@furongjun1999/dsh-memory 0.8.1 → 0.9.6

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.
Files changed (714) hide show
  1. package/README.md +662 -659
  2. package/codebuddy/README.md +1 -1
  3. package/data/policy.json +17 -6
  4. package/docs/README.md +18 -2
  5. package/docs/discipline/harnesses.yaml +1 -1
  6. package/docs/discipline/templates/spec.txt.tmpl +23 -0
  7. package/docs//350/247/204/350/214/203/AI/347/274/226/347/250/213/351/200/232/347/224/250/350/247/204/350/214/203.md +179 -0
  8. package/docs//350/247/204/350/214/203/AI/347/274/226/347/250/213/351/200/232/347/224/250/350/247/204/350/214/203_/347/201/265/346/236/242/351/200/202/351/205/215.md +45 -0
  9. package/docs//350/247/204/350/214/203/AI/347/274/226/347/250/213/351/200/232/347/224/250/350/247/204/350/214/203_/350/207/252/346/243/200/346/270/205/345/215/225.md +69 -0
  10. package/dsh/README.md +1 -1
  11. package/dsh/cordis.yml.example +1 -1
  12. package/hive/README.md +705 -0
  13. package/hive/id_charset_blocks.txt +147 -0
  14. package/lib/bridge.d.ts +3 -0
  15. package/lib/bridge.js +32 -11
  16. package/lib/cli.js +14 -17
  17. package/lib/hooks.d.ts +18 -3
  18. package/lib/hooks.js +147 -18
  19. package/lib/index.d.ts +2 -2
  20. package/lib/index.js +18 -4
  21. package/lib/init.js +21 -21
  22. package/lib/lib/datapath.d.ts +56 -0
  23. package/lib/lib/datapath.js +102 -9
  24. package/lib/lib/hook_audit.d.ts +0 -1
  25. package/lib/lib/hook_audit.js +13 -1
  26. package/lib/lib/mdcg_client.d.ts +4 -1
  27. package/lib/lib/mdcg_client.js +10 -2
  28. package/lib/lib/mutual.js +61 -15
  29. package/lib/lib/prompt_safety.js +16 -1
  30. package/lib/lib/roleplay_web.js +30 -2
  31. package/lib/lib/token_store.d.ts +25 -6
  32. package/lib/lib/token_store.js +224 -33
  33. package/lib/tools.d.ts +8 -2
  34. package/lib/tools.js +22 -3
  35. package/md_cg/_guard_kit.py +121 -0
  36. package/md_cg/admission.py +718 -718
  37. package/md_cg/anchor.py +114 -0
  38. package/md_cg/audit.py +22 -5
  39. package/md_cg/autonomy.py +433 -288
  40. package/md_cg/autonomy_modes.py +781 -642
  41. package/md_cg/bench_beir_bm25.py +154 -0
  42. package/md_cg/bench_beir_scifact.py +170 -0
  43. package/md_cg/bench_bm25_blockmax.py +134 -0
  44. package/md_cg/bench_bm25_incremental.py +246 -0
  45. package/md_cg/bench_ccg_heading.py +118 -0
  46. package/md_cg/bench_csre_correctness.py +145 -0
  47. package/md_cg/bench_excerpt_budget.py +94 -0
  48. package/md_cg/bench_fuzzy_substrings.py +299 -0
  49. package/md_cg/bench_memory_rsi.py +358 -0
  50. package/md_cg/bench_memory_rsi_latency.py +82 -0
  51. package/md_cg/bench_memory_rsi_public.py +116 -0
  52. package/md_cg/bench_memory_rsi_tools.py +229 -0
  53. package/md_cg/bench_rrf_candidates.py +212 -0
  54. package/md_cg/bench_rrf_public.py +100 -0
  55. package/md_cg/bench_text_projections.py +221 -0
  56. package/md_cg/bench_timeanchor.py +357 -0
  57. package/md_cg/bm25.py +601 -0
  58. package/md_cg/branches.py +404 -327
  59. package/md_cg/build_postings.py +74 -73
  60. package/md_cg/ccgc.py +4 -2
  61. package/md_cg/chain.py +572 -379
  62. package/md_cg/coldverify.py +531 -292
  63. package/md_cg/condaccept.py +263 -0
  64. package/md_cg/condgate.py +308 -0
  65. package/md_cg/config_registry.py +329 -0
  66. package/md_cg/config_registry_bulk.py +482 -0
  67. package/md_cg/config_registry_guard.py +487 -0
  68. package/md_cg/config_render.py +354 -0
  69. package/md_cg/config_validate.py +455 -0
  70. package/md_cg/conformance.py +21 -0
  71. package/md_cg/conformance_core.py +753 -0
  72. package/md_cg/consistency.py +1278 -955
  73. package/md_cg/consolidate.py +180 -20
  74. package/md_cg/crypto.py +794 -472
  75. package/md_cg/datapath.py +137 -12
  76. package/md_cg/eval_common.py +522 -575
  77. package/md_cg/evidence.py +946 -585
  78. package/md_cg/evidence_delivery.py +101 -0
  79. package/md_cg/experience_consolidation.py +317 -0
  80. package/md_cg/export.py +329 -222
  81. package/md_cg/forgetting.py +1050 -922
  82. package/md_cg/freshness.py +601 -548
  83. package/md_cg/fsutil.py +1120 -824
  84. package/md_cg/goal_gen.py +458 -0
  85. package/md_cg/hotcache.py +107 -9
  86. package/md_cg/hyperedge.py +14 -3
  87. package/md_cg/identity.py +573 -391
  88. package/md_cg/imgskill.py +1712 -0
  89. package/md_cg/insight.py +540 -521
  90. package/md_cg/interface_schema.py +146 -0
  91. package/md_cg/interop.py +20 -0
  92. package/md_cg/judgment_manifest.py +24 -1
  93. package/md_cg/lifecycle.py +86 -5
  94. package/md_cg/links.py +27 -6
  95. package/md_cg/logref.py +332 -327
  96. package/md_cg/mcp_server.py +5227 -4449
  97. package/md_cg/mdcg.py +1301 -203
  98. package/md_cg/mdcos.py +7104 -5440
  99. package/md_cg/memory_rsi.py +308 -0
  100. package/md_cg/metacognition.py +726 -591
  101. package/md_cg/nodefile.py +151 -3
  102. package/md_cg/op_table.py +1128 -0
  103. package/md_cg/postings.py +524 -300
  104. package/md_cg/predict.py +4 -1
  105. package/md_cg/probe_memory_rsi.py +95 -0
  106. package/md_cg/probe_memory_rsi_mcp.py +66 -0
  107. package/md_cg/protect.py +188 -28
  108. package/md_cg/protocol.py +492 -411
  109. package/md_cg/provenance.py +7 -1
  110. package/md_cg/qmd.py +477 -0
  111. package/md_cg/qmd_llm.py +334 -0
  112. package/md_cg/query_excerpt.py +171 -0
  113. package/md_cg/queue_health.py +130 -0
  114. package/md_cg/reach.py +470 -453
  115. package/md_cg/readcache.py +115 -3
  116. package/md_cg/reconcile.py +239 -228
  117. package/md_cg/refindex.py +1191 -1042
  118. package/md_cg/refine.py +12 -3
  119. package/md_cg/retrieve.py +128 -0
  120. package/md_cg/review_cli.py +204 -26
  121. package/md_cg/revise.py +140 -0
  122. package/md_cg/rollback.py +256 -19
  123. package/md_cg/rollback_cli.py +21 -1
  124. package/md_cg/rotate.py +347 -0
  125. package/md_cg/routing.py +62 -1
  126. package/md_cg/rrf_candidates.py +204 -0
  127. package/md_cg/run_tests.py +19 -1
  128. package/md_cg/scrub.py +177 -23
  129. package/md_cg/security.py +269 -17
  130. package/md_cg/self_state.py +1053 -1029
  131. package/md_cg/self_surface.py +50 -0
  132. package/md_cg/semantic/canonical.py +1 -1
  133. package/md_cg/semantic/export_en_zh_map.py +2 -2
  134. package/md_cg/semantic/zh_en_atoms.py +2 -0
  135. package/md_cg/session5.py +106 -0
  136. package/md_cg/signer.py +76 -2
  137. package/md_cg/sleep.py +1779 -1583
  138. package/md_cg/sources.py +273 -15
  139. package/md_cg/state_events.py +75 -10
  140. package/md_cg/state_slots.py +312 -221
  141. package/md_cg/stg.py +82 -16
  142. package/md_cg/stgidx.py +537 -281
  143. package/md_cg/subgraph.py +851 -729
  144. package/md_cg/sustain.py +458 -19
  145. package/md_cg/task_memory.py +367 -0
  146. package/md_cg/test_action_derive.py +210 -203
  147. package/md_cg/test_audit_heartbeat_20261015.py +417 -0
  148. package/md_cg/test_audit_protection_20261015.py +483 -0
  149. package/md_cg/test_audit_rotate_parity.py +280 -0
  150. package/md_cg/test_audit_state_20261015.py +253 -0
  151. package/md_cg/test_audit_write_20261015.py +161 -0
  152. package/md_cg/test_auto_defaults.py +43 -4
  153. package/md_cg/test_autonomy_emotion.py +328 -0
  154. package/md_cg/test_autonomy_modes.py +43 -3
  155. package/md_cg/test_b1_foreign_read.py +245 -0
  156. package/md_cg/test_bm25.py +438 -0
  157. package/md_cg/test_bm25_hydration_refill.py +265 -0
  158. package/md_cg/test_bm25_incremental.py +241 -0
  159. package/md_cg/test_bm25_incremental_mutations.py +160 -0
  160. package/md_cg/test_bm25_mutations.py +205 -0
  161. package/md_cg/test_bm25_pool_view.py +196 -0
  162. package/md_cg/test_boot_openblas_guard.py +218 -197
  163. package/md_cg/test_boundary_hit.py +1 -1
  164. package/md_cg/test_branches_sensitivity.py +375 -0
  165. package/md_cg/test_c8_search_rrf_gates.py +1 -1
  166. package/md_cg/test_canonical_timezones.py +211 -0
  167. package/md_cg/test_ccg_heading_parser.py +73 -0
  168. package/md_cg/test_chain_best_first.py +403 -0
  169. package/md_cg/test_chain_candidate_scope.py +227 -0
  170. package/md_cg/test_cn_recall_grams.py +206 -0
  171. package/md_cg/test_coldverify_drain.py +127 -0
  172. package/md_cg/test_condaccept.py +156 -0
  173. package/md_cg/test_condgate.py +141 -0
  174. package/md_cg/test_config_registry.py +55 -0
  175. package/md_cg/test_crypto_rotate.py +257 -0
  176. package/md_cg/test_csre_correctness.py +297 -0
  177. package/md_cg/test_csre_index_import_guard.py +161 -0
  178. package/md_cg/test_csre_mutations.py +34 -0
  179. package/md_cg/test_csre_session.py +69 -0
  180. package/md_cg/test_curiosity_budget.py +468 -0
  181. package/md_cg/test_d17_autostart_toolface.py +453 -0
  182. package/md_cg/test_datapath_tilde_home.py +271 -0
  183. package/md_cg/test_derived_cache_generations.py +256 -0
  184. package/md_cg/test_evidence_delivery_compact.py +54 -0
  185. package/md_cg/test_exact_text_cache.py +191 -0
  186. package/md_cg/test_excerpt_budget.py +188 -0
  187. package/md_cg/test_experience_consolidation.py +112 -0
  188. package/md_cg/test_export_default_collision.py +138 -0
  189. package/md_cg/test_export_snapshots.py +322 -0
  190. package/md_cg/test_fuzzy_substring_equivalence.py +127 -0
  191. package/md_cg/test_gain_gate.py +16 -5
  192. package/md_cg/test_goal_gen.py +355 -0
  193. package/md_cg/test_h2_session_view_norm.py +251 -233
  194. package/md_cg/test_h4_sustain_snapshot.py +1490 -1387
  195. package/md_cg/test_heartbeat_ledger.py +557 -0
  196. package/md_cg/test_hierarchy_audit_graph.py +237 -0
  197. package/md_cg/test_i50c_meta_passthrough.py +628 -608
  198. package/md_cg/test_i50e_readside_protection.py +6 -4
  199. package/md_cg/test_identity_attribution.py +96 -2
  200. package/md_cg/test_identity_contract.py +329 -0
  201. package/md_cg/test_imgskill.py +1444 -0
  202. package/md_cg/test_index_snapshot_shape.py +240 -0
  203. package/md_cg/test_interop_judgment.py +258 -228
  204. package/md_cg/test_issue111_gate_order.py +77 -0
  205. package/md_cg/test_issue111_scoring_contract.py +87 -0
  206. package/md_cg/test_issue121_calibration_bounds.py +293 -0
  207. package/md_cg/test_issue128_session_env_authority.py +246 -0
  208. package/md_cg/test_issue130_rrf_rank_credibility.py +217 -0
  209. package/md_cg/test_issue131_postings_content_staleness.py +424 -0
  210. package/md_cg/test_issue145_sleep_no_silent_rollback.py +168 -0
  211. package/md_cg/test_issue146_frontmatter_quarantine.py +428 -0
  212. package/md_cg/test_issue147_pack_id_escape.py +378 -0
  213. package/md_cg/test_issue148_merge_downgrade_leak.py +192 -0
  214. package/md_cg/test_issue149_accel_default_face.py +249 -0
  215. package/md_cg/test_issue150_empty_query_terms.py +328 -0
  216. package/md_cg/test_issue151_153_157_write_cache_index.py +152 -0
  217. package/md_cg/test_issue151b_base_add_importance_guard.py +166 -0
  218. package/md_cg/test_issue152_sealed_postings.py +309 -0
  219. package/md_cg/test_issue152b_decrypt_cache.py +218 -0
  220. package/md_cg/test_issue153b_cache_alias_and_canon.py +236 -0
  221. package/md_cg/test_issue153b_index_generation_invalidation.py +143 -0
  222. package/md_cg/test_issue157b_index_remainder.py +344 -0
  223. package/md_cg/test_issue189_conclusion_bidirectional.py +223 -0
  224. package/md_cg/test_issue190_tidy_idempotent.py +806 -0
  225. package/md_cg/test_issue191_import_trust.py +382 -0
  226. package/md_cg/test_issue192_move_layer_indices_guard.py +258 -0
  227. package/md_cg/test_issue193_consistency_knowledge_guard.py +524 -0
  228. package/md_cg/test_issue193_write_scope.py +378 -0
  229. package/md_cg/test_issue197_rank_credibility_property.py +201 -0
  230. package/md_cg/test_issue211_export_target.py +325 -0
  231. package/md_cg/test_issue212_writeface_gaps.py +349 -0
  232. package/md_cg/test_issue214_reject_sensitivity.py +267 -0
  233. package/md_cg/test_issue215_condition_source_gap.py +272 -0
  234. package/md_cg/test_issue216_clash_routes.py +214 -0
  235. package/md_cg/test_issue217_slot_routes.py +248 -0
  236. package/md_cg/test_issue218_postcondition_value_blind.py +222 -0
  237. package/md_cg/test_issue225_consistency_torn_log.py +433 -0
  238. package/md_cg/test_issue245_swarm_traversal.py +366 -0
  239. package/md_cg/test_issue258_pack_sensitivity.py +468 -0
  240. package/md_cg/test_issue265_coldverify_multiproc.py +438 -0
  241. package/md_cg/test_issue274_read_node_id_none.py +341 -0
  242. package/md_cg/test_issue275_layer_param.py +250 -0
  243. package/md_cg/test_issue277_node_id_traversal.py +384 -0
  244. package/md_cg/test_issue278_neg_separated_tension.py +271 -0
  245. package/md_cg/test_issue278_neg_window.py +294 -0
  246. package/md_cg/test_issue278_negative_layering.py +83 -0
  247. package/md_cg/test_issue279_signers_path_guard.py +652 -0
  248. package/md_cg/test_issue284_skip_not_pass.py +291 -0
  249. package/md_cg/test_issue294_flywheel_sensitivity.py +573 -0
  250. package/md_cg/test_issue295_sleep_root_binding.py +589 -0
  251. package/md_cg/test_issue297_rewrite_sensitivity.py +454 -0
  252. package/md_cg/test_issue298_consistency_ledger_sensitivity.py +421 -0
  253. package/md_cg/test_issue301_baseline_corpus_gate.py +245 -0
  254. package/md_cg/test_issue303_design_boundaries.py +262 -0
  255. package/md_cg/test_issue313_writelimit_sensitivity.py +404 -0
  256. package/md_cg/test_issue315_merge_branch_sensitivity.py +384 -0
  257. package/md_cg/test_issue315_sensitivity_passthrough_guard.py +514 -0
  258. package/md_cg/test_issue319_tenant_unregistered_bypass.py +257 -0
  259. package/md_cg/test_issue320_vstate_requires_verify.py +279 -0
  260. package/md_cg/test_issue321_mcp_deep_json.py +305 -0
  261. package/md_cg/test_issue322_node_write_lock.py +727 -0
  262. package/md_cg/test_issue323_aggregate_segment_boundary.py +365 -0
  263. package/md_cg/test_issue324_write_size_fields.py +215 -0
  264. package/md_cg/test_issue325_audit_poison.py +230 -0
  265. package/md_cg/test_issue326_append_jsonl_oserror.py +464 -0
  266. package/md_cg/test_issue326_ingest_auto_whitelist.py +331 -0
  267. package/md_cg/test_issue326_key_perms.py +355 -0
  268. package/md_cg/test_issue326_twin_parser_parity.py +323 -0
  269. package/md_cg/test_issue326_vm_call_frame.py +569 -0
  270. package/md_cg/test_issue345_writelimit_effective_sens.py +234 -0
  271. package/md_cg/test_issue346_overwrite_sensitivity_gate.py +424 -0
  272. package/md_cg/test_issue346_review_overwrite_guard.py +509 -0
  273. package/md_cg/test_issue348_insight_verify_op.py +200 -0
  274. package/md_cg/test_issue348_merge_rmw_lock.py +345 -0
  275. package/md_cg/test_issue350_casefold_collision.py +236 -0
  276. package/md_cg/test_issue351_ingest_root_fallback.py +202 -0
  277. package/md_cg/test_issue352_imgaudit.py +418 -0
  278. package/md_cg/test_issue352_pathsjson.py +537 -0
  279. package/md_cg/test_issue352_refindex.py +412 -0
  280. package/md_cg/test_issue352_rollback.py +604 -0
  281. package/md_cg/test_issue352_vstateinherit.py +196 -0
  282. package/md_cg/test_issue354_p1_writelimit_type.py +522 -0
  283. package/md_cg/test_issue354_p2_datapath_clock.py +381 -0
  284. package/md_cg/test_issue354_p2_writelimit_toctou.py +327 -0
  285. package/md_cg/test_issue354_rust_index_containment.py +571 -0
  286. package/md_cg/test_issue354_rust_wal_containment.py +606 -0
  287. package/md_cg/test_issue358_truncation_observable.py +247 -0
  288. package/md_cg/test_issue360_tidy_private_excerpt.py +196 -0
  289. package/md_cg/test_issue361_blindspot_not_silent_drop.py +494 -0
  290. package/md_cg/test_issue362_neg_condition.py +468 -0
  291. package/md_cg/test_issue363_lexical_idf.py +275 -0
  292. package/md_cg/test_issue370_readcache_assembly.py +488 -0
  293. package/md_cg/test_issue375_state_slot_retire.py +325 -0
  294. package/md_cg/test_issue376_heartbeat_failed.py +293 -0
  295. package/md_cg/test_issue377_layer_protect_stamp.py +324 -0
  296. package/md_cg/test_issue378_protected_floor.py +310 -0
  297. package/md_cg/test_issue383_state_event_write_gate.py +462 -0
  298. package/md_cg/test_issue397_sens_eff_wiring.py +435 -0
  299. package/md_cg/test_issue398_toolface_op_parity.py +511 -0
  300. package/md_cg/test_issue399_hive_json_shape.py +317 -0
  301. package/md_cg/test_issue400_hive_dirty_args.py +450 -0
  302. package/md_cg/test_issue401_wisdom_codegen.py +295 -0
  303. package/md_cg/test_issue404_heal_receipt.py +390 -0
  304. package/md_cg/test_issue405_git_pathspec.py +333 -0
  305. package/md_cg/test_issue408_eface_registry.py +291 -0
  306. package/md_cg/test_issue409_compact_index_completeness.py +509 -0
  307. package/md_cg/test_issue410_verify_demote_write.py +366 -0
  308. package/md_cg/test_issue411_scrub_idempotence.py +575 -0
  309. package/md_cg/test_issue412_discipline_producer_parity.py +312 -0
  310. package/md_cg/test_issue413_preimage_failclosed.py +239 -0
  311. package/md_cg/test_issue413_preimage_twin.py +340 -0
  312. package/md_cg/test_issue414_write_args_parity.py +613 -0
  313. package/md_cg/test_issue415_deps_gate_probe.py +358 -0
  314. package/md_cg/test_issue416_entry_key_subset.py +525 -0
  315. package/md_cg/test_issue417_twophase_merge_landed.py +423 -0
  316. package/md_cg/test_issue418_identity_role_type_gate.py +188 -0
  317. package/md_cg/test_issue418_role_type_gate.py +265 -0
  318. package/md_cg/test_issue419_pii_redaction.py +642 -0
  319. package/md_cg/test_issue424_merge_confirm_sensitivity.py +381 -0
  320. package/md_cg/test_issue425_ingest_library_path.py +331 -0
  321. package/md_cg/test_issue426a_dedup_receipt.py +242 -0
  322. package/md_cg/test_issue426b_root_clash_invisible.py +270 -0
  323. package/md_cg/test_issue426c_enqueue_not_silent.py +346 -0
  324. package/md_cg/test_issue426e_restore_overwrite_live.py +310 -0
  325. package/md_cg/test_issue426f_paths_gbk_failclosed.py +301 -0
  326. package/md_cg/test_issue427_serve_status_shape.py +191 -0
  327. package/md_cg/test_issue431_432_468_protect_single_point.py +322 -0
  328. package/md_cg/test_issue433_outlook_threshold_single_point.py +278 -0
  329. package/md_cg/test_issue434_435_float_gate.py +90 -0
  330. package/md_cg/test_issue439_sibling_sensitivity_inherit.py +448 -0
  331. package/md_cg/test_issue43_default_policy.py +13 -4
  332. package/md_cg/test_issue443a_state_event_length_gate.py +94 -0
  333. package/md_cg/test_issue443b_audit_record_length_gate.py +83 -0
  334. package/md_cg/test_issue443c_self_state_imgskill_length_gate.py +77 -0
  335. package/md_cg/test_issue444_scrub_evidence_plaintext.py +281 -0
  336. package/md_cg/test_issue445_metacognition_ledger_gate.py +454 -0
  337. package/md_cg/test_issue446_453_403_verified_evidence.py +382 -0
  338. package/md_cg/test_issue446_is_encrypted_envelope.py +166 -0
  339. package/md_cg/test_issue447_recent_sens_failclosed.py +270 -0
  340. package/md_cg/test_issue450_sensitivity_fallback_reason.py +51 -0
  341. package/md_cg/test_issue450_writegate_unknown_sens.py +203 -0
  342. package/md_cg/test_issue451_audit_actor.py +313 -0
  343. package/md_cg/test_issue452_ingest_root_hint.py +238 -0
  344. package/md_cg/test_issue452_ingest_root_sensitivity.py +299 -0
  345. package/md_cg/test_issue463_ghost_prune_guard.py +259 -0
  346. package/md_cg/test_issue466_basis_registry.py +469 -0
  347. package/md_cg/test_issue467_identity_sensitivity.py +204 -0
  348. package/md_cg/test_issue467_identity_sensitivity_parity.py +361 -0
  349. package/md_cg/test_issue469_scrub_override_passthrough.py +126 -0
  350. package/md_cg/test_issue470_wal_event_durable.py +427 -0
  351. package/md_cg/test_issue471_identity_readable.py +137 -0
  352. package/md_cg/test_issue472_audit_chain.py +91 -0
  353. package/md_cg/test_issue472_prune_ledger.py +106 -0
  354. package/md_cg/test_issue472_rollback_degraded.py +82 -0
  355. package/md_cg/test_issue491_compose_sandbox.py +460 -0
  356. package/md_cg/test_issue492_dup_readout_opt_in.py +292 -0
  357. package/md_cg/test_issue498_merge_trace_stale_index.py +280 -0
  358. package/md_cg/test_issue511_scoreboard_bad_files.py +256 -0
  359. package/md_cg/test_issue521_rotate_lock_suffix.py +110 -0
  360. package/md_cg/test_issue522_review_record_sensitivity.py +267 -0
  361. package/md_cg/test_issue524_orphan_noise_no_demote.py +88 -0
  362. package/md_cg/test_issue525_read_null_semantics.py +52 -0
  363. package/md_cg/test_issue538_hook_condition_space.py +97 -0
  364. package/md_cg/test_issue540_path_guard_whitespace.py +126 -0
  365. package/md_cg/test_issue541_tidy_protect_gate.py +141 -0
  366. package/md_cg/test_issue542_tenant_root_same.py +153 -0
  367. package/md_cg/test_issue543_issue_authorization.py +147 -0
  368. package/md_cg/test_issue544_sandbox_reflection_gate.py +175 -0
  369. package/md_cg/test_issue545_charset_vs_size_diagnostic.py +77 -0
  370. package/md_cg/test_issue546_ledger_concurrency.py +81 -0
  371. package/md_cg/test_issue547_append_jsonl_whitelist.py +115 -0
  372. package/md_cg/test_issue547_audit_durable.py +96 -0
  373. package/md_cg/test_issue549_durable_append.py +143 -0
  374. package/md_cg/test_issue557_r1_credential_value_forms.py +76 -0
  375. package/md_cg/test_issue557_r2_write_entries_size_gate.py +84 -0
  376. package/md_cg/test_issue557_r3_pem_block_redaction.py +75 -0
  377. package/md_cg/test_issue559_refine_prefix.py +70 -0
  378. package/md_cg/test_issue561_pending_review.py +109 -0
  379. package/md_cg/test_issue66_scrub_semantics.py +107 -0
  380. package/md_cg/test_issue67_scrub_control_pool.py +233 -0
  381. package/md_cg/test_issue69_audit_boundary.py +129 -0
  382. package/md_cg/test_issue71_queue_backlog.py +377 -0
  383. package/md_cg/test_issue72_coldverify_controlled_drain.py +420 -0
  384. package/md_cg/test_issue73_doc_chapter_refs.py +326 -0
  385. package/md_cg/test_issue75_neural_failure.py +582 -0
  386. package/md_cg/test_issue78_export_concurrency.py +334 -0
  387. package/md_cg/test_issue80_residual_flush.py +169 -0
  388. package/md_cg/test_issue82_credential_rules.py +271 -0
  389. package/md_cg/test_issue82_layer_and_size.py +120 -0
  390. package/md_cg/test_issue84_portability.py +95 -0
  391. package/md_cg/test_issue85_export_overwrite.py +113 -0
  392. package/md_cg/test_issue86_runtime_revoke.py +110 -0
  393. package/md_cg/test_issue87_default_sensitivity_single.py +201 -0
  394. package/md_cg/test_issue88_conclusion_face.py +280 -0
  395. package/md_cg/test_issue88_execution_conflicts.py +87 -0
  396. package/md_cg/test_issue89_zero_score.py +169 -0
  397. package/md_cg/test_issue89a_score_cap_tiebreak.py +279 -0
  398. package/md_cg/test_issue91_verifier_sandbox.py +314 -0
  399. package/md_cg/test_issue91_verifier_timeout.py +528 -0
  400. package/md_cg/test_issue93_remember_redaction.py +282 -0
  401. package/md_cg/test_issue96_key_fallback.py +102 -0
  402. package/md_cg/test_issue98_tenant_root.py +104 -0
  403. package/md_cg/test_jsonl_commit_watermark.py +221 -0
  404. package/md_cg/test_live_validity_cache.py +256 -0
  405. package/md_cg/test_logref.py +32 -0
  406. package/md_cg/test_mcp_call_argument_types.py +99 -0
  407. package/md_cg/test_md_access_parity.py +80 -5
  408. package/md_cg/test_md_cg_no_hive_dep.py +140 -0
  409. package/md_cg/test_memory_rsi.py +215 -0
  410. package/md_cg/test_memory_rsi_adversarial.py +91 -0
  411. package/md_cg/test_mode_parity.py +1413 -1393
  412. package/md_cg/test_mutation_residue.py +120 -0
  413. package/md_cg/test_mutation_rollback.py +62 -32
  414. package/md_cg/test_n123_derive_expiry_chain.py +10 -5
  415. package/md_cg/test_n130_verify_falsified_protect.py +185 -185
  416. package/md_cg/test_n209_verify_write_face_gates.py +7 -2
  417. package/md_cg/test_n230_dirty_replay.py +52 -9
  418. package/md_cg/test_n62_tenant_bind_failclosed.py +3 -3
  419. package/md_cg/test_neg_coverage_budget.py +500 -0
  420. package/md_cg/test_neg_tail_honesty.py +79 -9
  421. package/md_cg/test_p0.py +16 -2
  422. package/md_cg/test_p14_root_fallback.py +373 -0
  423. package/md_cg/test_p1x_ref_root.py +47 -10
  424. package/md_cg/test_p21_tokens.py +2 -1
  425. package/md_cg/test_p24_evidence.py +232 -227
  426. package/md_cg/test_p26_refindex.py +20 -3
  427. package/md_cg/test_p27_docindex.py +16 -1
  428. package/md_cg/test_p28_refcheck.py +45 -24
  429. package/md_cg/test_p29_session_ingest_export.py +6 -0
  430. package/md_cg/test_p2_mcp.py +474 -453
  431. package/md_cg/test_p2_six_elements.py +14 -1
  432. package/md_cg/test_p3.py +370 -261
  433. package/md_cg/test_p33_ccg_wiring.py +9 -2
  434. package/md_cg/test_p3_legacy_closure.py +6 -1
  435. package/md_cg/test_p40_refine_worklist.py +5 -2
  436. package/md_cg/test_p47_session_view.py +25 -9
  437. package/md_cg/test_p4_freshness.py +8 -1
  438. package/md_cg/test_p83_jsonrpc_parse_error.py +124 -0
  439. package/md_cg/test_p9c_dedup_hints.py +7 -2
  440. package/md_cg/test_policy_required_ccg.py +12 -0
  441. package/md_cg/test_postings_digest_memo.py +139 -0
  442. package/md_cg/test_predict_beta.py +138 -135
  443. package/md_cg/test_qmd.py +500 -0
  444. package/md_cg/test_query_excerpt.py +209 -0
  445. package/md_cg/test_reach.py +456 -378
  446. package/md_cg/test_reach_snapshot_admission.py +218 -0
  447. package/md_cg/test_readcache_prodpath.py +4 -0
  448. package/md_cg/test_recall_face_guards.py +854 -812
  449. package/md_cg/test_recall_metadata.py +422 -0
  450. package/md_cg/test_retr_s5.py +104 -5
  451. package/md_cg/test_retr_score_once.py +14 -2
  452. package/md_cg/test_review_cli_attribution.py +11 -7
  453. package/md_cg/test_review_queue_batch.py +123 -0
  454. package/md_cg/test_review_queue_index.py +130 -0
  455. package/md_cg/test_rrf_candidates.py +292 -0
  456. package/md_cg/test_rrf_qualified_topk.py +214 -0
  457. package/md_cg/test_rrf_unique_votes.py +218 -0
  458. package/md_cg/test_s1_ast_judge.py +110 -0
  459. package/md_cg/test_scrub_ledger.py +413 -0
  460. package/md_cg/test_security_audit_b26.py +34 -3
  461. package/md_cg/test_security_audit_v21.py +1 -2
  462. package/md_cg/test_sensitivity_predicate_single_point.py +184 -0
  463. package/md_cg/test_session5_identity.py +93 -0
  464. package/md_cg/test_session_attribution_consistency.py +813 -0
  465. package/md_cg/test_session_identity_tristate.py +486 -0
  466. package/md_cg/test_sleep_p1.py +35 -3
  467. package/md_cg/test_state_event_op.py +8 -1
  468. package/md_cg/test_state_projection_streaming.py +136 -0
  469. package/md_cg/test_stg_interval_candidates.py +280 -0
  470. package/md_cg/test_stg_time_blocks.py +239 -0
  471. package/md_cg/test_stgidx_index_parity.py +16 -12
  472. package/md_cg/test_subgraph_bounded_bfs.py +227 -0
  473. package/md_cg/test_sustain_bounded.py +30 -2
  474. package/md_cg/test_task_memory.py +394 -0
  475. package/md_cg/test_text_projection_mutations.py +101 -0
  476. package/md_cg/test_time_anchor.py +421 -0
  477. package/md_cg/test_time_core_lint.py +1024 -975
  478. package/md_cg/test_tool_face.py +200 -189
  479. package/md_cg/test_trust_bounded_frontier.py +248 -0
  480. package/md_cg/test_trust_edge_reachability.py +405 -0
  481. package/md_cg/test_w5_path_fingerprint.py +381 -0
  482. package/md_cg/test_w7_redlines.py +659 -0
  483. package/md_cg/test_write_return_shape.py +919 -0
  484. package/md_cg/test_writepipe.py +5 -3
  485. package/md_cg/timeanchor.py +534 -0
  486. package/md_cg/tokens.py +68 -5
  487. package/md_cg/tool_face.py +273 -260
  488. package/md_cg/trust.py +110 -26
  489. package/md_cg/twophase.py +74 -2
  490. package/md_cg/weights.py +642 -638
  491. package/md_cg/whitebox.py +9 -0
  492. package/md_cg/whitebox_kb/aeis_core/core.py +5026 -5026
  493. package/md_cg/whitebox_kb/wisdom/chat_engine.py +1504 -1506
  494. package/md_cg/whitebox_kb/wisdom/code_compose.py +876 -837
  495. package/md_cg/whitebox_kb/wisdom/compiler_code_units.py +3040 -3033
  496. package/md_cg/whitebox_kb/wisdom/compose_engine.py +3 -2
  497. package/md_cg/whitebox_kb/wisdom/core/condition_vm.py +22 -2
  498. package/md_cg/whitebox_kb/wisdom/core/parser.py +17 -0
  499. package/md_cg/whitebox_kb/wisdom/csre.py +412 -264
  500. package/md_cg/whitebox_kb/wisdom/csre_session.py +97 -0
  501. package/md_cg/whitebox_kb/wisdom/md_access.py +23 -3
  502. package/md_cg/whitebox_kb/wisdom/migrate_wisdom.py +21 -0
  503. package/md_cg/whitebox_kb/wisdom/mini_python.py +44 -7
  504. package/md_cg/whitebox_kb/wisdom/navigate.py +249 -248
  505. package/md_cg/whitebox_kb/wisdom/net_units.py +12 -7
  506. package/md_cg/whitebox_kb/wisdom/neural_retrieve.py +100 -10
  507. package/md_cg/whitebox_kb/wisdom/semantic_translate.py +13 -7
  508. package/md_cg/whitebox_kb/wisdom/test_ecosystem_assembly.py +158 -158
  509. package/md_cg/whitebox_kb/wisdom/test_safe_module_proxy.py +147 -0
  510. package/md_cg/whitebox_kb/wisdom/verifier.py +1918 -1297
  511. package/md_cg/writelimit.py +584 -390
  512. package/md_cg/writepipe.py +587 -76
  513. package/package.json +13 -9
  514. package/src/bridge.ts +34 -11
  515. package/src/cli.ts +75 -65
  516. package/src/hooks.ts +908 -776
  517. package/src/index.ts +19 -5
  518. package/src/init.ts +362 -361
  519. package/src/lib/datapath.ts +419 -326
  520. package/src/lib/hook_audit.ts +256 -243
  521. package/src/lib/mdcg_client.ts +10 -2
  522. package/src/lib/mutual.ts +481 -432
  523. package/src/lib/prompt_safety.ts +164 -149
  524. package/src/lib/roleplay_web.ts +30 -2
  525. package/src/lib/token_store.ts +274 -52
  526. package/src/tools.ts +249 -223
  527. package/zcode/README.md +1 -1
  528. 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 +0 -169
  529. 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 +0 -144
  530. 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 +0 -299
  531. 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 +0 -239
  532. package/docs/eval/AGI/344/270/203/347/273/264/350/257/204/345/210/206/346/212/245/345/221/212_/345/205/255/345/256/266/350/256/260/345/277/206/347/263/273/347/273/237/345/220/214/345/260/272/345/256/236/346/265/213_v3.1.md +0 -340
  533. package/docs/eval/DSH/346/227/245/345/277/227/347/264/242/345/274/225v2_/345/217/202/350/200/203dsh-TUI_v1.0.md +0 -248
  534. package/docs/eval/DSH/346/227/245/345/277/227/347/264/242/345/274/225/346/225/210/346/236/234/351/252/214/350/257/201_v1.0.md +0 -209
  535. package/docs/eval/DSH/347/253/257/347/274/272/351/231/267/344/270/223/351/241/271_v1.0.md +0 -254
  536. package/docs/eval/N225_/347/264/242/345/274/225/346/227/245/345/277/227/351/235/236/345/257/271/350/261/241/350/243/205/350/275/275/351/235/242/347/261/273/345/236/213/351/227/270_v1.0.md +0 -418
  537. package/docs/eval/P1b2_/350/257/273/351/235/242/344/273/243/351/231/205/344/277/256/345/244/215_v1.0.md +0 -124
  538. package/docs/eval/bench_lingshu_self/bench_self.py +0 -140
  539. package/docs/eval/bench_lingshu_self/self_bench_result.json +0 -404
  540. 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 +0 -38
  541. package/docs/eval/cons200_/345/206/262/347/252/201/346/243/200/346/265/213/351/200/211/351/235/242_/345/256/236/346/226/275/350/256/260/345/275/225_v1.0.md +0 -223
  542. package/docs/eval/cons200_/345/206/262/347/252/201/346/243/200/346/265/213/351/200/211/351/235/242_/345/256/236/346/226/275/350/256/260/345/275/225_v1.1.md +0 -340
  543. package/docs/eval/issue43_/351/273/230/350/256/244/347/255/226/347/225/245/344/270/216/351/224/256/347/261/273/345/236/213/351/227/270_v1.1.md +0 -441
  544. package/docs/eval/issue43_/351/273/230/350/256/244/347/255/226/347/225/245/345/212/240/350/275/275/344/270/216/345/207/255/346/215/256/346/230/216/346/226/207/351/230/237/345/210/227_v1.0.md +0 -398
  545. package/docs/eval/issue50_/345/205/203/346/225/260/346/215/256/351/200/217/344/274/240/344/270/216/345/205/234/345/272/225_/344/277/256/345/244/215/350/256/260/345/275/225_v1.0.md +0 -437
  546. package/docs/eval/issue50_/345/215/212/351/207/215/345/244/215/345/276/205/345/256/232/345/244/215/346/240/270_/344/277/256/345/244/215/350/256/260/345/275/225_v1.0.md +0 -488
  547. package/docs/eval/issue50_/345/276/205/345/256/232/345/244/215/346/240/270/345/205/245/351/230/237_/344/277/256/345/244/215/350/256/260/345/275/225_v1.0.md +0 -428
  548. package/docs/eval/issue50_/350/257/273/351/235/242/344/277/235/346/212/244/345/217/252/350/256/244/346/230/276/345/274/217/346/235/245/346/272/220_/344/277/256/345/244/215/350/256/260/345/275/225_v1.0.md +0 -271
  549. package/docs/eval/issue50_/351/207/215/350/246/201/345/272/246/345/220/214/346/272/220/344/270/216/344/277/235/346/212/244/350/257/255/344/271/211_/344/277/256/345/244/215/350/256/260/345/275/225_v1.0.md +0 -306
  550. package/docs/eval/issue51_/344/270/200/351/224/256/345/256/211/350/243/205/345/244/261/350/264/245_/345/275/222/345/261/236/345/210/244/345/256/232_v1.0.md +0 -58
  551. package/docs/eval/issue52_/346/235/241/344/273/266/345/205/210/350/241/214/344/270/216/346/210/252/346/226/255/345/217/257/350/247/202/346/265/213_/344/277/256/345/244/215/350/256/260/345/275/225_v1.0.md +0 -360
  552. package/docs/eval/issue64_replay_check_/346/240/270/351/252/214/344/270/216/344/277/256/345/244/215_v1.0.md +0 -284
  553. package/docs/eval/issue65_session_recall_/344/274/232/350/257/235/351/232/224/347/246/273_/346/240/270/351/252/214/344/270/216/344/277/256/345/244/215_v1.0.md +0 -329
  554. package/docs/eval//344/270/211/346/241/243/350/207/252/346/262/273_/346/255/245/351/252/244/342/221/241/346/241/243/344/275/215/345/215/225/344/270/200/345/205/245/345/217/243_/350/220/275/347/240/201/350/256/260/345/275/225_v1.0.md +0 -651
  555. package/docs/eval//344/270/211/346/241/243/350/207/252/346/262/273_/346/255/245/351/252/244/342/221/242/345/217/230/346/233/264/345/215/225/344/270/216/345/233/236/346/273/232/345/216/237/350/257/255_/350/220/275/347/240/201/350/256/260/345/275/225_v1.0.md +0 -439
  556. package/docs/eval//344/270/211/346/241/243/350/207/252/346/262/273_/346/255/245/351/252/244/342/221/243/345/207/206/345/205/245/350/257/273/346/225/260_/350/220/275/347/240/201/350/256/260/345/275/225_v1.0.md +0 -1337
  557. package/docs/eval//344/270/211/346/241/243/350/207/252/346/262/273_/346/255/245/351/252/244/342/221/244/346/224/266/345/256/230/344/270/216/345/205/250/351/223/276/351/252/214/346/224/266_/350/220/275/347/240/201/350/256/260/345/275/225_v1.0.md +0 -1817
  558. 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 +0 -185
  559. package/docs/eval//344/274/230/345/214/226/347/254/254/344/270/200/346/211/271_/346/216/245/347/272/277/344/270/216/347/255/211/344/273/267/345/217/230/346/215/242_v1.0.md +0 -498
  560. package/docs/eval//344/274/230/345/214/226/347/254/254/344/270/211/346/211/271_/351/227/250/347/246/201/350/275/254/346/255/243/344/270/216/350/260/203/345/272/246/346/255/242/350/241/200/344/270/216/351/227/250/346/216/247/346/224/266/345/217/243_v1.0.md +0 -669
  561. package/docs/eval//344/274/230/345/214/226/347/254/254/344/272/214/346/211/271_/344/276/235/350/265/226/351/200/217/344/274/240/344/270/216/345/257/271/346/213/215/345/217/243/345/276/204/344/270/216/350/264/237/347/274/223/345/255/230_v1.0.md +0 -683
  562. package/docs/eval//345/207/272/350/264/247/351/235/242/345/206/222/347/203/237_/350/277/233/350/264/247/351/227/250/347/246/201_v1.0.md +0 -460
  563. package/docs/eval//345/217/221/345/270/20306_/344/270/200/351/224/256/351/205/215/347/275/256/344/270/216DSH0172_v1.0.md +0 -171
  564. package/docs/eval//345/217/221/345/270/20307_/345/244/226/351/203/250/346/212/245/345/221/212/345/233/233/346/211/271/344/277/256/345/244/215/344/270/216/346/217/222/344/273/266/351/235/242/345/212/240/345/233/272_v1.0.md +0 -209
  565. package/docs/eval//345/217/221/345/270/20308_/350/207/252/350/277/255/344/273/243/344/270/216/347/235/241/347/234/240_/345/233/276/346/243/200/347/264/242/350/267/257/344/270/216/346/235/203/351/207/215_v1.0.md +0 -97
  566. package/docs/eval//345/217/221/345/270/20309_/346/243/200/347/264/242/351/235/242/344/270/211/346/211/271/346/224/266/345/217/243_v1.0.md +0 -66
  567. package/docs/eval//345/217/221/345/270/20310_/350/257/273/351/235/242/346/215/237/345/235/217UTF8/345/256/266/346/227/217/346/224/266/345/217/243_v1.0.md +0 -71
  568. package/docs/eval//345/217/221/345/270/20311_/345/244/226/346/212/245/345/205/255/350/277/236/344/277/256/344/270/216/344/270/226/347/225/214/346/250/241/345/236/213/345/272/225/345/272/247_v1.0.md +0 -91
  569. package/docs/eval//345/217/221/345/270/20312_/344/270/226/347/225/214/346/250/241/345/236/213/345/212/237/350/203/275/347/253/257_v1.0.md +0 -108
  570. package/docs/eval//345/217/221/345/270/20313_/344/270/212/344/270/213/346/226/207/350/207/252/347/256/241/347/220/206/346/234/272/345/210/266_v1.0.md +0 -114
  571. package/docs/eval//345/217/221/345/270/20314_/350/272/253/344/275/223/303/227/350/204/221/347/273/204/345/220/210_v1.0.md +0 -160
  572. package/docs/eval//345/217/221/345/270/20315_/344/274/232/350/257/235/351/232/224/347/246/273/344/277/256/345/244/215/344/270/216/350/272/253/344/275/223/345/272/223/345/261/225/347/244/272_v1.0.md +0 -88
  573. 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 +0 -174
  574. package/docs/eval//345/275/222/344/270/200/345/261/202/347/274/272/347/234/201/347/277/273/345/205/263_/344/277/256/345/244/215/350/256/260/345/275/225_v1.0.md +0 -458
  575. package/docs/eval//346/200/247/350/203/275/344/270/223/351/241/271_/345/206/267/346/237/245/350/257/242/344/270/216/345/206/205/345/255/230_v1.0.md +0 -218
  576. package/docs/eval//346/225/205/351/232/234/346/263/250/345/205/245/345/256/236/346/265/213_v1.0.md +0 -432
  577. 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 +0 -223
  578. package/docs/eval//347/224/250/346/210/267/350/243/201/345/206/263/345/244/204/347/275/256_/347/254/254/344/270/200/346/211/271_v1.0.md +0 -195
  579. package/docs/eval//347/224/250/346/210/267/350/243/201/345/206/263/345/244/204/347/275/256_/347/254/254/344/272/214/346/211/271B6_/346/234/254/346/234/272/350/267/257/345/276/204/347/233/270/345/257/271/345/214/226_v1.0.md +0 -78
  580. 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 +0 -100
  581. 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 +0 -197
  582. package/docs/eval//347/254/2543/345/261/202stg/347/273/223/346/236/204/347/264/242/345/274/225_/345/256/236/346/226/275/350/256/260/345/275/225_v1.0.md +0 -250
  583. package/docs/eval//347/254/254/344/270/211/346/226/271/351/252/214/350/257/201/346/212/245/345/221/212_LoCoMo_/347/201/265/346/236/242_.png +0 -0
  584. package/docs/eval//347/254/254/344/270/211/346/226/271/351/252/214/350/257/201/346/212/245/345/221/212_/347/201/265/346/236/242_vs_dejavu_/347/273/237/344/270/200/350/257/204/345/210/206_v7.md +0 -202
  585. package/docs/eval//347/274/226/347/240/201/351/235/242/345/211/215/347/275/256_/345/205/245/345/217/243/350/207/252/344/277/235/350/257/201UTF8/344/270/216/346/226/207/346/234/254open/345/256/210/345/215/253_v1.0.md +0 -646
  586. 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 +0 -156
  587. 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 +0 -210
  588. 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 +0 -227
  589. 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 +0 -203
  590. 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 +0 -233
  591. 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 +0 -191
  592. 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 +0 -213
  593. 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 +0 -214
  594. 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 +0 -199
  595. 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_v18.md +0 -183
  596. 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_v19.md +0 -207
  597. 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 +0 -156
  598. 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_v20.md +0 -283
  599. 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_v21.md +0 -224
  600. 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_v22.md +0 -223
  601. 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_v23.md +0 -293
  602. 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_v24.md +0 -303
  603. 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_v25.md +0 -230
  604. 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_v26.md +0 -196
  605. 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_v27.md +0 -259
  606. 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_v28.md +0 -217
  607. 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_v29.md +0 -270
  608. 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 +0 -152
  609. 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_v30.md +0 -231
  610. 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_v31.md +0 -281
  611. 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 +0 -128
  612. 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 +0 -114
  613. 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 +0 -192
  614. 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 +0 -187
  615. 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 +0 -207
  616. 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 +0 -203
  617. package/docs/eval//350/257/255/344/271/211/346/227/266/347/251/272/345/233/276/350/241/245/345/205/250_P1_/346/212/275/345/217/226/345/231/250v4_/344/270/211/347/216/207/350/257/273/346/225/260_v1.0.md +0 -231
  618. package/docs/eval//351/262/270/345/250/230/344/270/203/347/261/273/347/212/266/346/200/201/350/277/275/350/270/252_v1.0.md +0 -176
  619. package/docs/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 -385
  620. package/docs/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 +0 -165
  621. 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 +0 -76
  622. package/docs/hive//345/244/232/347/253/257harness/351/200/232/344/277/241/345/245/221/347/272/246_v0.1.md +0 -42
  623. package/docs/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 +0 -236
  624. package/docs/hive//345/256/211/345/205/250/345/256/241/350/256/241/345/256/236/351/224/232_v0.1.md +0 -202
  625. 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 +0 -43
  626. 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 +0 -210
  627. 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 +0 -152
  628. 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 +0 -47
  629. 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 +0 -519
  630. 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 +0 -85
  631. 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 +0 -191
  632. 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 +0 -90
  633. package/docs/images/lingshu-moonlight-covenant-poster-preview.jpg +0 -0
  634. package/docs/images/lingshu-moonlight-covenant-poster.png +0 -0
  635. package/docs/mdcg/D_meta_/345/267/245/347/250/213/345/214/226/346/226/271/346/241/210_v0.2.md +0 -216
  636. package/docs/mdcg/README/350/257/246/347/273/206/347/211/210_v0.4.10.md +0 -746
  637. package/docs/mdcg/guardrail-charter.md +0 -164
  638. package/docs/mdcg/lingshu_tutorial.html +0 -14956
  639. package/docs/mdcg/memory-assessment.html +0 -923
  640. package/docs/mdcg/memory_score.html +0 -213
  641. package/docs/mdcg/memory_score.png +0 -0
  642. package/docs/mdcg/release_v0.3.0.md +0 -36
  643. package/docs/mdcg/release_v0.4.11.md +0 -49
  644. package/docs/mdcg/release_v0.4.5.md +0 -55
  645. package/docs/mdcg/tool_table_v0.3.0.md +0 -117
  646. package/docs/mdcg//344/270/226/347/225/214/346/250/241/345/236/213/345/212/237/350/203/275/347/253/257_/344/275/277/347/224/250/344/270/216/350/277/220/347/273/264_v1.0.md +0 -384
  647. package/docs/mdcg//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_/345/245/221/347/272/246_v0.1.md +0 -82
  648. 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 +0 -600
  649. 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 +0 -266
  650. 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 +0 -275
  651. package/docs/mdcg//345/217/221/345/270/203/351/227/250/347/246/201/351/223/276_v0.1.md +0 -134
  652. 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 +0 -130
  653. 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 +0 -278
  654. package/docs/mdcg//347/201/265/346/236/242MCP/345/267/245/345/205/267/346/200/273/350/241/250_v3.4.md +0 -164
  655. package/docs/mdcg//347/201/265/346/236/242_/350/207/252/346/210/221/345/261/202/345/256/232/344/271/211.md +0 -43
  656. package/docs/mdcg//347/201/265/346/236/242/344/270/211/345/261/202/346/213/206/345/210/206/350/247/204/345/210/222_v0.1.md +0 -181
  657. 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 +0 -177
  658. package/docs/mdcg//347/216/257/344/272/214_/347/231/275/347/256/261/345/241/253/345/205/205/346/265/201/346/260/264/347/272/277_v0.1.md +0 -31
  659. package/docs/mdcg//347/235/241/347/234/240/345/221/250/346/234/237_/350/277/220/347/273/264/345/211/215/346/217/220/344/270/216/347/273/264/346/212/244/346/214/207/345/215/227_v1.0.md +0 -337
  660. package/docs/mdcg//347/274/272/345/217/243/345/215/225_P0/346/224/266/345/217/243_v0.1.md +0 -270
  661. 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 +0 -491
  662. package/docs/mdcg//350/256/244/347/237/245/345/233/276_MD/347/233/256/345/275/225/346/226/271/346/241/210_v0.1.md +0 -552
  663. 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 +0 -340
  664. 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 +0 -624
  665. package/docs/mdcg//350/256/244/347/237/245/345/233/276/344/275/234/344/270/272/350/256/260/345/277/206/346/223/215/344/275/234/347/263/273/347/273/237_/350/257/204/344/274/260/344/270/216/350/267/257/347/272/277_v0.1.md +0 -94
  666. package/docs/mdcg//350/256/260/345/277/206/346/223/215/344/275/234/347/263/273/347/273/237_MdCGOS/344/270/216MCP/346/216/245/345/205/245_v0.1.md +0 -281
  667. package/docs/mdcg//350/267/250/347/253/257/351/252/214/350/257/201/344/270/216/345/220/214/346/255/245/345/215/217/350/256/256_v0.1.md +0 -93
  668. 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 +0 -176
  669. package/docs/plans/stg/346/235/241/344/273/266/345/214/226/344/270/216/347/273/223/346/236/204/347/264/242/345/274/225_/350/256/276/350/256/241_v0.1.md +0 -124
  670. package/docs/plans//345/205/250/344/270/255/346/226/207/347/274/226/347/240/201/344/270/216/350/234/202/345/267/242/344/273/273/345/212/241/346/240/207/350/257/206/345/245/221/347/272/246_v2.0.md +0 -220
  671. 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 +0 -81
  672. package/docs/plans//345/244/232/344/270/273/344/275/223/344/270/226/347/225/214/346/250/241/345/236/213_/345/257/271/351/275/220/350/257/204/344/274/260/344/270/216/350/220/275/345/234/260/350/256/276/350/256/241_v0.1.md +0 -141
  673. package/docs/plans//347/201/265/346/236/242/350/272/253/344/275/223/303/227/350/204/221_/344/270/226/347/225/214/346/250/241/345/236/213/345/257/271/346/216/245/350/256/276/350/256/241_v0.1.md +0 -119
  674. package/docs/plans//347/234/237/346/272/220/347/264/242/345/274/225_/351/200/232/347/224/250/346/234/272/345/210/266_v0.3.md +0 -227
  675. package/docs/plans//347/235/241/347/234/240/344/270/216/350/207/252/350/277/255/344/273/243_/345/212/237/350/203/275/344/274/230/345/214/226/350/256/276/350/256/241_v0.4.md +0 -547
  676. package/docs/plans//347/237/245/350/257/206/345/233/276/350/260/261/351/241/271/347/233/256_/344/272/244/346/216/245/346/226/207/346/241/243_20260914.md +0 -308
  677. package/docs/plans//350/234/202/345/267/242/346/250/241/345/236/213/345/257/206/351/222/245/351/205/215/347/275/256/351/235/242_v1.0.md +0 -385
  678. package/docs/plans//350/256/260/345/277/206/350/207/252/345/244/204/347/220/206/344/270/211/346/241/243/350/207/252/346/262/273_/350/256/276/350/256/241_v0.2.md +0 -207
  679. package/docs/plans//350/257/255/344/271/211/346/227/266/347/251/272/345/233/276/350/241/245/345/205/250_/344/270/226/347/225/214/346/250/241/345/236/213/345/212/237/350/203/275/347/253/257_/350/256/276/350/256/241_v0.1.md +0 -107
  680. package/docs/promo/bilibili-cover.jpg +0 -0
  681. package/docs/promo/promo-visual-1.png +0 -0
  682. package/docs/promo/promo-visual-2.png +0 -0
  683. package/docs/promo/promo-visual-3.png +0 -0
  684. package/docs/promo/promo-visual-4.png +0 -0
  685. package/docs/promo/promo-visual-5.png +0 -0
  686. package/docs/swarm//350/234/202/347/276/244/344/272/222/350/201/224_v0.1.md +0 -704
  687. 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 +0 -242
  688. package/docs/swarm//350/234/202/347/276/244/345/244/232/346/231/272/350/203/275/344/275/223_/344/275/277/347/224/250/346/212/245/345/221/212_20260913.md +0 -68
  689. package/docs/swarm//350/234/202/347/276/244/345/244/232/346/231/272/350/203/275/344/275/223_/345/212/237/350/203/275/350/257/264/346/230/216_v0.6.md +0 -130
  690. package/docs/swarm//350/234/202/347/276/244/347/233/262/345/214/272/346/240/207/350/256/260_v1.0.md +0 -89
  691. 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 +0 -365
  692. package/docs/theory//344/277/241/346/201/257/345/267/256/344/270/272/344/273/200/344/271/210/345/277/205/347/204/266/345/255/230/345/234/250/344/270/224/350/207/252/347/204/266/346/211/251/345/244/247.md +0 -239
  693. 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 +0 -129
  694. 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 +0 -134
  695. package/docs/theory//346/231/272/350/203/275/347/232/204/345/205/254/347/220/206/345/214/226/345/237/272/347/237/263.md +0 -789
  696. package/docs/theory//346/231/272/350/203/275/347/232/204/350/256/244/347/237/245/350/277/207/347/250/213.md +0 -379
  697. package/docs/theory//346/231/272/350/203/275/350/256/2723.4.md +0 -5260
  698. 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 +0 -145
  699. 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 +0 -144
  700. 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 +0 -198
  701. package/docs/theory//347/231/275/347/256/261/346/231/272/350/203/275/346/230/257/344/273/200/344/271/210/357/274/237.md +0 -812
  702. package/docs/theory//347/231/275/347/256/261/346/231/272/350/203/275/347/263/273/345/210/227/302/267/347/254/254/344/272/224/347/257/207/357/274/232/350/256/251AI/347/234/237/346/255/243/350/243/205/344/270/212/350/256/260/345/277/206.md +0 -421
  703. 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 +0 -221
  704. 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 +0 -237
  705. package/docs//344/270/215/345/217/257/351/235/240/346/200/247/347/220/206/350/256/272_v0.1.md +0 -343
  706. 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 +0 -82
  707. package/docs//345/255/230/347/256/227/344/270/200/344/275/223/346/236/266/346/236/204/345/255/246/344/271/240_/345/244/226/351/203/250/347/220/206/350/256/272/345/257/271/347/205/247/344/270/216/350/267/257/347/272/277/344/272/244/346/216/245_20260916.md +0 -98
  708. package/docs//345/255/230/347/256/227/344/270/200/344/275/223/347/245/236/347/273/217/347/275/221/347/273/234/346/236/266/346/236/204_/346/200/273/347/272/262/344/270/216/344/272/244/346/216/245_20260916.md +0 -109
  709. package/docs//347/201/265/346/236/242/350/207/252/346/210/221/346/224/271/350/277/233/345/267/245/344/275/234/350/256/241/345/210/222_/345/244/226/351/203/250/347/240/224/347/251/266/347/263/273/345/210/227/345/220/270/346/224/266_v1_20260919.md +0 -286
  710. package/docs//347/231/275/347/256/261/345/214/226/347/272/262/351/242/206_/344/270/211/351/241/271/347/233/256/345/255/246/344/271/240/346/200/273/346/236/266/346/236/204/344/270/216Ghidra/344/272/244/346/216/245_20260916.md +0 -104
  711. package/docs//347/247/244_/350/256/260/345/277/206/347/263/273/347/273/237/350/257/204/346/265/213/350/247/204/350/214/203_v2.1.md +0 -225
  712. package/docs//350/256/260/345/277/206/347/263/273/347/273/237/345/215/263/347/231/275/347/256/261/345/214/226/346/231/272/350/203/275_/346/236/266/346/236/204/345/257/271/347/205/247_v0.2.md +0 -121
  713. package/docs//350/256/260/345/277/206/347/263/273/347/273/237/345/215/263/347/231/275/347/256/261/345/214/226/346/231/272/350/203/275_/346/236/266/346/236/204/345/257/271/347/205/247_v0.3.md +0 -117
  714. package/docs//350/256/260/345/277/206/350/257/204/345/256/241/347/263/273/347/273/237_/347/253/213/351/241/271/350/256/276/350/256/241/344/270/216/346/226/275/345/267/245/344/272/244/346/216/245_20260915.md +0 -183
@@ -1,299 +0,0 @@
1
- # AGI 七维评分报告 · md_cg(灵枢认知图记忆系统)v1.0
2
-
3
- > **评估对象**:`dsh-memory / md_cg`(条件空间驱动的认知图记忆系统)
4
- > **标尺**:`AEIS/docs/AGI记忆系统评分标尺_v1.0.md`
5
- > **前次基线**:标尺 §6.1 初评(2026-09-09)→ S7.0 / R7.0 / J7.5 / C7.5 / U6.0 / I4.5 / T7.5,综合 **5.8**,层次 **L4**
6
- > **本次性质**:**标尺 §2.3 四个门槛判别测试的实跑复核**(非印象打分)
7
- > **证据等级**:全部结论均有**落盘 JSON 实测数据**,逐项标注数据文件
8
- > **对照物**:同期完成的另一套记忆系统复核评价(**内部文档,不随仓库公开**,故不给路径)
9
-
10
- ---
11
-
12
- ## 一、评估条件空间
13
-
14
- ```
15
- [判定单]
16
- 条件空间:观测位置=灵枢侧自我实测(自评风险见 §二)
17
- 观测工具=标尺 §2.3 四门槛判别测试 + md_cg 公开 API(cg/recall/consistency/evolution/identity/self_state)
18
- 时间窗口=2026-09-12 实测快照
19
- 存在约束=自评系统;L4「新增条件分支」判据需区分「库结构行为」与「开发者行为」
20
- 白箱三问:依据=标尺 §2.3 判别测试表 + 本仓库实验目录全部 result_*.json
21
- 条件=分数只在同层内可比(§4.2);本次重点在**门槛是否通过**,不在分数微调
22
- 错后=发现反例即下调/标 DEFER 并归档(§七)
23
- 裁决:ACCEPT
24
- 依据:四门槛测试均已按 §2.3 口径实跑,数据落盘,可复现
25
- 缺失维度:U 维「同条件不同值」处理机制(S1 暴露);I 维「元认知工具」未实测
26
- (世界模型属身体层 AEIS,不计入大脑 I 维 —— 见 §十一.8 层错位声明)
27
- 验证路径:§六 S1 复现 + §十一 DEFER 项
28
- ```
29
-
30
- ---
31
-
32
- ## 二、自评风险声明与缓解
33
-
34
- 本报告是**系统评自己**,存在结构性利益冲突。三条缓解措施:
35
-
36
- 1. **判据外置**:全部通过标准取自标尺 §2.3 原文(hit@1 ≥ 80% / 负条件拒绝率 ≥ 70%),不由被测系统定义。
37
- 2. **语料外置**:对抗语料**直接复用**本地实验工作区(`docs/experiments/`,**不入库**)`deja-vu-negcond/corpus.py` 的同一批 20 组 `GROUPS`(不复制文本、不挑选利于己的样本)。
38
- 3. **反例优先**:凡实测未达判据处**照实记录并下调**(§六 S1、§十 U 维),不因"是自己"而豁免;发现的假阳性一并纠正(§七)。
39
-
40
- > **仍未消除的风险**:判据的解释权、观测位置的独立性。故 §十 对 L5 临界结论给 **DEFER** 而非 ACCEPT。
41
-
42
- ---
43
-
44
- ## 三、实验装置
45
-
46
- | 项 | 设置 |
47
- |---|---|
48
- | 被测件 | `md_cg`(`from md_cg.mdcos import MdCGOS, MdCGSecure`) |
49
- | 隔离 | 每组实验**独立库目录**(`_store/<name>`,跑前 `rmtree`),互不污染 |
50
- | 语料 | 复用本地实验工作区 `docs/experiments/deja-vu-negcond/corpus.py`(不入库)的 20 组 `GROUPS`(L3);L2 用 10 唯一目标 + 990 干扰 |
51
- | 流程纪律 | 先 smoke 验证 API 返回结构,再全量跑;结果落盘 `result_*.json` |
52
- | 命令 | 统一 python(`PYTHONUTF8=1`,argv 列表,不经 shell) |
53
-
54
- **脚本与产物清单**(`docs/experiments/md-cg-gate-probe/`):
55
-
56
- | 脚本 | 测试 | 产物 |
57
- |---|---|---|
58
- | `run_l2_retrieval.py` | L2 检索判别 | `result_l2.json`、`result_l2_strict.json` |
59
- | `run_l3_negcond2.py` | L3 负条件拒绝(四口径) | `result_l3v2_{A,B,C,D}.json` |
60
- | `run_l4_evolution.py` | L4 矛盾反馈 | `result_l4.json` |
61
- | `run_l4b_rollback.py` | L4 回滚链路 | `result_l4b.json` |
62
- | `run_l5_continuity.py` / `run_l5b_selfstate.py` | L5 身份/变更史 | `result_l5.json`、`result_l5b.json` |
63
- | `run_ct_probe.py` | C 预算 / T 墓碑 | `result_ct.json` |
64
- | `run_t2_probe.py` | T 写保护·密级 / I 版本链(补测) | `result_t2.json` |
65
-
66
- ---
67
-
68
- ## 四、L2 门槛 · 检索判别 —— **PASS**
69
-
70
- 标尺 §2.3 判据:`1000 条历史会话中检索「三月修过的 jwt 问题」→ hit@1 ≥ 80%`。
71
-
72
- | 口径 | 节点量 | 查询 | hit@1 | hit@5 | 数据 |
73
- |---|---|---|---|---|---|
74
- | 含歧义 | 1000(10 目标 + 990 干扰) | 10 | **0.90** | 1.00 | `result_l2.json` |
75
- | 去歧义(strict) | 同上 | 10 | **1.00** | 1.00 | `result_l2_strict.json` |
76
-
77
- **结论:两种口径均 ≥ 0.80,通过。**
78
-
79
- **唯一失败案例归因**(含歧义口径,`target_07`):
80
-
81
- - 查询「九月修过的 **限流** 问题」,目标被 3 条 filler 压到 rank 4。
82
- - 归因:目标中「限流」处于**组件位**(登录接口的限流阈值),而 filler 中「限流」处于**问题位**("如何解决限流")。md_cg 的词面召回**不消解语义槽位**,同词高权重者胜出。
83
- - 性质:属**评测集歧义**(含歧义口径);strict 口径移除该歧义后 hit@1 = 1.00。**这是真实边界,不是门檻未过。**
84
-
85
- ---
86
-
87
- ## 五、L3 门槛 · 负条件拒绝 —— **PASS**
88
-
89
- 标尺 §2.3 判据:`20 组「关键词相似但能力不同」对抗对,负条件拒绝率 ≥ 70%`。
90
-
91
- **四口径设计**(关键:未做 D 镜像就不能排除"无脑拒一切"):
92
-
93
- | 口径 | 写入 | 查询上下文 | reject_loose | reject_strict | accept_rate | hit@1 |
94
- |---|---|---|---|---|---|---|
95
- | A | 裸文本,无 context | 无 | 0.20 | 0.00 | 0.00 | 1.00 |
96
- | B | CCG,无 context | 无 | 0.20 | 0.00 | 1.00 | 0.90 |
97
- | **C** | CCG | +correct 条件 | **1.00** | 0.00 | 1.00 | 1.00 |
98
- | **D** | CCG | +bait 条件(镜像) | **1.00** | 0.00 | 1.00 | 1.00 |
99
-
100
- - **A/B ≈ 0.20**:无条件路由时,bait 与 correct 无法区分(基线)。
101
- - **C/D = 1.00 且完全对称**:给 correct 条件时拒 bait、给 bait 条件时**同样拒 correct**——**双向对称**排除了"凡有冲突一律拒"的作弊路径,证明是**真条件路由**。
102
- - 数据:`result_l3v2_A/B/C/D.json`。
103
-
104
- **两条诚实标注**:
105
-
106
- 1. `reject_strict = 0`:拒绝靠 **S 维条件空间物理分桶**(不同 `condition_space` 互不召回),而非 J 维语义层判定 REJECT。功能上等价,机制上需注明——它拒绝的是"条件不同",不是"条件相反"。
107
- 2. 该口径的拒绝率测量的是**结构化写入 + 条件查询**下的召回隔离,**不等于语义层的拒答率**;同批 `GROUPS` 下的横向对照评价属**内部文档,不随仓库公开**,故此处不引用其数据。
108
-
109
- **结论:1.00 ≥ 0.70,通过。**
110
-
111
- ---
112
-
113
- ## 六、L4 门槛 · 矛盾反馈 —— **部分通过(含 FAIL 项)**
114
-
115
- 标尺 §2.3 判据:`给一个与已有记忆矛盾的反馈 → 新增条件分支,而非追加矛盾条目`。
116
-
117
- | 场景 | 条件设计 | verdict | 是否满足判据 | 数据 |
118
- |---|---|---|---|---|
119
- | **S1** 同条件、不同值 | 同一 `condition_space` 下改值 | **ACCEPT + 静默写入** | **FAIL** | `result_l4.json` |
120
- | **S2** 不同条件 | 冲突但条件可区分 | DEFER(strength 1.0) | 部分 | 同上 |
121
- | **S3** 负条件互斥 | 条件互斥 | DEFER + 落 `unresolved` 层 | PASS | 同上 |
122
- | **S4** 自否定 | 不适用条件命中自身正文 | **REJECT**(条件自相矛盾) | PASS | 同上 |
123
-
124
- - 飞轮结构落盘:`unresolved` 3 条、`rejected` 0 条(`result_l4.json.structure`)。
125
- - **S1 是真实反例(已定位到具体缺口)**:调用**一致性引擎本体** `CONS.check(cg, B, condition_space=CS_PROD, auto_flywheel=True)`,对「生产环境 MySQL 下连接池 32 → 64」这一**同条件空间取值矛盾**返回 `verdict=ACCEPT / reason=无冲突 / strength=0.0`,随后按同一 id 覆盖写入。缺口在**引擎判定逻辑不覆盖"同条件不同取值"比对**(非"未调用检测"、非"飞轮未开"),这正是 §2.3 要防的"覆盖矛盾条目"。属 U 维「误差驱动结构精化」的边界暴露(§十)。
126
- - **S2 是假阳性**:条件可区分本应触发"补区分条件后分流",实测只给 DEFER。功能上安全(不误写),但未达成"新增分支"。
127
-
128
- **回滚链路(`result_l4b.json`)**:
129
-
130
- | 步骤 | 实测 |
131
- |---|---|
132
- | `evolution_record(before, after)` | 生成 `kind=condition_gap` 条目,含 `state.before / state.after` |
133
- | `evolution_rollback(entry_id, dry_run=True)` | ok,`would_change` 列出 7 项将回退字段 |
134
- | `evolution_rollback(entry_id)` 实执行 | ok,生成溯源条目 `rollback_of = <原 entry_id>` |
135
-
136
- **结论:S3/S4 通过、回滚链路成立;S1 未通过 → 门槛判为「部分通过」,并把 U 维置入 DEFER(§十一)。**
137
-
138
- ---
139
-
140
- ## 七、L5 门槛 · 跨会话身份 / 变更史 —— **PASS**
141
-
142
- 标尺 §2.3 判据:`跨时间/跨会话/跨副本问「你是谁、上个月相信什么、现在改了吗」→ 身份一致 + 能陈述变更与理由`。
143
-
144
- **阶段 A(建)→ 阶段 B(新实例问)→ 阶段 C(库副本问)**,`result_l5.json` / `result_l5b.json`:
145
-
146
- | 检查 | 结果 | 证据 |
147
- |---|---|---|
148
- | Q1 身份锚点跨会话持久 | ✓ | `identity_self_lingshu` 在 `self` 层,新实例可读回 |
149
- | Q3 跨会话召回 | ✓ | `session_recall` 取回会话 A 的要旨 |
150
- | Q4 变更史含**理由** | ✓ | `belief_pool`:pool 32→64,evidence「九月压测显示 QPS 翻倍需 64」 |
151
- | Q5 状态审计 | ✓(补测后) | `refresh` 后 `verdict=consistent`,`n_issues=0` |
152
- | Q7 跨副本身份一致 | ✓ | `l5_replica` 独立库问「你是谁」,答案与主库逐字一致 |
153
-
154
- **纠正一处既有假阳性**:上轮 `result_ct.json` 的 `L5_version_chain.chain_linked = true`,实为 `null == null`——第二次 `self_state_refresh` 因**幂等跳过**(状态未变),返回中**根本没有** `state_hash`。该 `true` 不可采信。
155
-
156
- **本轮 `run_t2_probe.py` 用 `force=True` 产生非空真链**(`result_t2.json`):
157
-
158
- | 步 | 动作 | version | hash | prev |
159
- |---|---|---|---|---|
160
- | 1 | `refresh` | 1 | `8881e9d6…` | null |
161
- | 2 | `refresh`(幂等) | 1 | `8881e9d6…` | null(`changed=false`) |
162
- | 3 | `refresh(force)` | 2 | `8881e9d6…` | `8881e9d6…` |
163
- | 4 | `refresh(force)` | 3 | `8881e9d6…` | `8881e9d6…` |
164
- | 5 | `refresh(dimensions=…)` | 4 | **`be8217b2…`** | `8881e9d6…` |
165
-
166
- - `link_3_to_1 / link_4_to_3 / link_5_to_4` 全为 `true`(非空链);`audit_verdict=consistent`。
167
- - 步 5 证明**内容变化 → 指纹变化**(hash 由 `8881e9d6…` 变为 `be8217b2…`),即版本链不是纯计数。
168
- - 步 2 证明**幂等**:状态未变则不产新版本(这也解释了上轮假阳性的来源)。
169
-
170
- **结论:身份一致 + 能陈述变更理由,通过。**
171
-
172
- ---
173
-
174
- ## 八、维度补测 · C 调用 / T 可信
175
-
176
- ### 8.1 C 维 · 预算装填(`result_ct.json`)
177
-
178
- | budget_tokens | tokens_used | 是否超预算 |
179
- |---|---|---|
180
- | 60 | 46 | 否 |
181
- | 120 | 118 | 否 |
182
- | 400 | 398 | 否 |
183
-
184
- `recall(budget_tokens=)` 严格不超预算,返回 `pack / skipped / tokens_used / provenance / meta`;条目带 `state`(四态)与 `provenance`(词法路径 + rank)。**7–8 档要件(预算装填 + 角色分层)成立。**
185
-
186
- ### 8.2 T 维 · 写保护与遗忘闭环(`result_t2.json`)
187
-
188
- | 链路 | 实测 |
189
- |---|---|
190
- | `forget` → 墓碑 | ok,tombstone `e47da7f6…`,检索排除 |
191
- | `restore`(无 force) | **拒绝**(`error=tombstoned`),墓碑保留 → 防静默复活 |
192
- | `restore(force=True)` | ok,节点回到检索 |
193
- | 恢复后墓碑 | **仍存在**(删除史不可抹除,符合设计) |
194
- | self 层锚点 `forget` | **抛 `ProtectionError`**:「不可遗忘(层保护:self)」 |
195
- | `forget(..., override=True)` | 放行,且落快照 `_protected_history/<id>/20260912-082301.md` + `_protected_audit.jsonl` 1 行 |
196
- | `protect_stats()` | `protected_count=1, immutable_count=1, by_layer={self:1}` |
197
-
198
- ### 8.3 T 维 · 密级阶梯读隔离(`MdCGSecure`,`result_t2.json`)
199
-
200
- | principal clearance | readable | 可见节点 | 检索结果 |
201
- |---|---|---|---|
202
- | private | public/internal/private | **3/3** | `sec_pub, sec_int, sec_priv` |
203
- | internal | public/internal | **2/3** | `sec_int, sec_pub`(**priv 不可见**) |
204
- | public | public | **1/3** | `sec_pub` |
205
-
206
- `priv_visible_to_internal=false`、`internal_visible_to_public=false` → **阶梯隔离成立**。
207
-
208
- > ⚠️ **重要边界**:读隔离与 `require_admin` 闸门**只存在于 `MdCGSecure` 子类**;`MdCGOS` 无读隔离。故 T 维的密级证据**限定于启用 Secure 层的部署**,不适用于裸 `MdCGOS`。
209
-
210
- ---
211
-
212
- ## 九、四个门槛判定汇总
213
-
214
- | 门槛 | §2.3 测试 | 通过标准 | 实测 | 判定 |
215
- |---|---|---|---|---|
216
- | **→ L2** | 1000 会话 hit@1 | ≥ 0.80 | **0.90**(含歧义)/ **1.00**(strict) | ✅ PASS |
217
- | **→ L3** | 20 组负条件拒绝率 | ≥ 0.70 | **1.00**(C/D 对称) | ✅ PASS |
218
- | **→ L4** | 矛盾 → 新增条件分支 | 不追加矛盾条目 | S3/S4 PASS;**S1 FAIL(静默写入)**;S2 DEFER | ⚠️ 部分 PASS |
219
- | **→ L5** | 跨会话/副本身份 + 变更史 | 身份一致 + 能陈述理由 | 全项一致;变更史含理由;版本链非空 | ✅ PASS |
220
-
221
- ---
222
-
223
- ## 十、七维评分与层次判定
224
-
225
- ### 10.1 逐维(实测支持度)
226
-
227
- | 维度 | 标尺初评 | 实测证据 | 本次判定 |
228
- |---|---|---|---|
229
- | **S 结构** | 7.0 | `condition_space` 物理分桶(C/D 对称 1.00 证明条件路由真成立) | **7.0 维持**(无跨档新证据) |
230
- | **R 检索** | 7.0 | L2 hit@1 0.90 / 1.00 | **7.0 维持** |
231
- | **J 判断** | 7.5 | L3 拒绝率 1.00;四态(`state`)随条目返回 | **7.5 维持**(注:拒绝由 S 维分桶承担,strict=0) |
232
- | **C 调用** | 7.5 | 预算严格不超(46/118/398) | **7.5 维持** |
233
- | **U 演化** | 6.0 | 回滚链路成立;**S1 静默覆盖矛盾值** | **6.0(带反例,见 10.3)** |
234
- | **I 连续** | 4.5 | self 锚点 + 变更史含理由 + 跨副本身份 + 状态版本链(4 版,链完整,指纹随内容变);**世界模型属身体层,不计入本维** | **4.5 → 6.0 上调** |
235
- | **T 可信** | 7.5 | 写保护 + 墓碑 + override 快照审计 + 密级阶梯(Secure 层) | **7.5 维持** |
236
-
237
- ### 10.2 门槛验算
238
-
239
- | 晋级 | 门槛 | 验算 | 结果 |
240
- |---|---|---|---|
241
- | → L1 | S ≥ 3 | S 7.0 | ✓ |
242
- | → L2 | R ≥ 5 | R 7.0 | ✓ |
243
- | → L3 | J ≥ 5 且 S ≥ 5 | J 7.5 / S 7.0 | ✓ |
244
- | → L4 | U ≥ 5 且 J ≥ 6 | U 6.0 / J 7.5 | ✓ |
245
- | → **L5** | **I ≥ 6 且 U ≥ 6 且 T ≥ 6** | I **6.0** ✓ / U **6.0** ⚠️ / T 7.5 ✓ | **临界** |
246
-
247
- ### 10.3 L5 判为 **DEFER**(不 ACCEPT、不 REJECT)
248
-
249
- **表面已达门槛**:I 由 4.5 上调至 6.0(本轮实测:跨副本身份成立、状态版本链成立、变更史含理由——恰是标尺 §八「md_cg 距 L5 只差 I」所指的缺失内容)。
250
-
251
- **但缺一个可裁决项**:门槛要求 `U ≥ 6`,而实测 **S1 暴露"同条件不同值被静默覆盖"**——这正落在 U 维 7–8 档「误差驱动结构精化」的判据上。若该行为被判定为「改答案而非改结构」,U 可能应为 5.5,则 **L5 不达**。
252
-
253
- ```
254
- 裁决:DEFER
255
- 缺失维度:U 维「同条件空间·不同取值」的比对分支(已定位:consistency 引擎漏检该场景)
256
- 已取证(非待测):CONS.check(..., condition_space=CS_PROD, auto_flywheel=True)
257
- 对 32→64 返回 ACCEPT / "无冲突" / strength=0.0;unresolved 层未新增该条
258
- 验证路径(可裁决的下一步):在 consistency 的条件级比对中加入"同条件空间取值不一致"
259
- 判定分支,复跑 S1;若 S1 转为 DEFER/REJECT 并落 unresolved 或生成条件分支
260
- → U ≥ 6 稳固成立 → L5 达成
261
- 关注信号:unresolved 层新增"S1 同条件冲突"条目;mem_A / mem_B 间出现 distinct_from 边
262
- ```
263
-
264
- **综合分(按标尺 §4.1 公式)**:min = 6.0,mean = 6.93 → `6.0×0.4 + 6.93×0.6 = 6.56`(前次 5.8,+0.76)。
265
-
266
- **层次结论:维持 L4(认知级),L5 临界 —— 卡点从 I 维转移到 U 维的一个具体反例。**
267
-
268
- ---
269
-
270
- ## 十一、诚实边界
271
-
272
- 1. **自评**:本报告由被测系统自身产出(§二)。判据虽外置,观测位置不独立。
273
- 2. **C/D 对称的可替代解释**:C/D 双高也可由"任何条件 context 都触发同一过滤"产生;本报告以 A/B(无 context)对照 + 条件互斥设计降低该可能,但未做"给无关条件 context"的第三组对照。
274
- 3. **strict = 0**:L3 的拒绝机制是 S 维物理分桶,非 J 维语义判定;J 维分数(7.5)的"判定"成分未经单独剥离验证。
275
- 4. **密级证据的适用范围**:仅 `MdCGSecure`;裸 `MdCGOS` 无读隔离(§8.3)。
276
- 5. **U 维 S1 是真实反例**,已在 §六 / §十 记录,未做豁免。
277
- 6. **"用户偏好画像 / 元认知"未实测**:I 维上调到 6.0 是**基于 5–6 档要件**(持久身份 + 画像 + 变更史)的判定,不代表 7–8 档成立。其中"画像"部分:`identity` 接口支持 `subject_kind=user`,但**库内 0 条 user 主体实证节点**(`self/` 26 条、`agent:*` 若干,`user` 无),故 5–6 档的"画像"要件目前由 `self` 锚点承担,user 画像为空槽位;元认知工具(`cg(op=metacognition)`,8 个 action)存在但未展开验证。
278
- 7. **测试规模**:L2 为 10 查询 / 1000 节点,非标尺原文的"完整 1000 条历史会话"语义;L3 为 20 组(与竞品同批语料)。
279
- 8. **I 维层错位不适用条件**:标尺 §三 I 维 7–8 档要件写作「自我锚点 + 元认知 + **世界模型**」。其中**世界模型属身体层**(AEIS `aeis/core.py` 的 `world_model` / `voxel_world` / `scene_simulator`;经 `md_cg/whitebox_kb/aeis_core/__init__.py:17-19` 明示"不带走、归 AEIS 库"),**不在本报告评估对象(大脑侧 md_cg)的能力边界内**,故**不作为 md_cg 的 I 维判分要件**。
280
- - 推论(对 md_cg 有利,但是事实):排除世界模型后,md_cg 的 I 维 7–8 档门槛要件收缩为「自我锚点 + 元认知」。自我锚点已实测成立(Q1/Q7);**剩余缺口仅"元认知未展开验证"**——而工具已存在(`cg(op=metacognition)`)。升级路径比标尺 §8 的原始判断更短。
281
- - 边界:本声明**只豁免 md_cg(大脑侧)**,不改变标尺对 AEIS 的 I 维判定(AEIS 内含身体,世界模型对它是本层能力,判定成立)。
282
-
283
- ---
284
-
285
- ## 十二、附:证据索引
286
-
287
- | 维度 | 核心证据 |
288
- |---|---|
289
- | L2 | `result_l2.json`、`result_l2_strict.json` |
290
- | L3 | `result_l3v2_A/B/C/D.json`、本地实验工作区 `docs/experiments/deja-vu-negcond/corpus.py`(同批 GROUPS,不入库) |
291
- | L4 | `result_l4.json`(S1–S4)、`result_l4b.json`(回滚 + `rollback_of`) |
292
- | L5 | `result_l5.json`、`result_l5b.json`、`result_t2.json`(真版本链)、`_store/l5_replica/` |
293
- | C | `result_ct.json`(budget 60/120/400) |
294
- | T | `result_t2.json`(墓碑/写保护/密级)、`_protected_audit.jsonl`、`_protected_history/` |
295
- | 源码 | `md_cg/mdcos.py`(`MdCGOS` / `MdCGSecure`)、`md_cg/protect.py`、`md_cg/self_state.py`、`md_cg/security.py` |
296
-
297
- ---
298
-
299
- *报告生成:2026-09-12 | 标尺版本 v1.0 | 复核性质:§2.3 四门槛实跑 | 证据等级:实测级(落盘 JSON)*
@@ -1,239 +0,0 @@
1
- # md_cg · AGI 七维评分报告 v2.0
2
-
3
- > 评分对象:**dsh-memory / md_cg(认知图 + 记忆 OS)** 当前实现(2026-09-10)
4
- > 评分标尺:`AEIS/docs/AGI记忆系统评分标尺_v1.0.md`(七维 · 层次优先 · 门槛晋级)
5
- > 综合公式:`综合 = min × 0.4 + mean × 0.6`
6
- > 取证:源码级(`md_cg/*.py` 逐模块核对)+ 测试级(**46 套 1970 项** + 并发锁 **300 项** 全绿)
7
- > + 基准级(`bench_membench` SNR 量化)+ 实库级(`AEIS/data/mdcg` **6339 节点** 只读盘点)
8
- > 立场:不虚高、不虚低。每分给依据;无法验证的项单列、不计分。
9
- > 上一版:`AGI七维评分报告_md_cg_v1.0.md`(综合 8.0)
10
-
11
- ---
12
-
13
- ## 一、结论
14
-
15
- | 维度 | v1.0 | **v2.0** | 档位 | 一句话依据 |
16
- |---|---|---|---|---|
17
- | **S 结构** | 8.5 | **8.5** | 认知图(四要素 3.5/4) | 新增层归位闭环(实库 228 条 contextual)+ 视觉证据面结构化;conditions 四维/execution schema 仍未强制 |
18
- | **R 检索** | 8.5 | **8.5** | 因果路由 ✓ / 意图理解 △ | 新增**召回分池/降权**修复结构性挤占(索引占比 44.2%→27.1%);意图理解仍是规则层 |
19
- | **J 判断** | 9.0 | **9.0** | 独立元认知 + 主动遗忘 | 新增抽检工单/只读预演/闸门;主判定 REJECT 仍简化版、元认知只建议不执行 |
20
- | **C 调用** | 8.5 | **9.0** ↑ | 精准注入 + **SNR 可量化** | **SNR 量化仪表盘落地**(40% 噪声率下 sigP 97.8% / noiseP 0.4%,带配对显著性 + 消融对照) |
21
- | **U 演化** | 7.5 | **8.5** ↑↑ | 飞轮 + **结构可版本回滚** | 两条扣分解除:`consolidate` 进 MCP 主循环;**通用结构变更账本 + 按条目回滚**(实库 238 条、已演练 4 次回滚) |
22
- | **I 连续** | 8.0 | **8.5** ↑ | 自我锚点 + 元认知 + 结构变更史 | 演化账本由「审计留痕」升级为**跨会话可读叙事 + 规律统计**;跨副本收敛仍未实测 |
23
- | **T 可信** | 8.5 | **8.5** | 加密 + 全链路审计 + 信任可度量 | 视觉证据脱敏纪律 + 写入预演/闸门/留痕;零信任仍未落地 |
24
-
25
- ```
26
- min = 8.5 (S/R/U/I/T 并列) ← v1.0 为 7.5(U)
27
- mean = 8.643 (S8.5 R8.5 J9.0 C9.0 U8.5 I8.5 T8.5)
28
- 综合 = 8.5 × 0.4 + 8.643 × 0.6 = 8.586 ≈ 8.6
29
- ```
30
-
31
- ### 层次判定:**L5(AGI 级记忆)· 稳固段**
32
-
33
- | 晋级 | 门槛 | 实测 | 结论 |
34
- |---|---|---|---|
35
- | → L3 | J ≥ 5 且 S ≥ 5 | 9.0 / 8.5 | ✅ |
36
- | → L4 | U ≥ 5 且 J ≥ 6 | 8.5 / 9.0 | ✅ |
37
- | → **L5** | **I ≥ 6 且 U ≥ 6 且 T ≥ 6** | **8.5 / 8.5 / 8.5** | ✅ |
38
-
39
- **五问定层法复核**(标尺 §2.2):
40
-
41
- ```
42
- Q1 能存吗? ✅ 五层 + 负记忆 + 目标槽 + 子图
43
- Q2 能从一堆里找对那条? ✅ 四路 RRF + 条件路由 + 因果链 + 分池配比
44
- Q3 能说「这条不该用」吗? ✅ 四态判定 / 条件级负路由 / 负记忆
45
- Q4 说错后改结构还是改答案?✅ 改结构,且**改结构本身可版本化回滚**(238 条账本 + dry_run 回滚)
46
- Q5 一个月后还是同一个「我」吗?✅ self/anchor 不可覆盖 + identity 锚点 + **结构变更史叙事**
47
- ```
48
-
49
- > 与 v1.0 的差别:v1.0 判「L5 入门」,短板唯一在 U(7.5)。本次 U 一次抬 1.0,**七维全部 ≥ 8.5,无 <8 维度**,故判 **L5 稳固段**。
50
- > 但「稳固段」≠「接近满分」:五个 8.5 维度的共同上限是**「机制齐备、门槛项未齐」**(零信任 / 开放域意图 / 条件四维强制 / 穷尽去污染),见 §五。
51
-
52
- ---
53
-
54
- ## 二、逐维依据(相对 v1.0 的增量)
55
-
56
- ### S · 结构 · 8.5(持平)
57
-
58
- **新增加分**
59
- - **层归位闭环**(G4):`consolidate.contextualize_prefixes` 判据(流水账形态 + 视觉产物落 `knowledge` 错位)→ `contextualize_rollback` / `contextualize_history`;**真实库已落地 228 条 `contextual`**(`_maintain.jsonl` = 229 `contextualize` + 1 `contextualize_rollback`)。
60
- - **层级对账精确吻合**:`knowledge 6102 + contextual 228 + self 5 + anchor 4 = 6339` = 索引节点数(无幽灵节点、无漏计)。
61
- - **视觉证据面**(G5):`vision_evidence.py` 把「部件节点证据面恒为 `{}`」的口径缺口补齐(6 字段结构化 + `cond_hash` 连接 + 图集编号脱敏),属 `modality` 之外的**证据面**结构化。
62
-
63
- **扣分(与 v1.0 相同,未动)**:① conditions 理论四维仍只有 `time_window` 强制;② execution 仍非一等 frontmatter(无 schema 校验);③ 衰减仍走抽样判陈旧,非连续时间因子。
64
-
65
- > 7–8 档「图结构 + 条件绑定」扎实达标,层归位是 9 档构成项之一 → 维持 **8.5**。
66
-
67
- ### R · 检索 · 8.5(持平,但结构性缺陷修复)
68
-
69
- **新增加分**
70
- - **召回分池/降权**(§七 条件化 ACCEPT 落地):`pooling.py` —— 三池(knowledge / index / negative)、**池额守恒**(`cap_ratio` 之和恒 1.0,最大余数法)、**显式可 A/B 权重表**、空池额度回流不增预算、池账审计面(`quota/cands/taken/lost`)。
71
- - **真实库实测(6339 节点,48 条确定性查询,关闭态 vs 启用态同口径)**:
72
-
73
- | 口径 | 关闭 | 启用 | delta |
74
- |---|---|---|---|
75
- | **index_share** | **0.4417** | **0.2708** | **−0.1708** |
76
- | **knowledge_share** | 0.5583 | 0.7292 | +0.1708 |
77
- | p95_scanned / truncation_rate / truncation_loss | 6339 / 0.2292 / 0.18289 | 同 | **0**(不敏感口径) |
78
-
79
- → 索引/产物类不再吃满召回额度,**且 `scanned` 与全局截断率不变**(无额外 IO、无额度膨胀)。
80
-
81
- **扣分**:意图理解仍是确定性规则(归一化 + 目标定向 + `route()` 建议),非开放域理解。分池修复的是**配比**,不是**理解**。
82
-
83
- > 9–10 档要求「因果路由 + 意图理解」齐备——因果达标、意图仅规则层 → 维持 **8.5**(分池不计入晋级,仅计入依据)。
84
-
85
- ### J · 判断 · 9.0(持平)
86
-
87
- **新增加分**:G6 `refine.py` —— **口径先定的抽检工单** → **只读预演(`apply=False`)** → **闸门** → 校准 → 留痕/回滚(真实库 `_refine.jsonl` 已在运行);G4/G5 的层位移与证据回填同样走「预演 → 闸门 → 应用 → 可回滚」。
88
-
89
- **扣分(未动)**:① 主判定路径 REJECT 仍是简化版——`mdcg.py:746` 自注「需条件对比,**简化版用关键词命中**」,与 `consistency.py` 条件级判定未统一;② 元认知只建议不执行;③ 预测反馈需显式调用。
90
-
91
- > 9–10 档门槛项「独立元认知 + 主动遗忘」齐备 → 维持 **9.0**。
92
-
93
- ### C · 调用 · 8.5 → **9.0**(+0.5,v1.0 扣分①解除)
94
-
95
- **新增加分 —— SNR 量化仪表盘落地**
96
- v1.0 扣分①原文「**无 SNR 量化仪表盘**」。现 `bench_membench.py` 提供可比基准(1000 条 = **250 signal / 400 noise / 350 unlabeled,噪声率 40%**,噪声带强重复模式),且不是「跑一个好看的数字」,而是**带统计判定的实验台**:
97
-
98
- | 臂 | self@1 | self@10 | MRR | **sigP** | **noiseP** |
99
- |---|---|---|---|---|---|
100
- | 四路 | 89.2% | 94.3% | 0.911 | **97.8%** | **0.4%** |
101
- | 五路 (+fuzzy) | 79.1% | 92.3% | 0.832 | 95.7% | 0.8% |
102
-
103
- - 具备 **配对 bootstrap 5000 次 + McNemar 精确单侧**显著性、`--ablate {idcheck,shuffle,random,pool}` 消融对照、`--zero-importance` 去标签泄漏对照、`--sweep` 融合扫描、`--per-path` 逐路诊断。
104
- - §七 另提供**召回构成量化**(`index_share` / `knowledge_share` / `pool_lost_rate` + 池账)与**口径敏感性自检**(显式回 `insensitive_keys`/`sensitive_keys`)。
105
-
106
- **扣分(未动)**:② 去污染仍是 12/轮抽样 + 确定性判据,非穷尽语义蕴含;③ `_contradiction` 词级比对,跨表述矛盾不可识别。
107
-
108
- > 9 档门槛项「精准注入 + 可量化降噪」齐备 → **9.0**;「零污染」未达,故停在 9.0 下沿。
109
- > **特别声明**:升级依据是「**可量化 + 可回归**」,不是「指标好看」——该基准恰恰**暴露了缺陷**(五路在 `drop` 模式 self@1 **−29.2%**,p<0.001),见 §六·4。
110
-
111
- ### U · 演化 · 7.5 → **8.5**(+1.0,v1.0 两条扣分解除)
112
-
113
- v1.0 三条扣分,逐条复核:
114
-
115
- | v1.0 扣分 | 现状 | 证据 |
116
- |---|---|---|
117
- | ① `consolidate` 仅 CLI,`cg(op=...)` 无对应 op | **已解除** | `mcp_server.py` `cg(op=maintain, action=…)` 暴露 `induce` / `contextualize` / `contextualize_rollback` / `promote` / `promote_rollback` / `vision_evidence` / `refine` / `refine_gate` / `refine_calibrate` 等 → **进 agent 主循环** |
118
- | ② 无通用结构变更版本表 / 一键回滚(`protect` 只覆盖受保护节点) | **已解除** | `evolution.py`:`_evolution/ledger.md` 账本,`KIND = condition_gap / layer_shift / general / rollback`;每条含 `before/after` 状态快照(`state_of` + `diff`)、`rollback(entry_id, dry_run)` 按条目回滚、`history/patterns/summary`。**真实库 238 条(223 KB)**,其中 `layer_shift 229 / condition_gap 7 / rollback 4 / general 1` → **回滚路径已真实演练 4 次** |
119
- | ③ 预测校准需人工喂反馈 | **未变** | `predict.feedback` 仍为显式调用 |
120
-
121
- **新增加分**:`sustain` 新增 `evolve` 演化巡检(候补盘点 → 修复动作归口 → **只放行确定性动作**,`auto_evolve=False` 默认只记账),把「谁该被修」变成周期驱动而非人工巡检。
122
-
123
- **剩余扣分(半程)**:LLM 依赖的固化动作**不自动执行**(需 `heal(allow_evolve=True)` 放行确定性动作)→ 「**飞轮自运转**」仍缺最后一环。
124
-
125
- > 7–8 档核心 + 结构版本可回滚已达标 → **8.5**;9 档「全自转」未达。
126
-
127
- ### I · 连续 · 8.0 → **8.5**(+0.5,v1.0 扣分①缓解)
128
-
129
- **新增加分**
130
- - **跨会话可读的自我变更叙事**:`_evolution/ledger.md` 每条含「**规律** / 缺失条件 / **动作** / 证据 / 来源 / 时间」+ 状态快照,是**叙述性**文本而非纯字段留痕;配合 `evolution.patterns/summary` 做「哪一维条件反复缺失、由谁触发、哪些规律重复出现」的**规律统计** → 由 v1.0 的「审计/留痕」升级为「叙事 + 规律」。
131
- - **派生血缘**(G8):`provenance.py` —— `derived_from` 单一真相源、台账、悬空巡检、`rebuild_ledger` **只重放显式声明、不从 id 前缀反推**(守住「不回填」纪律)。
132
-
133
- **扣分(未变)**:② 跨副本状态收敛 / 冲突合并无实测(指纹只解决「读不到」);③ 元认知不写事实(设计如此)。
134
-
135
- > 9–10 档「跨会话生命史叙述**合成**」仍未达(当前是账本 + 规律,未合成连贯生命史)→ **8.5**。
136
-
137
- ### T · 可信 · 8.5(持平)
138
-
139
- **新增加分**
140
- - **视觉证据脱敏纪律**(G5):不读取任何图像、不重跑视觉、证据内只落结构化字段;图集目录名**不进库**(一律 `图集_N` 编号引用)。
141
- - **写入侧四重闸门**:预演(dry-run)→ 闸门 → 留痕(`_maintain` / `_evolution` / `_refine` / `_vision_evidence`)→ 可回滚。
142
- - 实库审计面在跑:`_audit.jsonl` / `_crypto.jsonl` / `_consistency.jsonl` / `_crosscheck.jsonl` / `_forgetting.jsonl` / `_metacognition.jsonl` / `_scrub.jsonl`。
143
-
144
- **扣分(未变)**:① 零信任未落地(主密钥仍在服务端进程可及,无远程身份校验 / 客户端持钥);② 非严格 E2E(服务端可解密;密文不参与全文索引的工程折衷);③ 威胁模型自认未覆盖本地内存取证 / 侧信道。
145
- **另如实标注**:`_protected_audit.jsonl` 在实库**尚无条目**(受保护节点审计未产生数据)。
146
-
147
- > 7–8 档「隐私树 + 全链路审计 + 防污染」完全达标,零信任缺位封顶 → 维持 **8.5**。
148
-
149
- ---
150
-
151
- ## 三、v1.0 → v2.0 对比
152
-
153
- | 维度 | v1.0 | v2.0 | Δ | 变化来源 |
154
- |---|---|---|---|---|
155
- | S 结构 | 8.5 | **8.5** | — | 层归位闭环(实库 228 条)· 视觉证据面结构化(不足以触及扣分项) |
156
- | R 检索 | 8.5 | **8.5** | — | 召回分池修复结构性挤占(index 占比 −17.1pp)· 意图理解仍未开放域 |
157
- | J 判断 | 9.0 | **9.0** | — | 抽检工单 / 只读预演 / 闸门 · REJECT 主判定仍简化版 |
158
- | **C 调用** | 8.5 | **9.0** | **+0.5** | `bench_membench` SNR 量化(sigP 97.8% / noiseP 0.4% + 配对显著性 + 消融对照)→ 解除「无 SNR 仪表盘」 |
159
- | **U 演化** | 7.5 | **8.5** | **+1.0** | `consolidate` 进 MCP 主循环 + 通用结构变更账本与按条目回滚(实库 238 条 / 回滚演练 4 次) |
160
- | **I 连续** | 8.0 | **8.5** | **+0.5** | 演化账本升级为跨会话叙事 + 规律统计 · 派生血缘(`provenance`) |
161
- | T 可信 | 8.5 | **8.5** | — | 视觉脱敏纪律 + 写入四重闸门 · 零信任仍未落地 |
162
- | **min / mean** | 7.5 / 8.357 | **8.5 / 8.643** | | |
163
- | **综合** | **8.0** | **8.6** | **+0.6** | |
164
- | **层次** | L5 入门 | **L5 稳固段** | | 无 <8 维度;短板由「单点 7.5」变为「五点 8.5 共同上限」 |
165
-
166
- > 归因说明:本次评的是**当前库整体**,增量来自 v1.0 之后的全部工作(P20–P43:演化账本/令牌/理论/链接/证据/权重/引用索引/文档索引/核对/会话存取/维护/洞察/回填/CCG 接线/批量核对/条件化断言/kp_ 对齐/条件空间/层归位/视觉证据/抽检工单/演化巡检/派生溯源/召回分池)。
167
-
168
- ---
169
-
170
- ## 四、测试与实测取证
171
-
172
- **测试级:46 套功能测试 1970 项 + 并发锁 300 项,全绿**
173
-
174
- ```
175
- P0 25 · P1 38 · P2 36 · P2-MCP 63 · P3 33 · P4 44 · P5 37 · P6 72 · P7 36 · P8 34
176
- P9 37 · P10 24 · P11 27 · P12 32 · P13 42 · P14 44 · P15 53 · P16 46 · P17 55 · P18 20
177
- P19 13 · P20 46 · P21 31 · P22 43 · P23 57 · P24 29 · P25 36 · P26 55 · P27 56 · P28 45
178
- P29 48 · P30 55 · P31 65 · P32 58 · P33 58 · P34 51 · P35 47 · P36 34 · P37 42 · P38 37
179
- P39 53 · P40 46 · P41 36 · P42 52 · P43 79 → 合计 1970
180
- test_lock 300/300
181
- ```
182
-
183
- > v1.0 名单(P0–P19 共 21 套)本轮实测为 811 项(P1 37→38、P5 36→37,测试有增补);新增 25 套共 1159 项。
184
-
185
- **基准级:SNR 量化(`bench_membench`,1000 条 / 噪声率 40%)**
186
-
187
- ```
188
- 四路 self@1 89.2% self@10 94.3% MRR 0.911 sigP 97.8% noiseP 0.4%
189
- 五路 self@1 79.1% self@10 92.3% MRR 0.832 sigP 95.7% noiseP 0.8%
190
- 配套:配对 bootstrap 5000 次 + McNemar 精确单侧;--ablate / --sweep / --per-path / --zero-importance
191
- ```
192
-
193
- **实库级:`AEIS/data/mdcg` 只读盘点(6339 节点)**
194
-
195
- | 项 | 实测 |
196
- |---|---|
197
- | 层分布 | `knowledge 6102` + `contextual 228` + `self 5` + `anchor 4` = **6339**(与索引精确吻合) |
198
- | 层归位 | `_maintain.jsonl` 230 条 = `contextualize 229` + `contextualize_rollback 1` |
199
- | 演化账本 | `_evolution/ledger.md` **238 条 / 223 KB**(`layer_shift 229` / `condition_gap 7` / `rollback 4` / `general 1`),每条含状态快照 |
200
- | 召回分池 | `index_share 0.4417→0.2708`、`knowledge_share 0.5583→0.7292`;`p95_scanned` / `truncation_*` delta = 0 |
201
- | 留痕面 | `_audit` / `_crypto` / `_consistency` / `_crosscheck` / `_forgetting` / `_identity` / `_metacognition` / `_refine` / `_scrub` / `_sustain` / `_vision_evidence` 均在运行 |
202
- | 空面(如实标注) | `_protected_audit.jsonl` 无条目;`rejected` / `unresolved` / `structural` / `goals` 层 0 条(能力存在,该库未产生该类数据) |
203
-
204
- **维度 ← 用例映射**:J←P9/P11/P12/P34/P35 · I←P10/P12/P14/P16/P42 · T←P13/P39 · U←P6/P8/P17/P20/P30/P41 · S←P2/P4/P37/P38 · R←P2/P8/P36/P43 · C←P15/P24/P31/P43 + `bench_membench`
205
-
206
- ---
207
-
208
- ## 五、短板与到下一档的差距
209
-
210
- | 维度 | v2.0 | 到 9.5+ 需补 |
211
- |---|---|---|
212
- | S 结构 | 8.5 | conditions 四维强制 + execution 一等 schema;衰减改连续时间因子 |
213
- | R 检索 | 8.5 | 开放域意图理解(脱离规则层) |
214
- | U 演化 | 8.5 | **飞轮全自转**(LLM 固化动作自主执行);预测反馈自动联动 |
215
- | I 连续 | 8.5 | 跨会话**生命史叙述合成**;跨副本状态收敛实测 |
216
- | T 可信 | 8.5 | **零信任**(远程身份校验 + 客户端持钥);密文可检索方案 |
217
- | J 判断 | 9.0 | 主判定 REJECT 与 `consistency` 条件级判定统一;元认知建议可执行 |
218
- | C 调用 | 9.0 | 去污染抽样→**穷尽 + 语义蕴含级**;跨表述矛盾识别 |
219
-
220
- **优先级(按「最小改动换最大抬升」)**
221
-
222
- 1. **U · 飞轮全自转**:现只自动执行**确定性**动作(`auto_evolve=False`),LLM 依赖的固化动作仍需人工放行。而「闸门 + 版本账本 + 回滚」已就绪——缺的只是**自动放行策略**,属最小改动冲击 9.0。
223
- 2. **T · 零信任**:8.5 的封顶项,但需基础设施(客户端持钥 / 远程身份校验),成本最高。
224
- 3. **R · 开放域意图**:规则层 → 模型层,与「有意剔除神经嵌入」的既有取向冲突,属**路线决策**而非工程补丁。
225
-
226
- > **关键提示**:现在 min = 8.5,且五个维度并列最低——**抬升单维不再抬升综合**。要突破 8.6,必须同时抬升最低维中的多数,或把 min 整体推过 9.0。
227
-
228
- ## 六、诚实边界声明
229
-
230
- 1. **U / I 无公开基准**:标尺自述「U / I 无公开基准,只能行为观察」。本报告的 8.5 / 8.5 属**代码能力 + 实库留痕判定**,非长周期实证;「长期压力验证的 L5」仍需真实运行数月方可确认。
231
- 2. **打分含主观性**:0.5 分粒度由评估者按判据落位,不同评估者可能 ±0.5(尤其 C 的 9.0 与 I 的 8.5,均处档位下沿)。
232
- 3. **「能实现」≠「已验证」**:未验证项(跨副本收敛、长周期自我连续性、零信任)已单列 §五,**不计入分数**。
233
- 4. **基准的作用是暴露缺陷,不是修饰分数**:`bench_membench` 显示五路(+fuzzy)在 `drop` 模式 self@1 **−29.2%**(95% CI [−0.352, −0.236],p<0.001)——该回归是**已知未修缺陷**。C 升 9.0 的依据是「该量可测、可回归、可消融」,而非「指标优秀」。
234
- 5. **实库存在「空面」**:`_protected_audit.jsonl` 无条目、`rejected`/`unresolved`/`structural`/`goals` 层 0 条 → 相关能力由测试(P9/P11/P15)取证,**不由实库取证**;两者性质不同,已分别标注。
235
- 6. **可证伪**:任一维度若发现反例(功能代码存在但不可调用、测试造假、口径不可复现),对应分数应下调。
236
-
237
- ---
238
-
239
- *报告生成:2026-09-10 · 标尺:`AGI记忆系统评分标尺_v1.0` · 对照版:`AGI七维评分报告_md_cg_v1.0` · 取证脚本已于生成后清理*