/lifecycle.py +272 -272
- package/md_cg/linkref.py +280 -280
- package/md_cg/links.py +140 -107
- package/md_cg/mcp_server.py +403 -55
- package/md_cg/md_whitebox.py +345 -345
- package/md_cg/mdcg.py +783 -233
- package/md_cg/mdcos.py +500 -70
- package/md_cg/metacognition.py +591 -591
- package/md_cg/migrate.py +119 -119
- package/md_cg/migrate_aeis.py +221 -221
- package/md_cg/migrate_roleplay.py +293 -293
- package/md_cg/migrate_wisdom_graph.py +360 -360
- package/md_cg/mreview/__init__.py +25 -25
- package/md_cg/mreview/__main__.py +110 -110
- package/md_cg/mreview/bundle.py +178 -178
- package/md_cg/mreview/candidates.py +262 -262
- package/md_cg/mreview/govern.py +694 -693
- package/md_cg/mreview/locate.py +939 -939
- package/md_cg/mreview/pipeline.py +728 -728
- package/md_cg/mreview/rules/duplication.json +21 -21
- package/md_cg/mreview/rules/field_coverage.json +54 -54
- package/md_cg/mreview/rules/source_license.json +21 -21
- package/md_cg/mreview/rules/template_flow.json +21 -21
- package/md_cg/mreview/ruleset.py +252 -252
- package/md_cg/nodefile.py +575 -575
- package/md_cg/pooling.py +484 -472
- package/md_cg/postings.py +300 -298
- package/md_cg/predict.py +1100 -1100
- package/md_cg/progressive.py +123 -123
- package/md_cg/protect.py +272 -272
- package/md_cg/protocol/md_cg_gate.proto +33 -33
- package/md_cg/protocol.py +372 -372
- package/md_cg/provenance.py +582 -582
- package/md_cg/reach.py +453 -453
- package/md_cg/readcache.py +143 -0
- package/md_cg/reconcile.py +228 -0
- package/md_cg/refindex.py +833 -833
- package/md_cg/refine.py +604 -604
- package/md_cg/review_cli.py +170 -0
- package/md_cg/roleviews.py +89 -89
- package/md_cg/routing.py +393 -365
- package/md_cg/run_tests.py +211 -0
- package/md_cg/scrub.py +13 -3
- package/md_cg/security.py +128 -18
- package/md_cg/self_state.py +1029 -1029
- package/md_cg/selfreport.py +152 -151
- package/md_cg/semantic/__init__.py +10 -10
- package/md_cg/semantic/canonical.py +122 -122
- package/md_cg/semantic/en_normalizer.py +364 -364
- package/md_cg/semantic/en_zh_map.json +28694 -0
- package/md_cg/semantic/export_en_zh_map.py +64 -0
- package/md_cg/semantic/unify.py +45 -0
- package/md_cg/semantic/zh_en_atoms.py +139 -139
- package/md_cg/signer.py +7 -4
- package/md_cg/sources.py +816 -582
- package/md_cg/statushdr.py +179 -179
- package/md_cg/stg.py +54 -37
- package/md_cg/subgraph.py +729 -729
- package/md_cg/sustain.py +35 -5
- package/md_cg/tasks.py +470 -470
- package/md_cg/test_access_hints.py +147 -0
- package/md_cg/test_action_derive.py +203 -203
- package/md_cg/test_audit_rotate.py +270 -270
- package/md_cg/test_autonomy.py +143 -143
- package/md_cg/test_bench_governance.py +102 -102
- package/md_cg/test_blindspot_tickets.py +166 -166
- package/md_cg/test_branch_discard_tombstone.py +136 -0
- package/md_cg/test_branches.py +13 -3
- package/md_cg/test_ccg_perturb.py +184 -184
- package/md_cg/test_ccgc.py +433 -433
- package/md_cg/test_census_prune.py +81 -81
- package/md_cg/test_chain_read_isolate.py +168 -0
- package/md_cg/test_cond_compose_anchors.py +76 -76
- package/md_cg/test_cond_match.py +165 -165
- package/md_cg/test_condition_anchor.py +81 -81
- package/md_cg/test_d_meta.py +412 -412
- package/md_cg/test_datapath_device_name.py +203 -0
- package/md_cg/test_datapath_root.py +199 -199
- package/md_cg/test_emit_negtail_cache.py +156 -0
- package/md_cg/test_en_pipeline.py +22 -2
- package/md_cg/test_gain_gate.py +212 -212
- package/md_cg/test_govern_directread.py +421 -0
- package/md_cg/test_health_scale.py +173 -173
- package/md_cg/test_hive_ingest.py +285 -0
- package/md_cg/test_hot_cold.py +215 -215
- package/md_cg/test_hyperedge.py +245 -245
- package/md_cg/test_i26_empty_first_write.py +116 -0
- package/md_cg/test_i27_e041_identity.py +128 -0
- package/md_cg/test_i28_hotcache_prodpath.py +122 -0
- package/md_cg/test_i32_hotcache_env_key.py +218 -0
- package/md_cg/test_identity_attribution.py +96 -15
- package/md_cg/test_index_durability.py +17 -3
- package/md_cg/test_interop.py +95 -0
- package/md_cg/test_interop_judgment.py +228 -0
- package/md_cg/test_issue39_utf8_stdio.py +273 -0
- package/md_cg/test_lifecycle.py +309 -309
- package/md_cg/test_linkref.py +306 -306
- package/md_cg/test_links_concurrent_write.py +188 -0
- package/md_cg/test_lock.py +43 -43
- package/md_cg/test_md_access_parity.py +255 -255
- package/md_cg/test_md_writepath.py +345 -345
- package/md_cg/test_mdstore_search_parity.py +160 -0
- package/md_cg/test_merge_upsert.py +168 -0
- package/md_cg/test_mr_m2.py +587 -587
- package/md_cg/test_mr_m3.py +710 -710
- package/md_cg/test_mr_m4.py +485 -485
- package/md_cg/test_n123_derive_expiry_chain.py +205 -0
- package/md_cg/test_n130_verify_falsified_protect.py +185 -0
- package/md_cg/test_n131_merge_gate.py +205 -0
- package/md_cg/test_p0.py +250 -250
- package/md_cg/test_p1.py +316 -316
- package/md_cg/test_p10_identity.py +173 -173
- package/md_cg/test_p11_consistency.py +233 -233
- package/md_cg/test_p12_metacognition.py +212 -212
- package/md_cg/test_p13_encryption.py +241 -241
- package/md_cg/test_p14_sustain.py +249 -249
- package/md_cg/test_p15_scrub.py +280 -280
- package/md_cg/test_p16_self_state.py +301 -301
- package/md_cg/test_p17_predict.py +354 -354
- package/md_cg/test_p18_whitebox.py +171 -171
- package/md_cg/test_p19_migrate_roleplay.py +149 -149
- package/md_cg/test_p1x_ref_root.py +160 -0
- package/md_cg/test_p20_evolution.py +315 -315
- package/md_cg/test_p21_tokens.py +293 -270
- package/md_cg/test_p22_theory.py +175 -175
- package/md_cg/test_p23_links.py +311 -311
- package/md_cg/test_p24_evidence.py +227 -227
- package/md_cg/test_p25_weights.py +156 -156
- package/md_cg/test_p26_refindex.py +416 -416
- package/md_cg/test_p27_docindex.py +16 -7
- package/md_cg/test_p28_refcheck.py +305 -305
- package/md_cg/test_p29_session_ingest_export.py +354 -333
- package/md_cg/test_p2_mcp.py +3 -0
- package/md_cg/test_p3.py +11 -2
- package/md_cg/test_p30_maintain.py +330 -330
- package/md_cg/test_p31_insight.py +534 -534
- package/md_cg/test_p32_backfill.py +7 -1
- package/md_cg/test_p33_ccg_wiring.py +293 -293
- package/md_cg/test_p34_crosscheck.py +331 -331
- package/md_cg/test_p35_conditioned_claim.py +252 -252
- package/md_cg/test_p36_kp_align.py +230 -230
- package/md_cg/test_p37_condition_space.py +248 -248
- package/md_cg/test_p38_concurrent_flush.py +102 -0
- package/md_cg/test_p38_contextualize.py +273 -273
- package/md_cg/test_p39_verify_flow.py +153 -0
- package/md_cg/test_p39_vision_evidence.py +369 -369
- package/md_cg/test_p40_refine_worklist.py +241 -241
- package/md_cg/test_p41_evolve_patrol.py +224 -224
- package/md_cg/test_p42_provenance.py +269 -269
- package/md_cg/test_p43_pooling.py +412 -398
- package/md_cg/test_p44_md_whitebox.py +231 -231
- package/md_cg/test_p45_session_identity.py +219 -219
- package/md_cg/test_p46_unit_scope.py +272 -272
- package/md_cg/test_p47_session_view.py +316 -0
- package/md_cg/test_p4_fuzzy.py +223 -223
- package/md_cg/test_p5_semantic.py +226 -226
- package/md_cg/test_p6_consolidate.py +440 -387
- package/md_cg/test_p7_goals_recent.py +202 -202
- package/md_cg/test_p8_subgraph_chain.py +200 -200
- package/md_cg/test_p9_forget_protect.py +231 -231
- package/md_cg/test_predict_beta.py +135 -135
- package/md_cg/test_preflight_failclosed.py +100 -100
- package/md_cg/test_progressive.py +146 -146
- package/md_cg/test_propose_tail_index.py +157 -0
- package/md_cg/test_protocol.py +243 -243
- package/md_cg/test_reach.py +378 -378
- package/md_cg/test_reach_keys.py +201 -201
- package/md_cg/test_read_clip.py +141 -141
- package/md_cg/test_read_scope_b27.py +277 -0
- package/md_cg/test_readcache_default_on.py +168 -0
- package/md_cg/test_readcache_precise_inval.py +270 -0
- package/md_cg/test_readcache_prodpath.py +203 -0
- package/md_cg/test_reconcile_v0.py +294 -0
- package/md_cg/test_retr_gates_prodpath.py +140 -0
- package/md_cg/test_retr_s1.py +344 -340
- package/md_cg/test_retr_s1b.py +276 -209
- package/md_cg/test_retr_s3.py +194 -194
- package/md_cg/test_retr_s4.py +163 -163
- package/md_cg/test_retr_s5.py +200 -200
- package/md_cg/test_retr_s6.py +157 -157
- package/md_cg/test_retr_s7.py +392 -384
- package/md_cg/test_retr_s8_time.py +369 -316
- package/md_cg/test_retr_s9_edges.py +286 -286
- package/md_cg/test_retr_s9_entity_ctx.py +9 -3
- package/md_cg/test_retr_score_once.py +208 -0
- package/md_cg/test_review_conformance.py +367 -367
- package/md_cg/test_review_onepass.py +170 -0
- package/md_cg/test_role_views.py +354 -354
- package/md_cg/test_rrf_graph_seed_cache.py +154 -0
- package/md_cg/test_security_audit.py +155 -0
- package/md_cg/test_security_audit_b26.py +161 -0
- package/md_cg/test_security_audit_v21.py +250 -0
- package/md_cg/test_sem_noise.py +242 -242
- package/md_cg/test_semantic_canonical.py +16 -2
- package/md_cg/test_session_isolation.py +168 -0
- package/md_cg/test_snapshot_autoclose.py +187 -0
- package/md_cg/test_subproc_encoding.py +192 -192
- package/md_cg/test_sustain_mutual.py +153 -153
- package/md_cg/test_tail_watermark_race.py +208 -0
- package/md_cg/test_tasks.py +409 -409
- package/md_cg/test_tenant_env_override_warn.py +139 -0
- package/md_cg/test_tenant_registry_corrupt_warn.py +151 -0
- package/md_cg/test_tool_face.py +189 -189
- package/md_cg/test_transfer.py +180 -180
- package/md_cg/test_trust.py +361 -361
- package/md_cg/test_twophase.py +286 -286
- package/md_cg/test_v14_fixes.py +38 -20
- package/md_cg/test_validity_filter.py +280 -280
- package/md_cg/test_verify_answer.py +138 -138
- package/md_cg/test_verify_dirty_reconcile.py +157 -0
- package/md_cg/test_wisdom_md_store.py +292 -292
- package/md_cg/test_writelimit.py +197 -197
- package/md_cg/test_writepipe.py +214 -214
- package/md_cg/theory.py +6 -3
- package/md_cg/tokens.py +85 -14
- package/md_cg/tool_face.py +260 -260
- package/md_cg/trust.py +986 -950
- package/md_cg/twophase.py +231 -231
- package/md_cg/units.py +668 -667
- package/md_cg/vision_evidence.py +667 -666
- package/md_cg/weights.py +624 -624
- package/md_cg/whitebox.py +527 -527
- package/md_cg/whitebox_kb/__init__.py +37 -37
- package/md_cg/whitebox_kb/aeis_core/__init__.py +42 -42
- package/md_cg/whitebox_kb/aeis_core/semantic.py +280 -280
- package/md_cg/whitebox_kb/aeis_core/textutil.py +13 -13
- package/md_cg/whitebox_kb/engine.py +310 -310
- package/md_cg/whitebox_kb/seed_knowledge//346/231/272/350/203/275/350/256/2723.4.md +5260 -5260
- package/md_cg/whitebox_kb/wisdom/browser_units.py +2631 -2631
- package/md_cg/whitebox_kb/wisdom/causal_discover.py +432 -432
- package/md_cg/whitebox_kb/wisdom/chat_engine.py +1506 -1506
- package/md_cg/whitebox_kb/wisdom/compiler_code_units.py +3033 -3033
- package/md_cg/whitebox_kb/wisdom/condition_algebra.py +112 -112
- package/md_cg/whitebox_kb/wisdom/condition_frame.py +315 -315
- package/md_cg/whitebox_kb/wisdom/condition_kb.py +108 -108
- package/md_cg/whitebox_kb/wisdom/conflict_map.json +4445 -4445
- package/md_cg/whitebox_kb/wisdom/core/lexer.py +512 -512
- package/md_cg/whitebox_kb/wisdom/core/name_checker.py +1023 -1023
- package/md_cg/whitebox_kb/wisdom/cspmn.py +258 -258
- package/md_cg/whitebox_kb/wisdom/csre.py +264 -264
- package/md_cg/whitebox_kb/wisdom/danmaku_audit.py +252 -252
- package/md_cg/whitebox_kb/wisdom/distilled_condition_units.json +6417 -6417
- package/md_cg/whitebox_kb/wisdom/docs/WB-EVAL-20260902b.json +1950 -1950
- package/md_cg/whitebox_kb/wisdom/docs/WB-EVAL-20260902c.json +1296 -1296
- package/md_cg/whitebox_kb/wisdom/docs/whitebox_capability_graph_demo.json +59 -59
- package/md_cg/whitebox_kb/wisdom/graph_db_units.py +3089 -3089
- package/md_cg/whitebox_kb/wisdom/knowledge_points.py +318 -318
- package/md_cg/whitebox_kb/wisdom/md_access.py +470 -470
- package/md_cg/whitebox_kb/wisdom/md_store.py +276 -251
- package/md_cg/whitebox_kb/wisdom/migrate_wisdom.py +476 -476
- package/md_cg/whitebox_kb/wisdom/multilang_ir.py +128 -124
- package/md_cg/whitebox_kb/wisdom/navigate.py +248 -248
- package/md_cg/whitebox_kb/wisdom/neural_retrieve.py +224 -224
- package/md_cg/whitebox_kb/wisdom/os_units.py +2735 -2735
- package/md_cg/whitebox_kb/wisdom/pattern_separation.py +376 -376
- package/md_cg/whitebox_kb/wisdom/prereq_map.json +364 -364
- package/md_cg/whitebox_kb/wisdom/python_code_units.py +2766 -2766
- package/md_cg/whitebox_kb/wisdom/role_solidified.json +7 -7
- package/md_cg/whitebox_kb/wisdom/route_memory.py +244 -244
- package/md_cg/whitebox_kb/wisdom/scene_reconstruction.py +148 -148
- package/md_cg/whitebox_kb/wisdom/snr_report.json +40 -40
- package/md_cg/whitebox_kb/wisdom/test_code_compose_domains.py +7541 -7541
- package/md_cg/whitebox_kb/wisdom/test_compiler_self_bootstrap.py +354 -354
- package/md_cg/whitebox_kb/wisdom/test_ecosystem_assembly.py +158 -158
- package/md_cg/whitebox_kb/wisdom/test_ecosystem_demos.py +193 -193
- package/md_cg/whitebox_kb/wisdom/test_graph_db.py +292 -292
- package/md_cg/whitebox_kb/wisdom/test_python_self_bootstrap.py +160 -160
- package/md_cg/whitebox_kb/wisdom/trigger_words_index.json +4366 -4366
- package/md_cg/whitebox_kb/wisdom/verifier.py +1297 -1297
- package/md_cg/writelimit.py +356 -356
- package/md_cg/writepipe.py +20 -8
- package/package.json +101 -96
- package/skills/plugin.json +54 -54
- package/skills/skills/designer-perspective/SKILL.md +158 -158
- package/skills/skills/designer-perspective/references/01-observation-position.md +66 -66
- package/skills/skills/designer-perspective/references/02-structure-recognition.md +62 -62
- package/skills/skills/designer-perspective/references/03-direction-judgment.md +55 -55
- package/skills/skills/designer-perspective/references/04-qualification-verdict.md +72 -72
- package/skills/skills/designer-perspective/references/05-condition-attribution.md +74 -74
- package/skills/skills/designer-perspective/scripts/designer.py +545 -545
- package/skills/skills/designer-perspective/tests/cases.jsonl +17 -17
- package/skills/skills/designer-perspective/tests/selftest.py +61 -61
- package/skills/skills/lingshu-browser/SKILL.md +60 -60
- package/skills/skills/lingshu-compiler/SKILL.md +56 -56
- package/skills/skills/lingshu-compiler/units/analyze-type-infer/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/check-name-real/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compile-assign/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compile-expr-tree/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compile-full-pipeline/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compile-func-def/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compile-if-then/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compile-logic-expr/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compile-recursive/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compile-scope/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compile-type-check/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compile-while/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compiler-0010c4bf/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compiler-0355bffb/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compiler-0361708a/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compiler-054a0414/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compiler-05a1691a/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compiler-05eeed1e/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compiler-0622a1f6/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compiler-08b54217/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compiler-0a62b70c/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compiler-0ab24d00/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compiler-0e093688/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compiler-0e82b966/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compiler-0ee9b9b9/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compiler-0f3787e6/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compiler-1028685f/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compiler-11897630/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compiler-16661b9b/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compiler-1f722303/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compiler-25be1262/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compiler-2a76ba07/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compiler-2dbea54a/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compiler-2df52f16/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compiler-2ec2c9d2/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compiler-2f8c8f39/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compiler-38378ac7/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compiler-39457d2e/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compiler-39fb5926/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compiler-47f4fbfa/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compiler-4a8cd1f1/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compiler-4b230b6b/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compiler-4cbbda95/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compiler-4d5a68ff/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compiler-56dc9bdd/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compiler-57e76ebe/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compiler-61ff016b/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compiler-63eae588/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compiler-64c4224b/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compiler-66377955/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compiler-674ab4f6/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compiler-6af76fd6/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compiler-6d979db2/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compiler-6de326c0/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compiler-6f1c0eea/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compiler-724d6c9b/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compiler-7690177f/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compiler-7b229a7d/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compiler-7f471b3a/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compiler-815cad08/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compiler-83c61634/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compiler-8c798c81/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compiler-8dd747ac/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compiler-90324d0a/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compiler-94b8d72d/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compiler-94f12231/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compiler-95937c16/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compiler-98a5625b/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compiler-98b4c42f/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compiler-98e3894a/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compiler-9bdfe4b8/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compiler-9d9b4e83/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compiler-9f7be5ad/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compiler-a6daf076/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compiler-a7995a0c/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compiler-a8399248/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compiler-aabbd099/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compiler-afe169d8/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compiler-b0678bda/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compiler-b09bd196/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compiler-b1396e23/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compiler-b9b31ce0/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compiler-c10264a7/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compiler-cb1e8e4b/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compiler-ccafd438/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compiler-cda9c262/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compiler-ce648068/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compiler-cf5776a4/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compiler-d974e5d3/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compiler-e3979fd3/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compiler-eb1cf2b5/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compiler-ecb30d5b/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compiler-f8c8b24b/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compiler-f99fedbe/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compiler-fa8ff5f7/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compiler-fe1b058d/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/lex-chinese-program/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/lex-dao-de-jing/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/lex-nine-chapters/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/vm-arithmetic/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/vm-array-ops/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/vm-closure-call/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/vm-closure-create/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/vm-compare/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/vm-cond-jump/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/vm-cond-space/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/vm-exception/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/vm-func-call/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/vm-loop-run/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/vm-profiling/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/vm-refcount/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/vm-run-loop/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/vm-short-circuit/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/vm-stack-guard/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/vm-stack-ops/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/vm-trust-accum/SKILL.md +45 -45
- package/skills/skills/lingshu-graph/SKILL.md +63 -63
- package/skills/skills/lingshu-net/SKILL.md +48 -48
- package/skills/skills/lingshu-os/SKILL.md +64 -64
- package/skills/skills/lingshu-pylang/SKILL.md +71 -71
- package/src/bridge.ts +33 -0
- package/src/hooks.ts +38 -2
- package/src/index.ts +526 -518
- package/src/lib/datapath.ts +326 -326
- package/src/lib/mdcg_client.ts +413 -413
- package/src/lib/mutual.ts +428 -428
- package/src/lib/prompt_safety.ts +62 -62
- package/src/lib/python_path.ts +71 -71
- package/src/lib/roleplay_web.ts +116 -29
- package/src/lib/token_store.ts +13 -3
- package/src/tools.ts +212 -212
- package/zcode/AGENTS.md +11 -3
- package/zcode/README.md +41 -41
- /package/docs/{mdcg → hive}//344/270/273/344/273/243/347/220/206/345/255/220/344/273/243/347/220/206/350/256/260/345/277/206/346/236/266/346/236/204/350/256/276/350/256/241.md" +0 -0
|
@@ -1,288 +1,288 @@
|
|
|
1
|
-
#!/usr/bin/env python3
|
|
2
|
-
# -*- coding: utf-8 -*-
|
|
3
|
-
"""bench_progressive.py · 渐进式语义检索双实验(2026-09-14)
|
|
4
|
-
|
|
5
|
-
理论(使用者与 GPT 定稿):查询不是固定语义结构,而是不断收紧的约束集合;
|
|
6
|
-
从最小可靠语义开始宽检索,再按区分度逐步增加条件使语义收敛。
|
|
7
|
-
「语义解析的完整性与检索的必要性不是同一个问题。」
|
|
8
|
-
|
|
9
|
-
实验一(locomo-zh-500 · 原子面,与 bench_en_atoms_public 同 doc 侧口径):
|
|
10
|
-
G0 一次性全原子 —— 对照锚点(应 ≈ ② 96.8/99.8/99.8,同口径自校验)
|
|
11
|
-
G2 只宽检不细化 —— core 50% 一次性排序(「最小可靠语义」本身水平)
|
|
12
|
-
G1 渐进 —— Stage-1 宽检索 → 区分度引导加原子 → 重排
|
|
13
|
-
(max_stages=4;每阶段都被测量)
|
|
14
|
-
统计:initial(Stage-1)/ final(终态)hit@1/5/10+MRR、平均 refinement
|
|
15
|
-
steps(added 条件数)、一步收敛率。
|
|
16
|
-
诚实边界:池 567 全 gold 零干扰——G1≈G0 属预期(⑤⑥⑦已证宽检索
|
|
17
|
-
鲁棒);本实验量化「渐进不劣于一次性 + 步数可观测」,消歧增益见实验二。
|
|
18
|
-
|
|
19
|
-
实验二(受控干扰池 · 引擎侧,42+2 节点真实 MdCGOS):
|
|
20
|
-
在 test_sem_noise 语料上放 2 条条件冒充邻居(正条件与 gold 同款 →
|
|
21
|
-
judge 判 ACCEPT)——构造 False ACCEPT>0 的形态;渐进循环:
|
|
22
|
-
宽检索 → discriminative 在候选间选区分性条件词 → 带条件重排。
|
|
23
|
-
断言:渐进后 False ACCEPT=0 且 top1 ∈ gold——
|
|
24
|
-
「DEFER/冒充不是终点,加条件后语义收敛」的引擎侧实证。
|
|
25
|
-
条件词池本轮手工给定(示教);真实系统应由 atoms.json 标准词表驱动。
|
|
26
|
-
|
|
27
|
-
跑法:python -m md_cg.bench_progressive
|
|
28
|
-
"""
|
|
29
|
-
import io
|
|
30
|
-
import json
|
|
31
|
-
import os
|
|
32
|
-
import shutil
|
|
33
|
-
import sys
|
|
34
|
-
import tempfile
|
|
35
|
-
import time
|
|
36
|
-
|
|
37
|
-
HERE = os.path.dirname(os.path.dirname(os.path.abspath(__file__)))
|
|
38
|
-
if HERE not in sys.path:
|
|
39
|
-
sys.path.insert(0, HERE)
|
|
40
|
-
|
|
41
|
-
from md_cg import bench6_common as b6 # noqa: E402
|
|
42
|
-
from md_cg.bench_en_atoms_public import ( # noqa: E402
|
|
43
|
-
CORPUS, QUESTIONS, jaccard, load_char_atoms, node_atom_set,
|
|
44
|
-
round_dict, zh_map_en)
|
|
45
|
-
from md_cg.mdcos import MdCGOS # noqa: E402
|
|
46
|
-
from md_cg.progressive import discriminative, progressive_search # noqa: E402
|
|
47
|
-
from md_cg.test_sem_noise import ( # noqa: E402
|
|
48
|
-
GOLD, GOLD_COND, NEG_TEXTS, NEIGHBORS, QUERY, UNREL_DOCS, _doc)
|
|
49
|
-
|
|
50
|
-
|
|
51
|
-
# 生效条件:ev、ids 给定时按 ids 顺序找首个满足 nid in ev 的 1-based 位次返回,找不到(或 ids 为空、ev 为空)时返回 0。
|
|
52
|
-
def rank_of(ev, ids):
|
|
53
|
-
"""gold 在 id 排名中的位次(不在则 0)。"""
|
|
54
|
-
return next((r for r, nid in enumerate(ids, 1) if nid in ev), 0)
|
|
55
|
-
|
|
56
|
-
|
|
57
|
-
# 生效条件:rank 为 1 时 st["h1"] 加 1,rank 落在 0<rank<=5 时 st["h5"] 加 1,落在 0<rank<=10 时 st["h10"] 加 1,rank 为真值时 st["rr"] 加 1.0/rank(负 rank 亦进入该分支得负增量),无返回值。
|
|
58
|
-
def _agg(rank, st):
|
|
59
|
-
if rank == 1:
|
|
60
|
-
st["h1"] += 1
|
|
61
|
-
if 0 < rank <= 5:
|
|
62
|
-
st["h5"] += 1
|
|
63
|
-
if 0 < rank <= 10:
|
|
64
|
-
st["h10"] += 1
|
|
65
|
-
if rank:
|
|
66
|
-
st["rr"] += 1.0 / rank
|
|
67
|
-
|
|
68
|
-
|
|
69
|
-
# 生效条件:label、st、n 与可选 extra 给定时打印 label、st["h1"/"h5"/"h10"]*100.0/n、st["rr"]/n 及 extra 组成的行,n 为 0 时因源码直接相除抛 ZeroDivisionError(源码未校验 n)。
|
|
70
|
-
def _show(label, st, n, extra=""):
|
|
71
|
-
print(" %-22s hit@1=%5.1f%% hit@5=%5.1f%% hit@10=%5.1f%% "
|
|
72
|
-
"MRR=%.4f%s" % (label, st["h1"] * 100.0 / n,
|
|
73
|
-
st["h5"] * 100.0 / n, st["h10"] * 100.0 / n,
|
|
74
|
-
st["rr"] / n, extra))
|
|
75
|
-
|
|
76
|
-
|
|
77
|
-
# 生效条件:qa、docs 给定时对 docs 每项按 jaccard(qa, na) 降序、id 升序排序并返回其 id 序列,docs 为空返回空列表,qa 为空时全部分数相同故只按 id 升序。
|
|
78
|
-
def full_rank(qa, docs):
|
|
79
|
-
"""全池 Jaccard 排序 → id 序列(与 run_arm 同判据)。"""
|
|
80
|
-
scored = sorted(((jaccard(qa, na), nid) for nid, na in docs),
|
|
81
|
-
key=lambda t: (-t[0], t[1]))
|
|
82
|
-
return [nid for _s, nid in scored]
|
|
83
|
-
|
|
84
|
-
|
|
85
|
-
# 生效条件:无 required 形参;以模块级常量 CORPUS、QUESTIONS 逐行 json.loads 建语料与题目,每题取 evidence_turns 与 question 分词结果,qa 为空则跳过且不计入 n,最终返回含 g0/g2/g1i/g1f/n/fa10_conservative/avg_pool 的统计字典(分母均以 max(·,1) 兜零)。
|
|
86
|
-
def exp_locomo():
|
|
87
|
-
"""实验一:原子面渐进三形态。"""
|
|
88
|
-
t0 = time.time()
|
|
89
|
-
with io.open(CORPUS, encoding="utf-8") as f:
|
|
90
|
-
corpus = [json.loads(ln) for ln in f if ln.strip()]
|
|
91
|
-
with io.open(QUESTIONS, encoding="utf-8") as f:
|
|
92
|
-
questions = [json.loads(ln) for ln in f if ln.strip()]
|
|
93
|
-
char_atoms = load_char_atoms()
|
|
94
|
-
docs = [(c["id"], node_atom_set(c, char_atoms)) for c in corpus]
|
|
95
|
-
docs_map = dict(docs)
|
|
96
|
-
print("[progressive] 实验一 locomo-zh-500 · 语料 %d / 题 %d"
|
|
97
|
-
% (len(corpus), len(questions)))
|
|
98
|
-
|
|
99
|
-
pool_sizes = [] # G1 各阶段候选池大小(统计面)
|
|
100
|
-
|
|
101
|
-
# 生效条件:used 给定时对其 frozenset 调 full_rank 取前 20 条,把该池大小追加进 pool_sizes,返回由这些 id 与 0.0 组成的二元组列表。
|
|
102
|
-
def rank_fn(used):
|
|
103
|
-
ranked = full_rank(frozenset(used), docs)[:20]
|
|
104
|
-
pool_sizes.append(len(ranked))
|
|
105
|
-
return [(nid, 0.0) for nid in ranked]
|
|
106
|
-
|
|
107
|
-
st_g0 = {"h1": 0, "h5": 0, "h10": 0, "rr": 0.0}
|
|
108
|
-
st_g2 = dict(st_g0)
|
|
109
|
-
st_i = dict(st_g0)
|
|
110
|
-
st_f = dict(st_g0)
|
|
111
|
-
steps_sum = conv = n = fa10_sum = 0
|
|
112
|
-
for q in questions:
|
|
113
|
-
ev = set(q.get("evidence_turns") or [])
|
|
114
|
-
qa = [w for w in zh_map_en(str(q.get("question") or ""),
|
|
115
|
-
char_atoms).split() if w]
|
|
116
|
-
if not qa:
|
|
117
|
-
continue
|
|
118
|
-
n += 1
|
|
119
|
-
# G0 一次性全原子(对照锚点)
|
|
120
|
-
_agg(rank_of(ev, full_rank(frozenset(qa), docs)), st_g0)
|
|
121
|
-
# G2 只宽检不细化
|
|
122
|
-
core = qa[:max(1, int(len(qa) * 0.5))]
|
|
123
|
-
_agg(rank_of(ev, full_rank(frozenset(core), docs)), st_g2)
|
|
124
|
-
# G1 渐进
|
|
125
|
-
tr = progressive_search(rank_fn, qa, docs_map,
|
|
126
|
-
ratio=0.5, k=10, max_stages=4)
|
|
127
|
-
_agg(rank_of(ev, full_rank(frozenset(tr["stages"][0]["used"]),
|
|
128
|
-
docs)), st_i)
|
|
129
|
-
fin_all = full_rank(frozenset(tr["stages"][-1]["used"]), docs)
|
|
130
|
-
_agg(rank_of(ev, fin_all), st_f)
|
|
131
|
-
# 保守 false_accept 口径:final top10 内非 gold 条数(本语料无资格
|
|
132
|
-
# 判定面,非 gold 全计为假接受——上界,见打印行口径注明)
|
|
133
|
-
fa10_sum += sum(1 for nid in fin_all[:10] if nid not in ev)
|
|
134
|
-
steps_sum += len(tr["added"])
|
|
135
|
-
conv += 1 if (tr["converged"] or len(tr["added"]) < 3) else 0
|
|
136
|
-
|
|
137
|
-
_show("G0 一次性全原子", st_g0, n,
|
|
138
|
-
" ← 锚点应≈② 96.8/99.8/99.8")
|
|
139
|
-
_show("G2 只宽检(core50%)", st_g2, n)
|
|
140
|
-
_show("G1 渐进·initial", st_i, n)
|
|
141
|
-
_show("G1 渐进·final", st_f, n)
|
|
142
|
-
print(" refinement: 平均加条件数=%.2f 一步收敛率=%.1f%% (n=%d, %.0fs)"
|
|
143
|
-
% (steps_sum / max(n, 1), conv * 100.0 / max(n, 1), n,
|
|
144
|
-
time.time() - t0))
|
|
145
|
-
print(" false_accept(保守口径)=%.2f%% =G1 final top10 非 gold 平均占比"
|
|
146
|
-
"(上界:本语料无 CCG 注释面,非 gold 全计为假接受)"
|
|
147
|
-
% (fa10_sum * 10.0 / max(n, 1)))
|
|
148
|
-
print(" 候选池: G1 各阶段平均 top 池大小=%.1f (k=20 上限)"
|
|
149
|
-
% (sum(pool_sizes) / max(len(pool_sizes), 1)))
|
|
150
|
-
print(" 四态分布: 本实验为原子面(语料无 CCG 注释,judge_qualification "
|
|
151
|
-
"将全判 BLINDSPOT)——资格判定不适用,不硬造数字;四态实证见实验二")
|
|
152
|
-
return {"g0": st_g0, "g2": st_g2, "g1i": st_i, "g1f": st_f, "n": n,
|
|
153
|
-
"fa10_conservative": fa10_sum * 10.0 / max(n, 1),
|
|
154
|
-
"avg_pool": sum(pool_sizes) / max(len(pool_sizes), 1)}
|
|
155
|
-
|
|
156
|
-
|
|
157
|
-
# 生效条件:无必需形参;自建临时库(tempfile.mkdtemp + MdCGOS)注入 GOLD/NEIGHBORS 语料后跑渐进消歧,finally 清理临时目录;返回指标 dict,不依赖外部数据;
|
|
158
|
-
def exp_controlled():
|
|
159
|
-
"""实验二:受控干扰池——条件冒充形态下渐进消歧(引擎侧真检索)。"""
|
|
160
|
-
print("\n[progressive] 实验二 受控干扰池(42+2 节点真实 MdCGOS)")
|
|
161
|
-
tmp = tempfile.mkdtemp(prefix="mdcg_prog_")
|
|
162
|
-
try:
|
|
163
|
-
root = os.path.join(tmp, "root")
|
|
164
|
-
cg = MdCGOS(root)
|
|
165
|
-
gold_ids = [g for g, _ in GOLD]
|
|
166
|
-
for gid, fact in GOLD:
|
|
167
|
-
cg.add(gid, _doc(fact, GOLD_COND), layer="knowledge",
|
|
168
|
-
verification_basis="test")
|
|
169
|
-
# 2 条邻居改用 gold 同款生效条件 → 正条件命中 → judge 判 ACCEPT(冒充)
|
|
170
|
-
IMPOSTORS = ("n1a", "n3a")
|
|
171
|
-
for gid, lst in NEIGHBORS.items():
|
|
172
|
-
for nid, fact, _cond in lst:
|
|
173
|
-
cg.add(nid, _doc(fact, GOLD_COND if nid in IMPOSTORS
|
|
174
|
-
else _cond), layer="knowledge",
|
|
175
|
-
verification_basis="test")
|
|
176
|
-
for nid, fact in NEG_TEXTS:
|
|
177
|
-
cg.add(nid, _doc(fact, "问没做的事", extra_neg=True),
|
|
178
|
-
layer="knowledge", verification_basis="test",
|
|
179
|
-
non_applicable_conditions=["刚才在干什么"])
|
|
180
|
-
for nid, fact, cond in UNREL_DOCS:
|
|
181
|
-
cg.add(nid, _doc(fact, cond), layer="knowledge",
|
|
182
|
-
verification_basis="test")
|
|
183
|
-
|
|
184
|
-
KW = dict(paths=("lexical", "bucket", "entity", "graph", "fuzzy"),
|
|
185
|
-
fusion="sum", judge_ranking=True)
|
|
186
|
-
|
|
187
|
-
# 生效条件:used 给定时,used 为假值(空列表等)则查询串仅为 QUERY,否则为 QUERY 加空格再加空格连接的 used,再经 cg.search_rrf(k=10, judge=True, **KW) 返回 (id, 分值) 列表。
|
|
188
|
-
def engine_rank(used):
|
|
189
|
-
res, _m = cg.search_rrf(QUERY + (" " + " ".join(used)
|
|
190
|
-
if used else ""),
|
|
191
|
-
k=10, judge=True, **KW)
|
|
192
|
-
return [(r[0]["id"], r[1]) for r in res]
|
|
193
|
-
|
|
194
|
-
# 生效条件:res 给定时遍历其前 10 项,跳过 r[0]["id"] 属于外层 gold_ids 的项,对余项取 r[2](len(r)<=2 时取空 dict)并仅在 (ji or {}).get("state")=="ACCEPT" 时把该 id 收入,返回收集到的 out 列表。
|
|
195
|
-
def false_accepts(res):
|
|
196
|
-
out = []
|
|
197
|
-
for r in res[:10]:
|
|
198
|
-
nid = r[0]["id"]
|
|
199
|
-
if nid in gold_ids:
|
|
200
|
-
continue
|
|
201
|
-
ji = r[2] if len(r) > 2 else {}
|
|
202
|
-
if (ji or {}).get("state") == "ACCEPT":
|
|
203
|
-
out.append(nid)
|
|
204
|
-
return out
|
|
205
|
-
|
|
206
|
-
# 生效条件:res 给定时遍历其前 10 项,len(r)>2 时取 (r[2] or {}).get("state")、否则取 None,该值属于 ACCEPT/DEFER/REJECT/BLINDSPOT 四键之一则对应计数、否则计入「无judge」,返回只保留计数非零项的字典(全零则返回空 dict)。
|
|
207
|
-
def state_dist(res):
|
|
208
|
-
"""渐进循环四态分布(top10 内 judge state 计数,只列非零项)。
|
|
209
|
-
|
|
210
|
-
judge info 直接复用 search_rrf(judge=True) 注入的
|
|
211
|
-
judge_qualification 产物(结果第三元),不重复判定。
|
|
212
|
-
"""
|
|
213
|
-
d = {"ACCEPT": 0, "DEFER": 0, "REJECT": 0, "BLINDSPOT": 0,
|
|
214
|
-
"无judge": 0}
|
|
215
|
-
for r in res[:10]:
|
|
216
|
-
st = (r[2] or {}).get("state") if len(r) > 2 else None
|
|
217
|
-
d[st if st in d else "无judge"] += 1
|
|
218
|
-
return {k: v for k, v in d.items() if v}
|
|
219
|
-
|
|
220
|
-
# 生效条件:res、fa 给定时返回 100.0*len(fa)/max(1, min(10, len(res))),即 len(res)>=10 时除以 10、不足时除以 len(res),len(res) 为 0 时被 max(1,·) 兜为除数 1。
|
|
221
|
-
def fa_rate(res, fa):
|
|
222
|
-
return 100.0 * len(fa) / max(1, min(10, len(res)))
|
|
223
|
-
|
|
224
|
-
# 宽检索(不加任何条件;judgeinfo 直接取 results 第三元)
|
|
225
|
-
wide_res, _m = cg.search_rrf(QUERY, k=10, judge=True, **KW)
|
|
226
|
-
wide = [(r[0]["id"], r[1]) for r in wide_res]
|
|
227
|
-
fa_pre = false_accepts(wide_res)
|
|
228
|
-
top1_pre = wide[0][0] if wide else None
|
|
229
|
-
print(" 宽检索 top1=%s top10 内 False ACCEPT=%s(率=%.1f%%)"
|
|
230
|
-
% (top1_pre, fa_pre or "无", fa_rate(wide_res, fa_pre)))
|
|
231
|
-
print(" 四态分布(宽检索 top10)=%s 候选池=%d judge 剔除=%s"
|
|
232
|
-
% (state_dist(wide_res), len(wide_res),
|
|
233
|
-
(_m or {}).get("judge_filtered", 0)))
|
|
234
|
-
|
|
235
|
-
# 渐进:候选间区分性条件词 → 带条件重排
|
|
236
|
-
POOL = ["我", "喝水", "吃饭", "电脑", "书", "音乐", "出门",
|
|
237
|
-
"浇花", "下班"]
|
|
238
|
-
cand_chars = {}
|
|
239
|
-
for r in wide_res[:10]:
|
|
240
|
-
nid = r[0]["id"]
|
|
241
|
-
cand_chars[nid] = frozenset(str(r[0].get("content") or ""))
|
|
242
|
-
cond = discriminative([p for p in POOL],
|
|
243
|
-
[cand_chars.get(nid, frozenset())
|
|
244
|
-
for nid, _s in wide[:10]])
|
|
245
|
-
if cond:
|
|
246
|
-
refined_res, _m2 = cg.search_rrf(QUERY + " " + cond,
|
|
247
|
-
k=10, judge=True, **KW)
|
|
248
|
-
fa_post = false_accepts(refined_res)
|
|
249
|
-
top1_post = refined_res[0][0]["id"] if refined_res else None
|
|
250
|
-
print(" 区分性条件=%r → 重排 top1=%s top10 内 "
|
|
251
|
-
"False ACCEPT=%s(率=%.1f%%)"
|
|
252
|
-
% (cond, top1_post, fa_post or "无",
|
|
253
|
-
fa_rate(refined_res, fa_post)))
|
|
254
|
-
print(" 四态分布(重排 top10)=%s 候选池=%d judge 剔除=%s"
|
|
255
|
-
% (state_dist(refined_res), len(refined_res),
|
|
256
|
-
(_m2 or {}).get("judge_filtered", 0)))
|
|
257
|
-
ok1 = top1_post in gold_ids
|
|
258
|
-
no_new = set(fa_post) <= set(fa_pre)
|
|
259
|
-
print(" 断言①渐进后 top1∈gold:%s" % ("PASS" if ok1 else "FAIL"))
|
|
260
|
-
print(" 断言②渐进不引入新冒充(fa_post⊆fa_pre):%s"
|
|
261
|
-
% ("PASS" if no_new else "FAIL"))
|
|
262
|
-
if fa_post:
|
|
263
|
-
for r in refined_res[:10]:
|
|
264
|
-
if r[0]["id"] in fa_post and len(r) > 2:
|
|
265
|
-
print(" 冒充 %s judge reason:%s"
|
|
266
|
-
% (r[0]["id"],
|
|
267
|
-
(r[2] or {}).get("reason", "")))
|
|
268
|
-
print(" 边界(如实):条件面冒充(生效条件被伪造为 gold 同款)"
|
|
269
|
-
"不被内容词渐进消除——词面条件确认是必要非充分,"
|
|
270
|
-
"真伪判据=verification_basis/时间线冲突(证据强度分离,另案)")
|
|
271
|
-
return ok1 and no_new
|
|
272
|
-
print(" 无区分性条件可选(宽检索已无冒充)——如实记录")
|
|
273
|
-
return not fa_pre
|
|
274
|
-
finally:
|
|
275
|
-
shutil.rmtree(tmp, ignore_errors=True)
|
|
276
|
-
|
|
277
|
-
|
|
278
|
-
# 生效条件:无 required 形参;先执行 exp_locomo(),再按 exp_controlled() 返回值 ok 的真假返回 0(真)或 1(假)。
|
|
279
|
-
def main():
|
|
280
|
-
exp_locomo()
|
|
281
|
-
ok = exp_controlled()
|
|
282
|
-
print("\n[progressive] 完成。受控池消歧断言:%s"
|
|
283
|
-
% ("PASS" if ok else "FAIL"))
|
|
284
|
-
return 0 if ok else 1
|
|
285
|
-
|
|
286
|
-
|
|
287
|
-
if __name__ == "__main__":
|
|
1
|
+
#!/usr/bin/env python3
|
|
2
|
+
# -*- coding: utf-8 -*-
|
|
3
|
+
"""bench_progressive.py · 渐进式语义检索双实验(2026-09-14)
|
|
4
|
+
|
|
5
|
+
理论(使用者与 GPT 定稿):查询不是固定语义结构,而是不断收紧的约束集合;
|
|
6
|
+
从最小可靠语义开始宽检索,再按区分度逐步增加条件使语义收敛。
|
|
7
|
+
「语义解析的完整性与检索的必要性不是同一个问题。」
|
|
8
|
+
|
|
9
|
+
实验一(locomo-zh-500 · 原子面,与 bench_en_atoms_public 同 doc 侧口径):
|
|
10
|
+
G0 一次性全原子 —— 对照锚点(应 ≈ ② 96.8/99.8/99.8,同口径自校验)
|
|
11
|
+
G2 只宽检不细化 —— core 50% 一次性排序(「最小可靠语义」本身水平)
|
|
12
|
+
G1 渐进 —— Stage-1 宽检索 → 区分度引导加原子 → 重排
|
|
13
|
+
(max_stages=4;每阶段都被测量)
|
|
14
|
+
统计:initial(Stage-1)/ final(终态)hit@1/5/10+MRR、平均 refinement
|
|
15
|
+
steps(added 条件数)、一步收敛率。
|
|
16
|
+
诚实边界:池 567 全 gold 零干扰——G1≈G0 属预期(⑤⑥⑦已证宽检索
|
|
17
|
+
鲁棒);本实验量化「渐进不劣于一次性 + 步数可观测」,消歧增益见实验二。
|
|
18
|
+
|
|
19
|
+
实验二(受控干扰池 · 引擎侧,42+2 节点真实 MdCGOS):
|
|
20
|
+
在 test_sem_noise 语料上放 2 条条件冒充邻居(正条件与 gold 同款 →
|
|
21
|
+
judge 判 ACCEPT)——构造 False ACCEPT>0 的形态;渐进循环:
|
|
22
|
+
宽检索 → discriminative 在候选间选区分性条件词 → 带条件重排。
|
|
23
|
+
断言:渐进后 False ACCEPT=0 且 top1 ∈ gold——
|
|
24
|
+
「DEFER/冒充不是终点,加条件后语义收敛」的引擎侧实证。
|
|
25
|
+
条件词池本轮手工给定(示教);真实系统应由 atoms.json 标准词表驱动。
|
|
26
|
+
|
|
27
|
+
跑法:python -m md_cg.bench_progressive
|
|
28
|
+
"""
|
|
29
|
+
import io
|
|
30
|
+
import json
|
|
31
|
+
import os
|
|
32
|
+
import shutil
|
|
33
|
+
import sys
|
|
34
|
+
import tempfile
|
|
35
|
+
import time
|
|
36
|
+
|
|
37
|
+
HERE = os.path.dirname(os.path.dirname(os.path.abspath(__file__)))
|
|
38
|
+
if HERE not in sys.path:
|
|
39
|
+
sys.path.insert(0, HERE)
|
|
40
|
+
|
|
41
|
+
from md_cg import bench6_common as b6 # noqa: E402
|
|
42
|
+
from md_cg.bench_en_atoms_public import ( # noqa: E402
|
|
43
|
+
CORPUS, QUESTIONS, jaccard, load_char_atoms, node_atom_set,
|
|
44
|
+
round_dict, zh_map_en)
|
|
45
|
+
from md_cg.mdcos import MdCGOS # noqa: E402
|
|
46
|
+
from md_cg.progressive import discriminative, progressive_search # noqa: E402
|
|
47
|
+
from md_cg.test_sem_noise import ( # noqa: E402
|
|
48
|
+
GOLD, GOLD_COND, NEG_TEXTS, NEIGHBORS, QUERY, UNREL_DOCS, _doc)
|
|
49
|
+
|
|
50
|
+
|
|
51
|
+
# 生效条件:ev、ids 给定时按 ids 顺序找首个满足 nid in ev 的 1-based 位次返回,找不到(或 ids 为空、ev 为空)时返回 0。
|
|
52
|
+
def rank_of(ev, ids):
|
|
53
|
+
"""gold 在 id 排名中的位次(不在则 0)。"""
|
|
54
|
+
return next((r for r, nid in enumerate(ids, 1) if nid in ev), 0)
|
|
55
|
+
|
|
56
|
+
|
|
57
|
+
# 生效条件:rank 为 1 时 st["h1"] 加 1,rank 落在 0<rank<=5 时 st["h5"] 加 1,落在 0<rank<=10 时 st["h10"] 加 1,rank 为真值时 st["rr"] 加 1.0/rank(负 rank 亦进入该分支得负增量),无返回值。
|
|
58
|
+
def _agg(rank, st):
|
|
59
|
+
if rank == 1:
|
|
60
|
+
st["h1"] += 1
|
|
61
|
+
if 0 < rank <= 5:
|
|
62
|
+
st["h5"] += 1
|
|
63
|
+
if 0 < rank <= 10:
|
|
64
|
+
st["h10"] += 1
|
|
65
|
+
if rank:
|
|
66
|
+
st["rr"] += 1.0 / rank
|
|
67
|
+
|
|
68
|
+
|
|
69
|
+
# 生效条件:label、st、n 与可选 extra 给定时打印 label、st["h1"/"h5"/"h10"]*100.0/n、st["rr"]/n 及 extra 组成的行,n 为 0 时因源码直接相除抛 ZeroDivisionError(源码未校验 n)。
|
|
70
|
+
def _show(label, st, n, extra=""):
|
|
71
|
+
print(" %-22s hit@1=%5.1f%% hit@5=%5.1f%% hit@10=%5.1f%% "
|
|
72
|
+
"MRR=%.4f%s" % (label, st["h1"] * 100.0 / n,
|
|
73
|
+
st["h5"] * 100.0 / n, st["h10"] * 100.0 / n,
|
|
74
|
+
st["rr"] / n, extra))
|
|
75
|
+
|
|
76
|
+
|
|
77
|
+
# 生效条件:qa、docs 给定时对 docs 每项按 jaccard(qa, na) 降序、id 升序排序并返回其 id 序列,docs 为空返回空列表,qa 为空时全部分数相同故只按 id 升序。
|
|
78
|
+
def full_rank(qa, docs):
|
|
79
|
+
"""全池 Jaccard 排序 → id 序列(与 run_arm 同判据)。"""
|
|
80
|
+
scored = sorted(((jaccard(qa, na), nid) for nid, na in docs),
|
|
81
|
+
key=lambda t: (-t[0], t[1]))
|
|
82
|
+
return [nid for _s, nid in scored]
|
|
83
|
+
|
|
84
|
+
|
|
85
|
+
# 生效条件:无 required 形参;以模块级常量 CORPUS、QUESTIONS 逐行 json.loads 建语料与题目,每题取 evidence_turns 与 question 分词结果,qa 为空则跳过且不计入 n,最终返回含 g0/g2/g1i/g1f/n/fa10_conservative/avg_pool 的统计字典(分母均以 max(·,1) 兜零)。
|
|
86
|
+
def exp_locomo():
|
|
87
|
+
"""实验一:原子面渐进三形态。"""
|
|
88
|
+
t0 = time.time()
|
|
89
|
+
with io.open(CORPUS, encoding="utf-8") as f:
|
|
90
|
+
corpus = [json.loads(ln) for ln in f if ln.strip()]
|
|
91
|
+
with io.open(QUESTIONS, encoding="utf-8") as f:
|
|
92
|
+
questions = [json.loads(ln) for ln in f if ln.strip()]
|
|
93
|
+
char_atoms = load_char_atoms()
|
|
94
|
+
docs = [(c["id"], node_atom_set(c, char_atoms)) for c in corpus]
|
|
95
|
+
docs_map = dict(docs)
|
|
96
|
+
print("[progressive] 实验一 locomo-zh-500 · 语料 %d / 题 %d"
|
|
97
|
+
% (len(corpus), len(questions)))
|
|
98
|
+
|
|
99
|
+
pool_sizes = [] # G1 各阶段候选池大小(统计面)
|
|
100
|
+
|
|
101
|
+
# 生效条件:used 给定时对其 frozenset 调 full_rank 取前 20 条,把该池大小追加进 pool_sizes,返回由这些 id 与 0.0 组成的二元组列表。
|
|
102
|
+
def rank_fn(used):
|
|
103
|
+
ranked = full_rank(frozenset(used), docs)[:20]
|
|
104
|
+
pool_sizes.append(len(ranked))
|
|
105
|
+
return [(nid, 0.0) for nid in ranked]
|
|
106
|
+
|
|
107
|
+
st_g0 = {"h1": 0, "h5": 0, "h10": 0, "rr": 0.0}
|
|
108
|
+
st_g2 = dict(st_g0)
|
|
109
|
+
st_i = dict(st_g0)
|
|
110
|
+
st_f = dict(st_g0)
|
|
111
|
+
steps_sum = conv = n = fa10_sum = 0
|
|
112
|
+
for q in questions:
|
|
113
|
+
ev = set(q.get("evidence_turns") or [])
|
|
114
|
+
qa = [w for w in zh_map_en(str(q.get("question") or ""),
|
|
115
|
+
char_atoms).split() if w]
|
|
116
|
+
if not qa:
|
|
117
|
+
continue
|
|
118
|
+
n += 1
|
|
119
|
+
# G0 一次性全原子(对照锚点)
|
|
120
|
+
_agg(rank_of(ev, full_rank(frozenset(qa), docs)), st_g0)
|
|
121
|
+
# G2 只宽检不细化
|
|
122
|
+
core = qa[:max(1, int(len(qa) * 0.5))]
|
|
123
|
+
_agg(rank_of(ev, full_rank(frozenset(core), docs)), st_g2)
|
|
124
|
+
# G1 渐进
|
|
125
|
+
tr = progressive_search(rank_fn, qa, docs_map,
|
|
126
|
+
ratio=0.5, k=10, max_stages=4)
|
|
127
|
+
_agg(rank_of(ev, full_rank(frozenset(tr["stages"][0]["used"]),
|
|
128
|
+
docs)), st_i)
|
|
129
|
+
fin_all = full_rank(frozenset(tr["stages"][-1]["used"]), docs)
|
|
130
|
+
_agg(rank_of(ev, fin_all), st_f)
|
|
131
|
+
# 保守 false_accept 口径:final top10 内非 gold 条数(本语料无资格
|
|
132
|
+
# 判定面,非 gold 全计为假接受——上界,见打印行口径注明)
|
|
133
|
+
fa10_sum += sum(1 for nid in fin_all[:10] if nid not in ev)
|
|
134
|
+
steps_sum += len(tr["added"])
|
|
135
|
+
conv += 1 if (tr["converged"] or len(tr["added"]) < 3) else 0
|
|
136
|
+
|
|
137
|
+
_show("G0 一次性全原子", st_g0, n,
|
|
138
|
+
" ← 锚点应≈② 96.8/99.8/99.8")
|
|
139
|
+
_show("G2 只宽检(core50%)", st_g2, n)
|
|
140
|
+
_show("G1 渐进·initial", st_i, n)
|
|
141
|
+
_show("G1 渐进·final", st_f, n)
|
|
142
|
+
print(" refinement: 平均加条件数=%.2f 一步收敛率=%.1f%% (n=%d, %.0fs)"
|
|
143
|
+
% (steps_sum / max(n, 1), conv * 100.0 / max(n, 1), n,
|
|
144
|
+
time.time() - t0))
|
|
145
|
+
print(" false_accept(保守口径)=%.2f%% =G1 final top10 非 gold 平均占比"
|
|
146
|
+
"(上界:本语料无 CCG 注释面,非 gold 全计为假接受)"
|
|
147
|
+
% (fa10_sum * 10.0 / max(n, 1)))
|
|
148
|
+
print(" 候选池: G1 各阶段平均 top 池大小=%.1f (k=20 上限)"
|
|
149
|
+
% (sum(pool_sizes) / max(len(pool_sizes), 1)))
|
|
150
|
+
print(" 四态分布: 本实验为原子面(语料无 CCG 注释,judge_qualification "
|
|
151
|
+
"将全判 BLINDSPOT)——资格判定不适用,不硬造数字;四态实证见实验二")
|
|
152
|
+
return {"g0": st_g0, "g2": st_g2, "g1i": st_i, "g1f": st_f, "n": n,
|
|
153
|
+
"fa10_conservative": fa10_sum * 10.0 / max(n, 1),
|
|
154
|
+
"avg_pool": sum(pool_sizes) / max(len(pool_sizes), 1)}
|
|
155
|
+
|
|
156
|
+
|
|
157
|
+
# 生效条件:无必需形参;自建临时库(tempfile.mkdtemp + MdCGOS)注入 GOLD/NEIGHBORS 语料后跑渐进消歧,finally 清理临时目录;返回指标 dict,不依赖外部数据;
|
|
158
|
+
def exp_controlled():
|
|
159
|
+
"""实验二:受控干扰池——条件冒充形态下渐进消歧(引擎侧真检索)。"""
|
|
160
|
+
print("\n[progressive] 实验二 受控干扰池(42+2 节点真实 MdCGOS)")
|
|
161
|
+
tmp = tempfile.mkdtemp(prefix="mdcg_prog_")
|
|
162
|
+
try:
|
|
163
|
+
root = os.path.join(tmp, "root")
|
|
164
|
+
cg = MdCGOS(root)
|
|
165
|
+
gold_ids = [g for g, _ in GOLD]
|
|
166
|
+
for gid, fact in GOLD:
|
|
167
|
+
cg.add(gid, _doc(fact, GOLD_COND), layer="knowledge",
|
|
168
|
+
verification_basis="test")
|
|
169
|
+
# 2 条邻居改用 gold 同款生效条件 → 正条件命中 → judge 判 ACCEPT(冒充)
|
|
170
|
+
IMPOSTORS = ("n1a", "n3a")
|
|
171
|
+
for gid, lst in NEIGHBORS.items():
|
|
172
|
+
for nid, fact, _cond in lst:
|
|
173
|
+
cg.add(nid, _doc(fact, GOLD_COND if nid in IMPOSTORS
|
|
174
|
+
else _cond), layer="knowledge",
|
|
175
|
+
verification_basis="test")
|
|
176
|
+
for nid, fact in NEG_TEXTS:
|
|
177
|
+
cg.add(nid, _doc(fact, "问没做的事", extra_neg=True),
|
|
178
|
+
layer="knowledge", verification_basis="test",
|
|
179
|
+
non_applicable_conditions=["刚才在干什么"])
|
|
180
|
+
for nid, fact, cond in UNREL_DOCS:
|
|
181
|
+
cg.add(nid, _doc(fact, cond), layer="knowledge",
|
|
182
|
+
verification_basis="test")
|
|
183
|
+
|
|
184
|
+
KW = dict(paths=("lexical", "bucket", "entity", "graph", "fuzzy"),
|
|
185
|
+
fusion="sum", judge_ranking=True)
|
|
186
|
+
|
|
187
|
+
# 生效条件:used 给定时,used 为假值(空列表等)则查询串仅为 QUERY,否则为 QUERY 加空格再加空格连接的 used,再经 cg.search_rrf(k=10, judge=True, **KW) 返回 (id, 分值) 列表。
|
|
188
|
+
def engine_rank(used):
|
|
189
|
+
res, _m = cg.search_rrf(QUERY + (" " + " ".join(used)
|
|
190
|
+
if used else ""),
|
|
191
|
+
k=10, judge=True, **KW)
|
|
192
|
+
return [(r[0]["id"], r[1]) for r in res]
|
|
193
|
+
|
|
194
|
+
# 生效条件:res 给定时遍历其前 10 项,跳过 r[0]["id"] 属于外层 gold_ids 的项,对余项取 r[2](len(r)<=2 时取空 dict)并仅在 (ji or {}).get("state")=="ACCEPT" 时把该 id 收入,返回收集到的 out 列表。
|
|
195
|
+
def false_accepts(res):
|
|
196
|
+
out = []
|
|
197
|
+
for r in res[:10]:
|
|
198
|
+
nid = r[0]["id"]
|
|
199
|
+
if nid in gold_ids:
|
|
200
|
+
continue
|
|
201
|
+
ji = r[2] if len(r) > 2 else {}
|
|
202
|
+
if (ji or {}).get("state") == "ACCEPT":
|
|
203
|
+
out.append(nid)
|
|
204
|
+
return out
|
|
205
|
+
|
|
206
|
+
# 生效条件:res 给定时遍历其前 10 项,len(r)>2 时取 (r[2] or {}).get("state")、否则取 None,该值属于 ACCEPT/DEFER/REJECT/BLINDSPOT 四键之一则对应计数、否则计入「无judge」,返回只保留计数非零项的字典(全零则返回空 dict)。
|
|
207
|
+
def state_dist(res):
|
|
208
|
+
"""渐进循环四态分布(top10 内 judge state 计数,只列非零项)。
|
|
209
|
+
|
|
210
|
+
judge info 直接复用 search_rrf(judge=True) 注入的
|
|
211
|
+
judge_qualification 产物(结果第三元),不重复判定。
|
|
212
|
+
"""
|
|
213
|
+
d = {"ACCEPT": 0, "DEFER": 0, "REJECT": 0, "BLINDSPOT": 0,
|
|
214
|
+
"无judge": 0}
|
|
215
|
+
for r in res[:10]:
|
|
216
|
+
st = (r[2] or {}).get("state") if len(r) > 2 else None
|
|
217
|
+
d[st if st in d else "无judge"] += 1
|
|
218
|
+
return {k: v for k, v in d.items() if v}
|
|
219
|
+
|
|
220
|
+
# 生效条件:res、fa 给定时返回 100.0*len(fa)/max(1, min(10, len(res))),即 len(res)>=10 时除以 10、不足时除以 len(res),len(res) 为 0 时被 max(1,·) 兜为除数 1。
|
|
221
|
+
def fa_rate(res, fa):
|
|
222
|
+
return 100.0 * len(fa) / max(1, min(10, len(res)))
|
|
223
|
+
|
|
224
|
+
# 宽检索(不加任何条件;judgeinfo 直接取 results 第三元)
|
|
225
|
+
wide_res, _m = cg.search_rrf(QUERY, k=10, judge=True, **KW)
|
|
226
|
+
wide = [(r[0]["id"], r[1]) for r in wide_res]
|
|
227
|
+
fa_pre = false_accepts(wide_res)
|
|
228
|
+
top1_pre = wide[0][0] if wide else None
|
|
229
|
+
print(" 宽检索 top1=%s top10 内 False ACCEPT=%s(率=%.1f%%)"
|
|
230
|
+
% (top1_pre, fa_pre or "无", fa_rate(wide_res, fa_pre)))
|
|
231
|
+
print(" 四态分布(宽检索 top10)=%s 候选池=%d judge 剔除=%s"
|
|
232
|
+
% (state_dist(wide_res), len(wide_res),
|
|
233
|
+
(_m or {}).get("judge_filtered", 0)))
|
|
234
|
+
|
|
235
|
+
# 渐进:候选间区分性条件词 → 带条件重排
|
|
236
|
+
POOL = ["我", "喝水", "吃饭", "电脑", "书", "音乐", "出门",
|
|
237
|
+
"浇花", "下班"]
|
|
238
|
+
cand_chars = {}
|
|
239
|
+
for r in wide_res[:10]:
|
|
240
|
+
nid = r[0]["id"]
|
|
241
|
+
cand_chars[nid] = frozenset(str(r[0].get("content") or ""))
|
|
242
|
+
cond = discriminative([p for p in POOL],
|
|
243
|
+
[cand_chars.get(nid, frozenset())
|
|
244
|
+
for nid, _s in wide[:10]])
|
|
245
|
+
if cond:
|
|
246
|
+
refined_res, _m2 = cg.search_rrf(QUERY + " " + cond,
|
|
247
|
+
k=10, judge=True, **KW)
|
|
248
|
+
fa_post = false_accepts(refined_res)
|
|
249
|
+
top1_post = refined_res[0][0]["id"] if refined_res else None
|
|
250
|
+
print(" 区分性条件=%r → 重排 top1=%s top10 内 "
|
|
251
|
+
"False ACCEPT=%s(率=%.1f%%)"
|
|
252
|
+
% (cond, top1_post, fa_post or "无",
|
|
253
|
+
fa_rate(refined_res, fa_post)))
|
|
254
|
+
print(" 四态分布(重排 top10)=%s 候选池=%d judge 剔除=%s"
|
|
255
|
+
% (state_dist(refined_res), len(refined_res),
|
|
256
|
+
(_m2 or {}).get("judge_filtered", 0)))
|
|
257
|
+
ok1 = top1_post in gold_ids
|
|
258
|
+
no_new = set(fa_post) <= set(fa_pre)
|
|
259
|
+
print(" 断言①渐进后 top1∈gold:%s" % ("PASS" if ok1 else "FAIL"))
|
|
260
|
+
print(" 断言②渐进不引入新冒充(fa_post⊆fa_pre):%s"
|
|
261
|
+
% ("PASS" if no_new else "FAIL"))
|
|
262
|
+
if fa_post:
|
|
263
|
+
for r in refined_res[:10]:
|
|
264
|
+
if r[0]["id"] in fa_post and len(r) > 2:
|
|
265
|
+
print(" 冒充 %s judge reason:%s"
|
|
266
|
+
% (r[0]["id"],
|
|
267
|
+
(r[2] or {}).get("reason", "")))
|
|
268
|
+
print(" 边界(如实):条件面冒充(生效条件被伪造为 gold 同款)"
|
|
269
|
+
"不被内容词渐进消除——词面条件确认是必要非充分,"
|
|
270
|
+
"真伪判据=verification_basis/时间线冲突(证据强度分离,另案)")
|
|
271
|
+
return ok1 and no_new
|
|
272
|
+
print(" 无区分性条件可选(宽检索已无冒充)——如实记录")
|
|
273
|
+
return not fa_pre
|
|
274
|
+
finally:
|
|
275
|
+
shutil.rmtree(tmp, ignore_errors=True)
|
|
276
|
+
|
|
277
|
+
|
|
278
|
+
# 生效条件:无 required 形参;先执行 exp_locomo(),再按 exp_controlled() 返回值 ok 的真假返回 0(真)或 1(假)。
|
|
279
|
+
def main():
|
|
280
|
+
exp_locomo()
|
|
281
|
+
ok = exp_controlled()
|
|
282
|
+
print("\n[progressive] 完成。受控池消歧断言:%s"
|
|
283
|
+
% ("PASS" if ok else "FAIL"))
|
|
284
|
+
return 0 if ok else 1
|
|
285
|
+
|
|
286
|
+
|
|
287
|
+
if __name__ == "__main__":
|
|
288
288
|
raise SystemExit(main())
|