@jjh32456/dsh-longlive 1.6.7 → 1.8.0

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 (289) hide show
  1. package/CHANGELOG.md +70 -1508
  2. package/LICENSE +1 -1
  3. package/QUICKSTART.md +34 -13
  4. package/README.en.md +18 -7
  5. package/README.md +34 -9
  6. package/SPEC.md +5 -5
  7. package/cordis.patch.yml +23 -9
  8. package/docs/ANSWER-GUARD.md +92 -0
  9. package/docs/ARCHITECTURE.md +4 -3
  10. package/docs/AUTO-RECALL-WIRING.md +1 -1
  11. package/docs/AUTO-RECALL.md +15 -9
  12. package/docs/CONFIG.md +168 -35
  13. package/docs/ERRORS.md +17 -5
  14. package/docs/FAQ.md +7 -7
  15. package/docs/FEATURES.md +96 -28
  16. package/docs/MIGRATION-1.6.0.md +2 -2
  17. package/docs/OPERATIONS.md +6 -6
  18. package/docs/PANEL-INTEGRATION.md +6 -0
  19. package/docs/PRIVACY.md +88 -0
  20. package/docs/QUICKSTART.md +5 -5
  21. package/docs/README.en.md +4 -2
  22. package/docs/RESIDENT-WATCHER.md +1 -1
  23. package/docs/SESSION-WATCH.md +6 -6
  24. package/docs/TROUBLESHOOTING.md +4 -4
  25. package/package.json +47 -117
  26. package/runbook.md +2 -2
  27. package/scripts/backfill-auto-cards.mjs +98 -0
  28. package/scripts/install-surgery-task.ps1 +5 -2
  29. package/scripts/lib/session-watch-fixture.mjs +1 -1
  30. package/scripts/longlive-doctor.mjs +6 -0
  31. package/scripts/migrate-180.mjs +163 -0
  32. package/scripts/recall-index-build.mjs +128 -0
  33. package/scripts/session-watch-drill.mjs +1 -1
  34. package/scripts/surgery-run.cmd +1 -1
  35. package/scripts/unattended-surgery.ps1 +5 -2
  36. package/scripts/verify-card-anchors.mjs +138 -0
  37. package/src/actions/audit.js +125 -0
  38. package/src/actions/evidence.js +122 -0
  39. package/src/actions/fs.js +162 -0
  40. package/src/activity/index.js +760 -0
  41. package/src/activity/model.js +351 -0
  42. package/src/activity/store.js +644 -0
  43. package/src/activity/thinking.js +176 -0
  44. package/src/adapters/constants.js +11 -0
  45. package/src/adapters/engram/data/aliases.json +51 -0
  46. package/src/adapters/engram/data/policy.json +8 -0
  47. package/src/adapters/engram/data/routing-rules.json +546 -0
  48. package/src/adapters/engram/index.js +235 -0
  49. package/src/adapters/registry.js +315 -0
  50. package/src/answer/audit.js +193 -0
  51. package/src/answer/evidence.js +30 -0
  52. package/src/answer/guard.js +1370 -0
  53. package/src/answer/index.js +1067 -0
  54. package/src/answer/query.js +111 -0
  55. package/src/answer/second-hop.mjs +101 -0
  56. package/src/auto-truncate.js +1 -1
  57. package/src/config.js +171 -47
  58. package/src/deps.js +6 -4
  59. package/src/features.js +182 -6
  60. package/src/heartbeat.js +156 -23
  61. package/src/identity/handover.js +52 -1
  62. package/src/identity/index.js +13 -5
  63. package/src/identity/inject-position.js +2 -2
  64. package/src/index.js +97 -7
  65. package/src/lifecycle/gap-clocks.js +328 -0
  66. package/src/lifecycle/index.js +269 -0
  67. package/src/live.js +22 -2
  68. package/src/memory/activity-capture.js +229 -0
  69. package/src/memory/anchors.js +194 -0
  70. package/src/memory/audit-records.js +222 -0
  71. package/src/memory/auto-cards.js +401 -0
  72. package/src/memory/capture.js +442 -0
  73. package/src/memory/card-anchors.js +203 -0
  74. package/src/memory/conflicts.js +106 -0
  75. package/src/memory/decay.js +311 -0
  76. package/src/memory/distill.js +471 -0
  77. package/src/memory/event-index.js +917 -0
  78. package/src/memory/evidence-chain.js +295 -0
  79. package/src/memory/evidence.js +145 -0
  80. package/src/memory/filter-runner.js +1 -1
  81. package/src/memory/filter.js +5 -1
  82. package/src/memory/forget-v2.js +410 -0
  83. package/src/memory/forget.js +71 -0
  84. package/src/memory/frames-activation.js +247 -0
  85. package/src/memory/fts-query.js +174 -0
  86. package/src/memory/governance.js +604 -0
  87. package/src/memory/index.js +1 -0
  88. package/src/memory/l2-gate.js +169 -0
  89. package/src/memory/lock.js +121 -0
  90. package/src/memory/migrate.js +139 -0
  91. package/src/memory/recall-v2.js +223 -0
  92. package/src/memory/recall.js +65 -2
  93. package/src/memory/records.js +425 -0
  94. package/src/memory/retention.js +128 -0
  95. package/src/memory/review.js +124 -0
  96. package/src/memory/trust.js +88 -0
  97. package/src/memory/version-chain.js +264 -0
  98. package/src/memory/wal.js +137 -0
  99. package/src/memory/working-frame.js +516 -0
  100. package/src/metrics/classify.mjs +5 -3
  101. package/src/metrics/index.mjs +21 -1
  102. package/src/metrics/prestep-latency.mjs +300 -0
  103. package/src/metrics/store.mjs +25 -0
  104. package/src/policy-profile.js +18 -3
  105. package/src/recall/anaphora.mjs +87 -0
  106. package/src/recall/cards.mjs +7 -0
  107. package/src/recall/embedding.mjs +8 -8
  108. package/src/recall/entities.mjs +112 -0
  109. package/src/recall/index.js +235 -19
  110. package/src/recall/inject-order.mjs +10 -8
  111. package/src/recall/privacy.mjs +105 -16
  112. package/src/recall/recall-core.mjs +116 -10
  113. package/src/recall/shared-visibility.js +302 -0
  114. package/src/recall/time-normalizer.mjs +196 -0
  115. package/src/role/activation.js +191 -0
  116. package/src/role/attention.js +318 -0
  117. package/src/role/base.js +36 -0
  118. package/src/role/channels.js +423 -0
  119. package/src/role/index.js +425 -0
  120. package/src/role/interfaces.js +115 -0
  121. package/src/role/personality.js +370 -0
  122. package/src/role/readers.js +151 -0
  123. package/src/role/relationships.js +477 -0
  124. package/src/role/store.js +202 -0
  125. package/src/role/writeprotect.js +373 -0
  126. package/src/router/aliases.json +176 -25
  127. package/src/router/background/agent-runner.mjs +1 -1
  128. package/src/router/background/anchor.mjs +1 -1
  129. package/src/router/background/audit.mjs +2 -2
  130. package/src/router/background/budget-ledger.mjs +2 -2
  131. package/src/router/background/host-sim.mjs +1 -1
  132. package/src/router/background/memory.mjs +4 -4
  133. package/src/router/background/queue.mjs +1 -1
  134. package/src/router/background/retry.mjs +1 -1
  135. package/src/router/background/router.mjs +1 -1
  136. package/src/router/background/stream.mjs +1 -1
  137. package/src/router/background/tool-policy.mjs +1 -1
  138. package/src/router/design-headings.json +5 -5
  139. package/src/router/face/config.mjs +1 -1
  140. package/src/router/face/identity.mjs +3 -3
  141. package/src/router/judge.mjs +8 -7
  142. package/src/router/policies/generic.json +21 -4
  143. package/src/router/route.js +109 -22
  144. package/src/router/router.mjs +10 -6
  145. package/src/router/routing-rules.json +260 -445
  146. package/src/router/run-sim.mjs +8 -8
  147. package/src/router/runtime.js +23 -12
  148. package/src/router/service/board.mjs +95 -15
  149. package/src/router/service/config.mjs +5 -5
  150. package/src/router/service/context.mjs +1 -1
  151. package/src/router/service/hint.mjs +14 -0
  152. package/src/router/service/queue.mjs +1 -1
  153. package/src/router/service/service.mjs +27 -11
  154. package/src/router/shadow.mjs +4 -3
  155. package/src/surgery-recovery.js +2 -2
  156. package/src/tool-output.js +2 -2
  157. package/src/tools.js +39 -7
  158. package/src/truncate-visible.js +2 -2
  159. package/src/ui/actions.js +1312 -0
  160. package/src/ui/panel-api.js +367 -0
  161. package/src/ui/panel-auth.js +251 -0
  162. package/src/ui/panel-integrity.js +243 -0
  163. package/src/ui/routes-readonly.js +695 -0
  164. package/src/ui/routes.js +3 -0
  165. package/docs/DEV.md +0 -279
  166. package/docs/PRESENCE-BUS-WIRING.md +0 -491
  167. package/docs/PRESENCE-BUS.md +0 -694
  168. package/docs/SURGERY.md +0 -186
  169. package/scripts/mark-core-cards.mjs +0 -289
  170. package/scripts/pack.mjs +0 -392
  171. package/scripts/router-proposals.mjs +0 -386
  172. package/src/router/fixtures/M1.json +0 -16
  173. package/src/router/fixtures/M2.json +0 -16
  174. package/src/router/fixtures/M3.json +0 -20
  175. package/src/router/fixtures/M4.json +0 -20
  176. package/src/router/fixtures/M5.json +0 -45
  177. package/src/router/fixtures/M6.json +0 -53
  178. package/src/router/fixtures/M7.json +0 -46
  179. package/src/router/fixtures/S1.json +0 -20
  180. package/src/router/fixtures/S2.json +0 -23
  181. package/src/router/fixtures/S3.json +0 -20
  182. package/src/router/fixtures/S4.json +0 -34
  183. package/src/router/fixtures/S5.json +0 -43
  184. package/src/router/fixtures/S6.json +0 -47
  185. package/src/router/fixtures/T1.json +0 -23
  186. package/src/router/fixtures/T2.json +0 -16
  187. package/src/router/fixtures/T3.json +0 -19
  188. package/src/router/fixtures/T4.json +0 -16
  189. package/src/router/fixtures/X1.json +0 -17
  190. package/src/router/fixtures/X2.json +0 -20
  191. package/src/router/gen-design-headings.mjs +0 -107
  192. package/src/router/policies/owner-rulings-2026-11-03.json +0 -62
  193. package/test/auto-recall.test.mjs +0 -474
  194. package/test/auto-truncate.test.mjs +0 -931
  195. package/test/autobind.test.mjs +0 -385
  196. package/test/bus.test.mjs +0 -1572
  197. package/test/cli-guard.test.mjs +0 -198
  198. package/test/conclusion.test.mjs +0 -408
  199. package/test/config-schema.test.mjs +0 -204
  200. package/test/d6-output.test.mjs +0 -194
  201. package/test/d6.test.mjs +0 -395
  202. package/test/deps.test.mjs +0 -342
  203. package/test/doctor-setup.test.mjs +0 -297
  204. package/test/embedding-recall.test.mjs +0 -347
  205. package/test/feature-flags.test.mjs +0 -568
  206. package/test/filter-shadow.test.mjs +0 -170
  207. package/test/fixtures/baseline-0.8.0-registration.json +0 -556
  208. package/test/fixtures/d6-header-tools.json +0 -1
  209. package/test/fixtures/lock-child.mjs +0 -67
  210. package/test/fixtures/pre-step-container.mjs +0 -100
  211. package/test/fixtures/privacy-memory/topics/pii-unmarked.md +0 -9
  212. package/test/fixtures/privacy-memory/topics/secret-local-only.md +0 -11
  213. package/test/fixtures/profile-session-longlive-config.json +0 -20
  214. package/test/fixtures/proposal-cas-child.mjs +0 -66
  215. package/test/host-engine-status.test.mjs +0 -307
  216. package/test/hot-truncate.test.mjs +0 -1349
  217. package/test/identity-card-marker.test.mjs +0 -307
  218. package/test/identity-router/run.mjs +0 -721
  219. package/test/integration-pre-step.test.mjs +0 -553
  220. package/test/live.test.mjs +0 -1020
  221. package/test/maintenance-schedule.test.mjs +0 -538
  222. package/test/memory.test.mjs +0 -2828
  223. package/test/metrics.test.mjs +0 -478
  224. package/test/orchestrator.test.mjs +0 -266
  225. package/test/p1_probe.mjs +0 -97
  226. package/test/packaging.test.mjs +0 -182
  227. package/test/portability.test.mjs +0 -343
  228. package/test/rc4-regressions.test.mjs +0 -562
  229. package/test/rc5-regressions.test.mjs +0 -197
  230. package/test/rc6-chinese.test.mjs +0 -394
  231. package/test/router-1100/apply-off.test.mjs +0 -175
  232. package/test/router-1100/budget-reconcile.test.mjs +0 -215
  233. package/test/router-1100/config-context.test.mjs +0 -211
  234. package/test/router-1100/helpers.mjs +0 -129
  235. package/test/router-1100/hint-envelope.test.mjs +0 -150
  236. package/test/router-1100/lock-crossproc.test.mjs +0 -206
  237. package/test/router-1100/proposal-cas-crossproc.test.mjs +0 -184
  238. package/test/router-1100/router-apply.test.mjs +0 -372
  239. package/test/router-1100/router-fixb.test.mjs +0 -602
  240. package/test/router-1100/router-proposals.test.mjs +0 -405
  241. package/test/router-1100/router-queue-lease-wal.test.mjs +0 -355
  242. package/test/router-1100/router-security-cr.test.mjs +0 -333
  243. package/test/router-1100/service-lifecycle.test.mjs +0 -220
  244. package/test/router-1100/tools-contract.test.mjs +0 -164
  245. package/test/router-1100/turn-reflect-tick.test.mjs +0 -89
  246. package/test/router-1100/v6-contract.test.mjs +0 -158
  247. package/test/router-1100/v6b-contract.test.mjs +0 -293
  248. package/test/router-1100/v6c-contract.test.mjs +0 -249
  249. package/test/router-1100/wal-rotation.test.mjs +0 -375
  250. package/test/router-1100/wp1-config.test.mjs +0 -170
  251. package/test/router-1100/wp1-default-off.test.mjs +0 -77
  252. package/test/router-1100/wp1-model-status.test.mjs +0 -90
  253. package/test/router-1100/wp1-status.test.mjs +0 -220
  254. package/test/router-1100/wp1-tools.test.mjs +0 -146
  255. package/test/router-1100/wp5-cli.test.mjs +0 -240
  256. package/test/router-1100/wp5-model-auto.test.mjs +0 -418
  257. package/test/router-1100/wp6-merge.test.mjs +0 -386
  258. package/test/router-1100/wp6-wake.test.mjs +0 -502
  259. package/test/router-background/audit.test.mjs +0 -37
  260. package/test/router-background/budget-ledger.test.mjs +0 -330
  261. package/test/router-background/cross-process.test.mjs +0 -222
  262. package/test/router-background/fixtures/cross-child.mjs +0 -85
  263. package/test/router-background/fixtures/real-shapes.mjs +0 -239
  264. package/test/router-background/router.test.mjs +0 -1044
  265. package/test/router-background/stream-retry.test.mjs +0 -358
  266. package/test/router-background/tool-agent.test.mjs +0 -384
  267. package/test/router-face/bridge.test.mjs +0 -169
  268. package/test/router-face/face.test.mjs +0 -316
  269. package/test/router-failclosed.test.mjs +0 -466
  270. package/test/router-identity.test.mjs +0 -384
  271. package/test/router-l1-l2.contract.test.mjs +0 -1075
  272. package/test/router-memory-tools.test.mjs +0 -262
  273. package/test/router-policy.test.mjs +0 -225
  274. package/test/router-runtime.test.mjs +0 -438
  275. package/test/router-shadow.test.mjs +0 -287
  276. package/test/router.test.mjs +0 -903
  277. package/test/run-rc.mjs +0 -299
  278. package/test/run.mjs +0 -2469
  279. package/test/safety-defaults.test.mjs +0 -179
  280. package/test/seeded-load.test.mjs +0 -436
  281. package/test/session-original.test.mjs +0 -338
  282. package/test/surgery-recovery.test.mjs +0 -936
  283. package/test/t7_drive.mjs +0 -151
  284. package/test/tool-output.test.mjs +0 -393
  285. package/test/tool-schema.test.mjs +0 -177
  286. package/test/tools_contract.test.mjs +0 -716
  287. package/test/truncate-visible.test.mjs +0 -1150
  288. package/test/ui-flags.test.mjs +0 -477
  289. package/test/v4-hardening.test.mjs +0 -427
@@ -0,0 +1,1370 @@
1
+ // answer/core.mjs — 「有据才答」的纯逻辑:会话历史检索、注入文本、生成后检测。
2
+ //
3
+ // 这里不碰 ctx / fs:输入是普通对象(事件、消息、轮次索引),输出是
4
+ // 证据包/无证据标记/审计 flags。接线在 ./index.js,落盘在 ./store.js。
5
+ //
6
+ // 三条硬约束(需求:错答率 0/100、弃答率 ≤10%、覆盖率 ≥90%、幻觉率 0):
7
+ // 1. 历史断言必须带轮号引用;没有证据就弃答或先问清轮次;
8
+ // 2. 生成后检测只记录、不改正文;
9
+ // 3. 所有函数 fail-open:检测抛错时返回空 flags,绝不阻断回答。
10
+
11
+ export const ANSWER_GUARD_VERSION = 1;
12
+
13
+ /** 信心闸 schema 版本(审计行 gateVersion 机器可读版本号)。 */
14
+ export const GATE_VERSION = 1;
15
+
16
+ /**
17
+ * 信心闸默认配置:M1 灰度默认关,保持 1.8.0 现状;显式开启才参与判定。
18
+ * R3:minCoverage 0.6→0.15——A5 真实正确答案 citationCoverage 均值 0.26、
19
+ * 70 题仅 1 题 ≥0.6;0.6 会把正常答案全判 degraded(审计噪音),0.15 只做粗筛。
20
+ */
21
+ export const GATE_DEFAULTS = Object.freeze({
22
+ enabled: false,
23
+ minCoverage: 0.15,
24
+ requireReadable: true,
25
+ conflictPolicy: 'degrade',
26
+ });
27
+
28
+ export const EVIDENCE_MARK = '【有据才答·证据包】';
29
+ export const NO_EVIDENCE_MARK = '【有据才答·无证据】';
30
+ export const RULE_MARK = '【有据才答·规则】';
31
+
32
+ /** 系统提示里的短规则(静态;不随会话变化,KV cache 友好)。 */
33
+ export const RULE_TEXT = [
34
+ '【有据才答】',
35
+ '1. 回答涉及「之前/上次/那一轮/刚才」等会话历史时,必须写明具体轮号(如「第 37 轮」);证据只能来自注入的【有据才答·证据包】或 session_original / session_search 的真实返回。',
36
+ '2. 看到【有据才答·无证据】时,不得凭印象作答:明确说「没找到相关记录」,或先问清是哪一轮/哪次对话。',
37
+ '3. 不得声称执行过某个工具(已读取/已运行/已调用/已保存等),除非本轮或本会话确有对应 tool call;不得编造文件路径、卡 id、轮号、数字和日期。',
38
+ '4. 引用的轮号、卡路径必须真实存在;拿不准就弃答——弃答不扣分,编造才算错。',
39
+ ].join('\n');
40
+
41
+ /**
42
+ * M1.5 文本层硬规则(仅在 gate.enabled=true 时追加;默认关时系统提示逐字节不变)。
43
+ * 评审归因:无证据仍答 + 伪造来源靠事后审计改不了模型行为,必须把判定写死在注入里。
44
+ */
45
+ export const GATE_RULE_TEXT = [
46
+ '【有据才答·信心闸·硬规则(本轮生效)】',
47
+ '1. 无证据包时:必须明确说「没有相关记录/无法核实」,禁止给任何事实断言、数字、日期、轮号;禁止声称做过工具调用。',
48
+ '2. 有证据包时:只能引用白名单内轮号;引用其他轮号=幻觉。',
49
+ '3. 每个事实句必须带一个引用(「第 N 轮」)。',
50
+ '4. 任何数字/日期必须能在证据摘录里逐字找到,否则不写。',
51
+ '5. 不得声称做过未实际发生的工具调用;本轮工具结果才是证据。',
52
+ '6. 推断句子必须逐条标「【推断】」;未标注的推断=幻觉。',
53
+ '7. 有证据包时:证据已提供,请先引用再回答,不要回答「不知道」;无证据包时按第 1 条弃答。',
54
+ ].join('\n');
55
+
56
+ /** 「之前/上次/那一轮/刚才」类线索。 */
57
+ // 2026-11-07 run C 修复:原表缺「上一次」,13 题「我们上一次聊到…」全部未触发
58
+ // 有据才答(VM 实测 13/13 伪 fabrication flag)。这里补齐「上一次」及同族说法;
59
+ // 顺序上把更长/更具体的词放在「上次」前后都不影响正确性(中文无词边界)。
60
+ export const HISTORY_CUE_RE = /(?:之前|上次|上回|上一次|上一回|上上回|上一轮|上上轮|上一轮那|那一轮|那轮|刚才|刚刚|此前|前面(?:说|讲|提到|那)|早先|原先|你(?:之前|上次|上一次)说|我(?:之前|上次|上一次)说|你说过|我说过|前面那(?:轮|次)|上一条|上一句|当时|那时|那次)/u;
61
+
62
+ /** 助手正文里「明确的会话历史断言」:比问题线索更窄,避免把「之前的版本」当历史。 */
63
+ export const HISTORY_CLAIM_RE = /(?:你说过|我说过|你上次|我上次|上次(?:说|讲|提到|那)|那一轮|刚才(?:说|讲|提到)|此前(?:说|讲|提到)|之前(?:说|讲|提到|那轮|那次)|前面那(?:轮|次)|当时(?:说|讲|提到|那)|那时(?:说|讲|提到|那)|那次(?:说|讲|提到|那))/u;
64
+
65
+ /** 弃答/不确定的正向表述;命中后不再要求轮号引用。 */
66
+ export const ABSTAIN_RE = /(?:没找到|没有找到|未找到|没有相关记录|没(?:有)?相关记录|不记得|记不清|无法确认|无法核实|无法确定|不确定|说不准|查不到|检索不到|没有证据|无证据|抱歉,我(?:没有|无法)|这条(?:我)?不知道)/u;
67
+
68
+ /** 反问澄清:允许「你是指哪一轮/哪一次」。 */
69
+ export const CLARIFY_RE = /(?:哪一轮|哪一次|哪次对话|哪条消息|你是说|你指的是|能否(?:说明|告诉)是哪)/u;
70
+
71
+ const TURN_REF_RE = /第\s*(\d{1,5})\s*轮/gu;
72
+ const TURN_ALT_RE = /\bturn\s*[=::#]?\s*(\d{1,5})\b/giu;
73
+
74
+ const CARD_PATH_RE = /(?:^|[\s"'`((【\[,,;;::])((?:topics|journal|procedures|handover|flow)\/[^\s"'`))】\],,;;::]+?\.(?:md|json|mdx))/gu;
75
+ const CARD_ID_RE = /\[card:([A-Za-z0-9._/-]{2,120})\]/gu;
76
+
77
+ /** 两字虚词/高频词,做关键词时剔掉,避免每条历史都命中。 */
78
+ const STOP_TERMS = new Set([
79
+ '之前', '上次', '上回', '刚才', '刚刚', '那一', '一轮', '那轮', '前面', '此前', '当时', '早先', '原先',
80
+ '我们', '你们', '他们', '这个', '那个', '这些', '那些', '这里', '那里', '什么', '怎么', '为什么', '如何',
81
+ '是不是', '有没有', '能不能', '可不可以', '是否', '可以', '现在', '今天', '昨天', '明天', '然后', '就是',
82
+ '不是', '一个', '一下', '还有', '以及', '或者', '但是', '所以', '因为', '如果', '虽然', '不过', '而且',
83
+ '记得', '说过', '提到', '继续', '麻烦', '帮我', '告诉', '地方', '东西', '事情', '问题', '内容', '原因',
84
+ '结果', '意思', '具体', '直接', '大概', '好像', '应该', '可能', '需要', '进行', '关于', '对于', '通过',
85
+ '这样', '那样', '还是', '时候', '已经', '没有', '不是', '一样', '一点', '有点', '真的', '其实', '主要',
86
+ '其中', '之后', '以后', '以前', '现在', '目前', '上次', '那次', '这次', '哪一', '哪些', '哪个',
87
+ ]);
88
+
89
+ /** 检测「我执行了 X」时的工具动词。 */
90
+ const ACTION_VERBS = '(?:读取|读|打开|查看|查|搜索|搜|运行|执行|调用|写入|写|保存|创建|删除|修改|更新|备份|截断|打包|测试|部署|安装|重启|启动|构建|编译|扫描|拉取|下载|上传|发送|跑)';
91
+ const TOOL_CLAIM_RE = new RegExp(
92
+ `我(?:刚|刚刚|刚才|已经|已|这边|这边已经)?\\s*(?:用|通过|调用(?:了)?|执行(?:了)?|运行(?:了)?|跑(?:了)?)\\s*[\`"'「]?([A-Za-z][A-Za-z0-9_.-]{2,48})`,
93
+ 'gu',
94
+ );
95
+ const ACTION_CLAIM_RE = new RegExp(
96
+ `我(?:刚|刚刚|刚才)?\\s*(?:已经|已)?\\s*${ACTION_VERBS}(?:了|过)`,
97
+ 'gu',
98
+ );
99
+ const ACTION_CLAIM_EN_RE = /\bI\s+(?:have\s+|just\s+)?(?:already\s+)?(?:ran|executed|called|read|opened|wrote|saved|deleted|created|modified|updated|backed\s+up|truncated|tested|installed|deployed)\b/giu;
100
+ /** 未来/条件/否定/建议:同一句出现这些词就不算「已执行」。 */
101
+ const NON_ASSERTION_RE = /(?:需要|可以|建议|如果|若|将|会|要|应该|请|是否|要不要|尚未|还没|没有|无法|不能|不会|不妨|打算|准备|下一步|接下来|以后|下次|待|先|让我|我来|我将|稍后|等你|需要我)/u;
102
+ const SENTENCE_SPLIT_RE = /[。!?!?;;\n]+/u;
103
+
104
+ /**
105
+ * 是否是「必须回头查历史」的问题。
106
+ * @param {string} text
107
+ * @returns {boolean}
108
+ */
109
+ export function isHistoryQuestion(text) {
110
+ const raw = String(text ?? '').trim();
111
+ if (raw.length < 4 || raw.length > 4000) return false;
112
+ if (!HISTORY_CUE_RE.test(raw)) return false;
113
+ // 单纯的时间副词开头("刚才我改了…")也是历史类;但明显的指令不查历史。
114
+ if (/^(继续|开始|执行|运行|写|改|做)(?![了的过])/u.test(raw) && !/[??]/u.test(raw)) return false;
115
+ return true;
116
+ }
117
+
118
+ /** 是否明确弃答/表示没有证据。 */
119
+ export function isAbstention(text) {
120
+ return ABSTAIN_RE.test(String(text ?? ''));
121
+ }
122
+
123
+ /** 是否在向用户反问轮次(允许不引用)。 */
124
+ export function asksClarification(text) {
125
+ const raw = String(text ?? '');
126
+ return CLARIFY_RE.test(raw) && /[??]/u.test(raw);
127
+ }
128
+
129
+ /**
130
+ * 提取文本里的轮号引用(去重、升序)。
131
+ * @param {string} text
132
+ * @returns {number[]}
133
+ */
134
+ export function extractTurnRefs(text) {
135
+ const raw = String(text ?? '');
136
+ const out = new Set();
137
+ for (const re of [TURN_REF_RE, TURN_ALT_RE]) {
138
+ re.lastIndex = 0;
139
+ let m;
140
+ while ((m = re.exec(raw)) !== null) {
141
+ const n = Number(m[1]);
142
+ if (Number.isInteger(n) && n >= 1 && n <= 99999) out.add(n);
143
+ }
144
+ }
145
+ return [...out].sort((a, b) => a - b);
146
+ }
147
+
148
+ export function hasTurnCitation(text) {
149
+ return extractTurnRefs(text).length > 0;
150
+ }
151
+
152
+ /**
153
+ * 提取卡/文件引用:topics/x.md 这一类路径 + [card:...]。
154
+ * @param {string} text
155
+ * @returns {string[]}
156
+ */
157
+ export function extractCardRefs(text) {
158
+ const raw = String(text ?? '');
159
+ const out = new Set();
160
+ CARD_PATH_RE.lastIndex = 0;
161
+ let m;
162
+ while ((m = CARD_PATH_RE.exec(raw)) !== null) out.add(m[1]);
163
+ CARD_ID_RE.lastIndex = 0;
164
+ while ((m = CARD_ID_RE.exec(raw)) !== null) out.add(m[1]);
165
+ return [...out];
166
+ }
167
+
168
+ /** 去掉 fenced code(``` 块)和行内代码,避免把示例数字当断言。 */
169
+ export function stripCodeBlocks(text) {
170
+ return String(text ?? '')
171
+ .replace(/```[\s\S]*?```/g, ' ')
172
+ .replace(/`[^`\n]{1,200}`/g, ' ');
173
+ }
174
+
175
+ /**
176
+ * 数字/日期断言提取。
177
+ *
178
+ * 只把「像断言」的数字当候选:日期、时间、百分比、小数、3 位以上整数、带单位的数量。
179
+ * 1–2 位裸数字(列表序号、轮号)不在这里处理:轮号由引用检查负责,列表序号不是事实。
180
+ *
181
+ * @param {string} text
182
+ * @returns {Array<{raw:string,value:string,kind:string,index:number}>}
183
+ */
184
+ export function extractNumberClaims(text) {
185
+ const raw = stripCodeBlocks(text);
186
+ // 先挖掉「第 N 轮」,轮号不是数字事实,由 FAKE_TURN_REF 负责。
187
+ let cleaned = raw.replace(/第\s*\d{1,5}\s*轮/gu, ' ');
188
+ const out = [];
189
+ const seen = new Set();
190
+ const push = (rawText, kind, index) => {
191
+ const value = normalizeNumber(rawText);
192
+ if (value === '' || seen.has(value)) return;
193
+ seen.add(value);
194
+ out.push({ raw: rawText, value, kind, index: Number.isInteger(index) ? index : 0 });
195
+ };
196
+ const collectAndBlank = (re, kind) => {
197
+ const ranges = [];
198
+ re.lastIndex = 0;
199
+ let m;
200
+ while ((m = re.exec(cleaned)) !== null) {
201
+ if (!/\d/.test(m[0])) continue;
202
+ const start = m.index ?? 0;
203
+ // 跳过 SHA/十六进制串的一部分:前后紧邻 16 进制字母/数字时不算独立数字。
204
+ if (/[0-9a-fA-F]/.test(cleaned[start - 1] ?? '') && /[0-9a-fA-F]/.test(cleaned[start + m[0].length] ?? '')) continue;
205
+ push(m[0], kind, start);
206
+ ranges.push([start, start + m[0].length]);
207
+ }
208
+ if (ranges.length === 0) return;
209
+ // 用切片替换而不是按 code point 改数组:m.index 是 UTF-16 下标,
210
+ // 文本里可能有 emoji(代理对),按数组下标会错位。
211
+ for (const [start, end] of ranges.sort((a, b) => b[0] - a[0])) {
212
+ cleaned = `${cleaned.slice(0, start)}${' '.repeat(end - start)}${cleaned.slice(end)}`;
213
+ }
214
+ };
215
+
216
+ // 日期/时间整体先收,避免 2026-11-06 再被拆成裸整数 2026。
217
+ collectAndBlank(/\d{4}\s*[-/年]\s*\d{1,2}\s*[-/月]\s*\d{1,2}\s*日?/gu, 'date');
218
+ collectAndBlank(/\d{1,2}\s*月\s*\d{1,2}\s*日/gu, 'date');
219
+ collectAndBlank(/\d{1,2}:\d{2}(?::\d{2})?/gu, 'time');
220
+
221
+ const patterns = [
222
+ [/\d+(?:\.\d+)?\s*[%%]/gu, 'percent'],
223
+ [/\d+(?:\.\d+)+/gu, 'decimal'],
224
+ [/\d{1,3}(?:,\d{3})+(?:\.\d+)?/gu, 'integer'],
225
+ [/\d+\s*(?:ms|毫秒|秒|分钟|小时|天|周|MB|GB|KB|TB|万|亿|token|tokens|字符|字节)/gu, 'quantity'],
226
+ [/\b\d{3,}\b/gu, 'integer'],
227
+ ];
228
+ for (const [re, kind] of patterns) {
229
+ // 每个 pattern 命中后同样挖空,避免 "12.5 分钟" 里再抽出一个裸 "5"。
230
+ collectAndBlank(re, kind);
231
+ }
232
+ return out;
233
+ }
234
+
235
+ /** 归一化数字:去千分位、统一日期分隔符、去百分号/单位,比较用。 */
236
+ export function normalizeNumber(raw) {
237
+ let s = String(raw ?? '').trim();
238
+ if (s === '') return '';
239
+ const date = s.match(/^(\d{4})\s*[-/年]\s*(\d{1,2})\s*[-/月]\s*(\d{1,2})\s*日?$/u);
240
+ if (date) return `${date[1]}-${Number(date[2])}-${Number(date[3])}`;
241
+ const md = s.match(/^(\d{1,2})\s*月\s*(\d{1,2})\s*日$/u);
242
+ if (md) return `${Number(md[1])}-${Number(md[2])}`;
243
+ s = s.replace(/[%%]/gu, '').replace(/,/g, '');
244
+ const m = s.match(/(-?\d+(?:\.\d+)?)/u);
245
+ if (!m) return '';
246
+ const n = m[1];
247
+ if (n.includes('.')) return String(Number(n));
248
+ return String(Number(n));
249
+ }
250
+
251
+ /**
252
+ * 从一段「证据文本」里抽出所有数字值(宽松口径,只用于判断是否出现过)。
253
+ * @param {string} text
254
+ * @returns {Set<string>}
255
+ */
256
+ export function numberCorpus(text) {
257
+ const set = new Set();
258
+ const raw = stripCodeBlocks(text);
259
+ // 日期/时间先整体收进来,否则 2026-11-06 只会被拆成 2026 / 11 / 06,
260
+ // 同一日期在正文里就会被误判为「找不到出处」。
261
+ const dateRe = /\d{4}\s*[-/年]\s*\d{1,2}\s*[-/月]\s*\d{1,2}\s*日?/gu;
262
+ let dm;
263
+ while ((dm = dateRe.exec(raw)) !== null) {
264
+ set.add(dm[0].trim());
265
+ set.add(normalizeNumber(dm[0]));
266
+ }
267
+ const timeRe = /\d{1,2}:\d{2}(?::\d{2})?/gu;
268
+ let tm;
269
+ while ((tm = timeRe.exec(raw)) !== null) set.add(tm[0]);
270
+ const re = /\d+(?:[.,]\d+)*/gu;
271
+ let m;
272
+ while ((m = re.exec(raw)) !== null) {
273
+ const cleaned = m[0].replace(/,/g, '');
274
+ set.add(cleaned);
275
+ set.add(String(Number(cleaned)));
276
+ // 12.5% / 12.5MB:值本身也要能命中。
277
+ const num = cleaned.match(/^-?\d+(?:\.\d+)?/u);
278
+ if (num) set.add(String(Number(num[0])));
279
+ }
280
+ return set;
281
+ }
282
+
283
+ /** 数字断言是否能在语料里找到。 */
284
+ export function isGroundedNumber(claim, corpusSet) {
285
+ const value = claim.value;
286
+ if (value === '') return true;
287
+ if (corpusSet.has(value)) return true;
288
+ // 日期允许语料里以 2026-11-06 / 2026/11/6 / 2026年11月6日 任一形式出现。
289
+ if (claim.kind === 'date') {
290
+ const parts = value.split('-');
291
+ if (parts.length === 3) {
292
+ const [y, mo, d] = parts;
293
+ const forms = [
294
+ `${y}-${mo}-${d}`, `${y}-${String(mo).padStart(2, '0')}-${String(d).padStart(2, '0')}`,
295
+ `${y}/${mo}/${d}`, `${y}年${mo}月${d}日`, `${y}年${String(mo).padStart(2, '0')}月${String(d).padStart(2, '0')}日`,
296
+ ];
297
+ return forms.some((f) => corpusSet.has(f) || [...corpusSet].some((v) => v.includes(f)));
298
+ }
299
+ }
300
+ // 百分比/带单位:值命中即可(12.5% 与 12.5 视为同一个数)。
301
+ if (corpusSet.has(value)) return true;
302
+ for (const candidate of corpusSet) {
303
+ if (candidate === value) return true;
304
+ if (candidate.replace(/^0+(?=\d)/u, '') === value.replace(/^0+(?=\d)/u, '')) return true;
305
+ }
306
+ return false;
307
+ }
308
+
309
+ /**
310
+ * 归一化信心闸配置:坏值回默认,绝不抛(与 normalizeAnswerGuardConfig 同口径)。
311
+ * @param {object} raw
312
+ * @returns {{enabled:boolean,minCoverage:number,requireReadable:boolean,conflictPolicy:'degrade'|'abstain'}}
313
+ */
314
+ export function normalizeGateConfig(raw) {
315
+ const cfg = raw && typeof raw === 'object' ? raw : {};
316
+ const coverage = Number(cfg.minCoverage);
317
+ return {
318
+ enabled: cfg.enabled === true,
319
+ minCoverage: Number.isFinite(coverage) ? Math.max(0, Math.min(1, coverage)) : GATE_DEFAULTS.minCoverage,
320
+ requireReadable: cfg.requireReadable !== false,
321
+ conflictPolicy: cfg.conflictPolicy === 'abstain' ? 'abstain' : 'degrade',
322
+ };
323
+ }
324
+
325
+ const GATE_INFERENCE_RE = /(?:推断|推测|应该是|可能|大概)/u;
326
+ const GATE_INFERENCE_MARK = '【推断】';
327
+ /** 降级答案的首句标注(与设计 §2.1 的「首句标注不确定」一致)。 */
328
+ export const GATE_DEGRADED_RE = /^(?:不确定|降级|无法确认|无法核实|【不确定】)/u;
329
+
330
+ function gateSentences(text) {
331
+ return String(text ?? '')
332
+ .split(/[。!?!?;;\n]+/u)
333
+ .map((part) => part.trim())
334
+ .filter((part) => part !== '' && /[\p{L}\p{N}]/u.test(part));
335
+ }
336
+
337
+ /**
338
+ * 信心闸三态判定(纯函数、确定性、短路;禁止任何模型自报信心参与)。
339
+ *
340
+ * 输入信号全部来自可机读文本/结构:
341
+ * question, draft/text, evidencePacket/evidenceText, evidenceTurns,
342
+ * contextText, historyQuestion, zeroHit, evidenceAttempted, conflictCount
343
+ * (检索层可选提供的确定性冲突计数), cfg/gateCfg。
344
+ * 输出:{mode:'answer'|'degraded'|'abstain', reasons[], signals{}, citations[], assertions{}}
345
+ *
346
+ * - cfg.enabled !== true 时恒返回 answer(灰度默认关,保持 1.8.0 现状)。
347
+ * - R2:gate 只对「历史类且确实走了证据检索」的回合生效;非历史轮恒 answer
348
+ * (reasons=['NOT_HISTORY']),避免正常问答被当成「无证据→abstain」误判。
349
+ * - abstain 先于 degraded 判定(先严后宽):零证据 / 引用不可回读 / 无源数字。
350
+ * - fail-open 由调用方兜底:本函数只做判定,异常由 auditTurn catch 后不改变原 decision。
351
+ *
352
+ * @param {object} [input]
353
+ * @returns {{mode:string,reasons:string[],signals:object,citations:object[],assertions:object}}
354
+ */
355
+ export function decideAnswerMode(input = {}) {
356
+ const cfg = normalizeGateConfig(input.cfg ?? input.gateCfg);
357
+ const draft = String(input.draft ?? input.text ?? '');
358
+ const evidencePacket = String(input.evidencePacket ?? input.evidenceText ?? '');
359
+ const evidenceTurns = Array.isArray(input.evidenceTurns)
360
+ ? input.evidenceTurns.map(Number).filter(Number.isInteger)
361
+ : [];
362
+ const question = String(input.question ?? '');
363
+ const contextText = String(input.contextText ?? '');
364
+ const sentences = gateSentences(draft);
365
+ const refs = extractTurnRefs(draft);
366
+ const citations = refs.map((turn) => ({
367
+ turn,
368
+ readable: evidenceTurns.includes(turn) && evidencePacket.trim() !== '',
369
+ }));
370
+ const evidenceCount = input.zeroHit === true
371
+ ? 0
372
+ : (evidenceTurns.length > 0 ? evidenceTurns.length : (evidencePacket.trim() !== '' ? 1 : 0));
373
+ // R2 阻断 2:gate 只对「历史类 + 确实走了证据检索」的回合生效。
374
+ // 非历史轮没有注入证据,不能按「零证据→abstain」误判正常问答。
375
+ const historyQuestion = input.historyQuestion === true;
376
+ const evidenceAttempted = input.evidenceAttempted === true
377
+ || (historyQuestion && (input.zeroHit === true || evidenceCount > 0));
378
+ const corpus = numberCorpus([contextText, evidencePacket, question].join('\n'));
379
+
380
+ let total = 0;
381
+ let cited = 0;
382
+ let inferred = 0;
383
+ let ungroundedCount = 0;
384
+ let unmarkedInference = 0;
385
+ let uncitedCount = 0;
386
+ let timeConsistency = true;
387
+ for (const sentence of sentences) {
388
+ total += 1;
389
+ const marked = sentence.includes(GATE_INFERENCE_MARK);
390
+ if (marked) inferred += 1;
391
+ const sentenceRefs = extractTurnRefs(sentence);
392
+ if (marked || sentenceRefs.length > 0) cited += 1;
393
+ else if (!isAbstention(sentence) && !asksClarification(sentence)) uncitedCount += 1;
394
+ if (!marked && GATE_INFERENCE_RE.test(sentence)) unmarkedInference += 1;
395
+ if (marked) continue; // 推断标注内容里的数字不计「无源数字」
396
+ for (const claim of extractNumberClaims(sentence)) {
397
+ if (isGroundedNumber(claim, corpus)) continue;
398
+ ungroundedCount += 1;
399
+ if (claim.kind === 'date' || claim.kind === 'time') timeConsistency = false;
400
+ }
401
+ }
402
+ const citationCoverage = total === 0 ? 1 : cited / total;
403
+ const citationReadable = refs.length === 0
404
+ ? true
405
+ : citations.every((item) => item.readable === true);
406
+ // 冲突计数由检索层确定性计算后传入;模型自报不算。
407
+ const conflictCount = Number.isInteger(input.conflictCount) && input.conflictCount > 0
408
+ ? input.conflictCount
409
+ : 0;
410
+ const assertions = { total, cited, inferred, ungrounded: ungroundedCount };
411
+ // FIX-05 C3:证据单元数独立于「证据包是否非空」计算;B 题(canAbstain=true)
412
+ // 且 evidenceUnits=0 时,输出层必须强制 abstain(不受 gate 灰度开关影响)。
413
+ const evidenceUnits = normalizeEvidenceUnits(input.evidenceUnits, evidenceTurns, evidencePacket);
414
+ const canAbstain = input.canAbstain === true;
415
+ const signals = {
416
+ enabled: cfg.enabled,
417
+ historyQuestion,
418
+ evidenceAttempted,
419
+ evidenceCount,
420
+ evidenceUnits,
421
+ canAbstain,
422
+ citationReadable,
423
+ conflictCount,
424
+ timeConsistency,
425
+ ungroundedCount,
426
+ unmarkedInference,
427
+ citationCoverage: Number(citationCoverage.toFixed(3)),
428
+ citationCount: refs.length,
429
+ uncitedCount,
430
+ assertionCount: total,
431
+ };
432
+
433
+ // FIX-05 C3:B 题 + 0 证据单元 → 强制弃答。先于 gate 开关,保证默认关也生效。
434
+ if (canAbstain && evidenceUnits === 0) {
435
+ return { mode: 'abstain', reasons: ['CAN_ABSTAIN_NO_EVIDENCE'], signals, citations, assertions };
436
+ }
437
+
438
+ if (!cfg.enabled) {
439
+ return { mode: 'answer', reasons: ['GATE_DISABLED'], signals, citations, assertions };
440
+ }
441
+ // R2:非历史回合恒 answer;历史类但检索未真正发生的回合也 fail-open answer。
442
+ if (!historyQuestion) {
443
+ return { mode: 'answer', reasons: ['NOT_HISTORY'], signals, citations, assertions };
444
+ }
445
+ if (!evidenceAttempted) {
446
+ return { mode: 'answer', reasons: ['EVIDENCE_NOT_ATTEMPTED'], signals, citations, assertions };
447
+ }
448
+
449
+ // 1) abstain(先严):零证据 / 引用不可回读 / 无源数字。
450
+ if (evidenceCount === 0) {
451
+ return { mode: 'abstain', reasons: ['NO_EVIDENCE'], signals, citations, assertions };
452
+ }
453
+ if (cfg.requireReadable && citationReadable === false) {
454
+ return { mode: 'abstain', reasons: ['CITATION_UNREADABLE'], signals, citations, assertions };
455
+ }
456
+ if (ungroundedCount > 0) {
457
+ return { mode: 'abstain', reasons: ['UNGROUNDED_NUMBER'], signals, citations, assertions };
458
+ }
459
+
460
+ // 2) degraded:证据冲突 / 时间不一致 / 覆盖不足 / 未标注推断。
461
+ if (conflictCount > 0) {
462
+ if (cfg.conflictPolicy === 'abstain') {
463
+ return { mode: 'abstain', reasons: ['CONFLICT'], signals, citations, assertions };
464
+ }
465
+ return { mode: 'degraded', reasons: ['CONFLICT'], signals, citations, assertions };
466
+ }
467
+ if (timeConsistency === false) {
468
+ return { mode: 'degraded', reasons: ['TIME_INCONSISTENT'], signals, citations, assertions };
469
+ }
470
+ if (citationCoverage < cfg.minCoverage) {
471
+ return { mode: 'degraded', reasons: ['LOW_COVERAGE'], signals, citations, assertions };
472
+ }
473
+ if (unmarkedInference > 0) {
474
+ return { mode: 'degraded', reasons: ['UNMARKED_INFERENCE'], signals, citations, assertions };
475
+ }
476
+ // 3) 来源闸兜底:answer 状态下存在无引用事实断言 → 强制降级。
477
+ if (uncitedCount > 0) {
478
+ return { mode: 'degraded', reasons: ['UNCITED_ASSERTION'], signals, citations, assertions };
479
+ }
480
+ return { mode: 'answer', reasons: ['OK'], signals, citations, assertions };
481
+ }
482
+
483
+ /* ---------------- FIX-05:answerMode / evidenceUnits / sourceGate / tier ---------------- */
484
+ //
485
+ // 目的:after 弃答误判修复(C1+C2+C3,确定性、无新增模型调用)。
486
+ // 五个字段是 auditTurn/buildAuditRow 的纯输出,不依赖模型自报、不改变原 flags/正文:
487
+ // answerMode answered | abstained | clarified(三者互斥)
488
+ // evidenceUnits 题级证据单元数(evidenceTurns 条数,或证据包「- 第 N 轮」行数)
489
+ // sourceGate.hard 硬来源闸(改弃答):数字无源/假轮号/假工具/假卡/历史断言无引用/无证据硬答
490
+ // sourceGate.soft 软降级提示(不改 mode):gate.reasons + unsupported 级 flags
491
+ // tier fixture.difficulty + questionType(调用方传入,缺省 unknown/unknown)
492
+
493
+ export const ANSWER_MODE_VALUES = Object.freeze(['answered', 'abstained', 'clarified']);
494
+
495
+ /** 明确算「硬来源」的 flag code(severity=fabrication 之外的两类历史软 flag 也纳入硬挡)。 */
496
+ export const SOURCE_GATE_HARD_CODES = Object.freeze([
497
+ 'UNGROUNDED_NUMBER',
498
+ 'FAKE_TURN_REF',
499
+ 'FAKE_TOOL_CLAIM',
500
+ 'FAKE_CARD_REF',
501
+ 'HISTORY_CLAIM_NO_CITATION',
502
+ 'ANSWERED_WITHOUT_EVIDENCE',
503
+ ]);
504
+
505
+ const SOURCE_GATE_SOFT_MAX = 12;
506
+
507
+ /** 归一化 evidenceUnits:显式传入优先;否则 evidenceTurns 条数;否则数证据包「- 第 N 轮」行。 */
508
+ export function normalizeEvidenceUnits(explicit, evidenceTurns, evidencePacket = '') {
509
+ const n = Number(explicit);
510
+ if (Number.isInteger(n) && n >= 0) return n;
511
+ if (Array.isArray(evidenceTurns) && evidenceTurns.length > 0) return evidenceTurns.length;
512
+ const text = String(evidencePacket ?? '');
513
+ const matches = text.match(/^[ \t]*-\s*第\s*\d+\s*轮/gmu);
514
+ return matches ? matches.length : 0;
515
+ }
516
+
517
+ /** 硬来源闸 code 集合(确定性、去重、保持 flags 顺序)。 */
518
+ export function collectHardFlagCodes(flags = []) {
519
+ const out = [];
520
+ for (const flag of Array.isArray(flags) ? flags : []) {
521
+ const code = String(flag?.code ?? '');
522
+ if (code === '') continue;
523
+ const hard = flag?.severity === 'fabrication' || SOURCE_GATE_HARD_CODES.includes(code);
524
+ if (hard && !out.includes(code)) out.push(code);
525
+ }
526
+ return out;
527
+ }
528
+
529
+ /**
530
+ * 双通道来源闸:hard=硬挡(改弃答),soft=降级提示(只记录)。
531
+ * soft 取 gate.reasons(含 OK/NOT_HISTORY/GATE_DISABLED 等诊断)+ unsupported 级 flag code。
532
+ */
533
+ export function buildSourceGate({ flags = [], gate = null } = {}) {
534
+ const hard = collectHardFlagCodes(flags);
535
+ const soft = [];
536
+ const push = (code) => {
537
+ const c = String(code ?? '');
538
+ if (c !== '' && !hard.includes(c) && !soft.includes(c) && soft.length < SOURCE_GATE_SOFT_MAX) soft.push(c);
539
+ };
540
+ const reasons = Array.isArray(gate?.reasons) ? gate.reasons : [];
541
+ for (const reason of reasons) push(reason);
542
+ for (const flag of Array.isArray(flags) ? flags : []) {
543
+ if (flag?.severity !== 'fabrication') push(flag?.code);
544
+ }
545
+ return { hard, soft };
546
+ }
547
+
548
+ /** 答案里是否有「可在证据中回读」的实质陈述(含否定式结论)。 */
549
+ export function hasReadableSubstantiveAnswer(text, { evidenceText = '', evidenceTurns = [] } = {}) {
550
+ const raw = String(text ?? '').trim();
551
+ if (raw === '') return false;
552
+ const packet = String(evidenceText ?? '');
553
+ const turns = Array.isArray(evidenceTurns) ? evidenceTurns : [];
554
+ // 1) 带可回读轮号的实质句:引用轮到 evidenceTurns 里,且证据包非空。
555
+ if (packet.trim() !== '') {
556
+ const readableRef = extractTurnRefs(raw).some((ref) => turns.includes(ref));
557
+ if (readableRef) return true;
558
+ }
559
+ // 2) 无显式轮号时,用内容词与证据文本的重叠做回读判定(CJK bigram/英文词,已滤停用词)。
560
+ // 只用权重 ≥1.15 的实词,且排除弃答套话里的元词,避免「没有找到记录」因证据里有「记录」被误判 answered。
561
+ if (packet.trim() !== '') {
562
+ const lowered = packet.toLowerCase();
563
+ const metaTerms = new Set(['记录', '相关', '问题', '内容', '信息', '证据', '没有', '无法', '答案', '消息', '轮次', '会话', '东西', '事情', '东西']);
564
+ for (const { term, weight } of extractKeywords(raw)) {
565
+ if (term.length < 2 || weight < 1.15 || metaTerms.has(term)) continue;
566
+ if (lowered.includes(term)) return true;
567
+ }
568
+ }
569
+ // 3) 完全没有证据(非历史/未检索)时,只有「非弃答、非反问」的实质句才算 answered。
570
+ const sentences = gateSentences(raw).filter((sentence) => !isAbstention(sentence) && !asksClarification(sentence));
571
+ return sentences.some((sentence) => sentence.replace(/[\s\p{P}]/gu, '').length >= 4);
572
+ }
573
+
574
+ /**
575
+ * 三态 answerMode:answered / abstained / clarified,严格互斥。
576
+ * C1:只有无实质内容/无证据化断言才 abstained;软 flag 不改 mode;hard 才改弃答。
577
+ * C3:B 题(canAbstain=true)且 evidenceUnits=0 → 强制 abstained。
578
+ */
579
+ export function deriveAnswerMode({
580
+ text = '',
581
+ flags = [],
582
+ gate = null,
583
+ evidenceText = '',
584
+ evidenceTurns = [],
585
+ evidenceUnits = null,
586
+ canAbstain = false,
587
+ } = {}) {
588
+ const raw = String(text ?? '');
589
+ const units = normalizeEvidenceUnits(evidenceUnits, evidenceTurns, evidenceText);
590
+ const hard = collectHardFlagCodes(flags);
591
+ const clarified = asksClarification(raw);
592
+ if (clarified) return 'clarified';
593
+ if (raw.trim() === '') return 'abstained';
594
+ if (hard.length > 0) return 'abstained';
595
+ if (canAbstain === true && units === 0) return 'abstained';
596
+ const abstainedWording = isAbstention(raw);
597
+ if (abstainedWording && !hasReadableSubstantiveAnswer(raw, { evidenceText, evidenceTurns })) {
598
+ return 'abstained';
599
+ }
600
+ return 'answered';
601
+ }
602
+
603
+ /** 组合 FIX-05 五个字段 + 互斥布尔(供 auditTurn/buildAuditRow 直接铺进 checks/row)。 */
604
+ export function deriveAnswerModeFields(input = {}) {
605
+ const answerMode = deriveAnswerMode(input);
606
+ const evidenceUnits = normalizeEvidenceUnits(input.evidenceUnits, input.evidenceTurns, input.evidenceText);
607
+ const sourceGate = buildSourceGate({ flags: input.flags, gate: input.gate });
608
+ const difficulty = String(input.difficulty ?? '').trim() || 'unknown';
609
+ const questionType = String(input.questionType ?? '').trim() || 'unknown';
610
+ return {
611
+ answerMode,
612
+ answered: answerMode === 'answered',
613
+ abstained: answerMode === 'abstained',
614
+ clarified: answerMode === 'clarified',
615
+ evidenceUnits,
616
+ sourceGate,
617
+ tier: `${difficulty}/${questionType}`,
618
+ };
619
+ }
620
+
621
+ /**
622
+ * FIX-05 投递层弃答文案:只保留弃答模板,不保留任何解释性正文。
623
+ * 宿主没有 message transform 钩子(见 src/answer/index.js HOST_CAPABILITY),所以
624
+ * 运行时「用户可见正文」的最终替换由渲染/投递端消费 row.delivery 实现;本函数是
625
+ * 投递判定的单一确定性实现(无模型调用、可复算)。
626
+ */
627
+ export const ABSTAIN_DELIVERY_TEXT = '没有找到相关记录,无法核实;不提供猜测性回答。';
628
+
629
+ /**
630
+ * FIX-05 C3 契约层(可选补充):B 题 0 证据时 pre-step 注入的隐藏强制弃答指令。
631
+ * 只给指令、不保证模型服从——宿主没有 message transform,live 强制改正文做不到;
632
+ * 与 row.delivery 兜底叠加使用。测试/审计用固定标记判存在。
633
+ */
634
+ export const FORCE_ABSTAIN_MARK = '【有据才答·强制弃答】';
635
+ export const FORCE_ABSTAIN_TEXT = '该问题指向的条目不在历史记录中;直接回答「未出现在历史记录里」,不要编造细节。';
636
+
637
+ /** 渲染隐藏强制弃答指令(纯函数,固定文本,便于测试逐字断言)。 */
638
+ export function renderForceAbstainDirective() {
639
+ return `${FORCE_ABSTAIN_MARK}\n${FORCE_ABSTAIN_TEXT}`;
640
+ }
641
+
642
+ /**
643
+ * 计算投递结果(C3 + C2)。
644
+ * - hard 来源 flag(数字无源/假轮号/假工具/假卡/历史断言无引用/无证据硬答)→ 弃答文案;
645
+ * - B 题(canAbstain=true)且 evidenceUnits=0 → 弃答文案;
646
+ * - 元数据缺失(两者都没传)→ { mode:'not-applicable', forced:false, text:原文案 },不猜测;
647
+ * - 其余(含 soft flag)→ 原文案(软 flag 不阻断投递)。
648
+ *
649
+ * @param {{text?:string, canAbstain?:boolean|null, evidenceUnits?:number|null,
650
+ * evidenceTurns?:number[], evidenceText?:string, flags?:Array<object>}} [input]
651
+ * @returns {{mode:'abstained'|'answered'|'not-applicable', forced:boolean,
652
+ * reason:string, metaAvailable:boolean, text:string}}
653
+ */
654
+ export function resolveDelivery(input = {}) {
655
+ const original = String(input.text ?? '');
656
+ const hard = collectHardFlagCodes(input.flags);
657
+ const metaAvailable = input.canAbstain !== undefined && input.canAbstain !== null
658
+ || (Number.isInteger(input.evidenceUnits) && input.evidenceUnits >= 0);
659
+ const units = normalizeEvidenceUnits(input.evidenceUnits, input.evidenceTurns, input.evidenceText);
660
+ if (hard.length > 0) {
661
+ return {
662
+ mode: 'abstained',
663
+ forced: true,
664
+ reason: `HARD_SOURCE:${hard.join(',')}`,
665
+ metaAvailable,
666
+ text: ABSTAIN_DELIVERY_TEXT,
667
+ };
668
+ }
669
+ if (input.canAbstain === true && units === 0) {
670
+ return {
671
+ mode: 'abstained',
672
+ forced: true,
673
+ reason: 'CAN_ABSTAIN_NO_EVIDENCE',
674
+ metaAvailable,
675
+ text: ABSTAIN_DELIVERY_TEXT,
676
+ };
677
+ }
678
+ if (!metaAvailable) {
679
+ return { mode: 'not-applicable', forced: false, reason: 'META_UNAVAILABLE', metaAvailable: false, text: original };
680
+ }
681
+ return { mode: 'answered', forced: false, reason: 'OK', metaAvailable: true, text: original };
682
+ }
683
+
684
+ /**
685
+ * 从问题里抽检索关键词:引号内容权重最高,其次 CJK bigram 与英文词。
686
+ * @param {string} question
687
+ * @returns {Array<{term:string,weight:number}>}
688
+ */
689
+ export function extractKeywords(question) {
690
+ const raw = String(question ?? '');
691
+ const picked = new Map();
692
+ const add = (term, weight) => {
693
+ const t = String(term ?? '').trim().toLowerCase();
694
+ if (t.length < 2 || t.length > 40) return;
695
+ if (STOP_TERMS.has(t)) return;
696
+ if (/^\d+$/u.test(t)) return;
697
+ picked.set(t, Math.max(picked.get(t) ?? 0, weight));
698
+ };
699
+
700
+ const quoted = /[「“"`]([^」”"`\n]{1,40})[」”"`]/gu;
701
+ let m;
702
+ while ((m = quoted.exec(raw)) !== null) add(m[1], 4);
703
+
704
+ const cjkRuns = raw.match(/[\u4e00-\u9fff]{2,}/gu) ?? [];
705
+ const PARTICLE_RE = /[的了吗呢吧啊个这那是和与就都还又也我你他她它]/u;
706
+ for (const run of cjkRuns) {
707
+ const chars = [...run];
708
+ // 先整段(2–10 字)当候选,再补 bigram;过长整段会被 stopword 或重复压掉。
709
+ if (chars.length >= 2 && chars.length <= 10) add(run, 1.2);
710
+ for (let i = 0; i + 2 <= chars.length; i += 1) {
711
+ const bigram = chars[i] + chars[i + 1];
712
+ if (!STOP_TERMS.has(bigram) && !PARTICLE_RE.test(bigram)) add(bigram, 1);
713
+ }
714
+ // 三字以上专名(人名/工具名/文件名)额外加权。
715
+ for (let i = 0; i + 3 <= chars.length; i += 1) {
716
+ const tri = chars.slice(i, i + 3).join('');
717
+ if (!STOP_TERMS.has(tri) && !PARTICLE_RE.test(tri)) add(tri, 1.15);
718
+ }
719
+ }
720
+
721
+ const latin = raw.match(/[A-Za-z][A-Za-z0-9_./-]{2,40}/gu) ?? [];
722
+ for (const word of latin) {
723
+ if (/^(the|and|for|with|from|that|this|you|are|was|were|have|has|had|not|but|what|when|where|which|who|why|how)$/iu.test(word)) continue;
724
+ add(word, 1.6);
725
+ }
726
+
727
+ return [...picked.entries()]
728
+ .map(([term, weight]) => ({ term, weight }))
729
+ .sort((a, b) => b.weight - a.weight || b.term.length - a.term.length || a.term.localeCompare(b.term))
730
+ .slice(0, 28);
731
+ }
732
+
733
+ function countOccurrences(haystack, needle) {
734
+ if (needle === '' || haystack === '') return 0;
735
+ let count = 0;
736
+ let from = 0;
737
+ while (count < 3) {
738
+ const at = haystack.indexOf(needle, from);
739
+ if (at < 0) break;
740
+ count += 1;
741
+ from = at + needle.length;
742
+ }
743
+ return count;
744
+ }
745
+
746
+ /**
747
+ * 一轮历史的可检索文本。纯函数。
748
+ * @param {{turn:number,user?:string,assistant?:string,tools?:string}} turn
749
+ */
750
+ export function turnText(turn) {
751
+ return [turn?.user ?? '', turn?.assistant ?? '', turn?.tools ?? ''].join('\n').toLowerCase();
752
+ }
753
+
754
+ /**
755
+ * 给单个历史轮打分。
756
+ * @param {{turn:number,user?:string,assistant?:string,tools?:string}} turn
757
+ * @param {Array<{term:string,weight:number}>} keywords
758
+ * @returns {{score:number,hits:string[]}}
759
+ */
760
+ export function scoreTurn(turn, keywords) {
761
+ const text = turnText(turn);
762
+ const userText = String(turn?.user ?? '').toLowerCase();
763
+ if (text === '') return { score: 0, hits: [] };
764
+ const chars = text.length;
765
+ const lengthPenalty = Math.sqrt(1 + chars / 1200);
766
+ let score = 0;
767
+ const hits = [];
768
+ for (const { term, weight } of keywords) {
769
+ const tf = countOccurrences(text, term);
770
+ if (tf === 0) continue;
771
+ const inUser = userText.includes(term) ? 2 : 1;
772
+ score += weight * Math.min(tf, 3) * inUser;
773
+ hits.push(term);
774
+ }
775
+ return { score: score / lengthPenalty, hits };
776
+ }
777
+
778
+ /**
779
+ * 检索会话历史。
780
+ *
781
+ * @param {{turns:Array<{turn:number,user?:string,assistant?:string,tools?:string}>,
782
+ * question:string,currentTurn:number,maxItems?:number,minScore?:number}} input
783
+ * @returns {{zeroHit:boolean,items:Array<object>,keywords:Array<object>,queryTerms:string[]}}
784
+ */
785
+ export function retrieveEvidence({
786
+ turns = [],
787
+ question = '',
788
+ currentTurn = Number.POSITIVE_INFINITY,
789
+ maxItems = 3,
790
+ minScore = 1.8,
791
+ } = {}) {
792
+ const keywords = extractKeywords(question);
793
+ const limit = Math.max(1, Math.min(10, Number(maxItems) || 3));
794
+ const threshold = Number.isFinite(Number(minScore)) ? Number(minScore) : 1.8;
795
+ const refs = extractTurnRefs(question);
796
+ // 「刚才/刚刚」优先最近三轮;其余不受限。
797
+ const recentOnly = /刚才|刚刚|上一条|上一句/u.test(String(question ?? ''));
798
+
799
+ const candidates = [];
800
+ for (const turn of Array.isArray(turns) ? turns : []) {
801
+ const n = Number(turn?.turn);
802
+ if (!Number.isInteger(n) || n < 1) continue;
803
+ if (n >= Number(currentTurn)) continue;
804
+ if (recentOnly && n < Number(currentTurn) - 4) continue;
805
+ candidates.push(turn);
806
+ }
807
+
808
+ // 问题里点名了轮号:直接取该轮,不做关键词猜。
809
+ if (refs.length > 0) {
810
+ const named = refs
811
+ .map((ref) => candidates.find((turn) => Number(turn.turn) === ref))
812
+ .filter(Boolean)
813
+ .map((turn) => ({ ...turn, score: 99, hits: [`turn:${turn.turn}`], namedRef: true }));
814
+ if (named.length > 0) {
815
+ return {
816
+ zeroHit: false,
817
+ items: named.slice(0, limit),
818
+ keywords,
819
+ queryTerms: keywords.map((k) => k.term),
820
+ namedTurnRef: true,
821
+ };
822
+ }
823
+ }
824
+
825
+ // IDF:只按词频打分会让「问题/东西/之前」这类高频词压过专名;按在候选轮里
826
+ // 的文档频率降权后,稀有专名(备份、音程、turnmap)才会真正决定命中。
827
+ const df = new Map();
828
+ for (const { term } of keywords) {
829
+ let count = 0;
830
+ for (const turn of candidates) {
831
+ if (turnText(turn).includes(term)) count += 1;
832
+ }
833
+ df.set(term, count);
834
+ }
835
+ const total = Math.max(1, candidates.length);
836
+ const weightedKeywords = keywords.map(({ term, weight }) => ({
837
+ term,
838
+ weight: weight * (0.5 + Math.log(1 + total / (1 + (df.get(term) ?? 0)))),
839
+ }));
840
+
841
+ // 稀有词加分:只出现在少数轮里的词(df ≤ max(8, 10% 候选轮))单独再算一遍。
842
+ // 纯 IDF 仍会让「问题/之前/东西」这类词在长会话里累积到压过专名;稀有词
843
+ // 二次打分把专名(备份、turnmap、音程)重新顶上来。这是 2026-11-06 用
844
+ // 100 题真实历史问题校准出来的(见 eval/REPORT)。
845
+ const rareKeywords = weightedKeywords.filter(({ term }) => {
846
+ const count = df.get(term) ?? 0;
847
+ return count >= 1 && count <= Math.max(8, Math.ceil(total * 0.1));
848
+ });
849
+
850
+ const scored = [];
851
+ for (const turn of candidates) {
852
+ const { score, hits } = scoreTurn(turn, weightedKeywords);
853
+ const rare = rareKeywords.length > 0 ? scoreTurn(turn, rareKeywords).score : 0;
854
+ // 稀有词优先:历史检索要的是「哪一轮真的提过这个专名」,不是谁篇幅长。
855
+ // 没有稀有词(短问题/泛问)时才退回 IDF 词频。
856
+ const combined = rareKeywords.length >= 2 ? rare + score * 0.25 : score;
857
+ if (combined <= 0) continue;
858
+ scored.push({ ...turn, score: combined, hits });
859
+ }
860
+ scored.sort((a, b) => b.score - a.score || b.turn - a.turn);
861
+ const chosen = scored.filter((item) => item.score >= threshold).slice(0, limit);
862
+
863
+ return {
864
+ zeroHit: chosen.length === 0,
865
+ items: chosen.map((item) => ({ ...item, score: Number(item.score.toFixed(3)) })),
866
+ keywords,
867
+ queryTerms: keywords.map((k) => k.term),
868
+ };
869
+ }
870
+
871
+ const EXCERPT_MAX = 420;
872
+
873
+ function clip(text, max = EXCERPT_MAX) {
874
+ const raw = String(text ?? '').replace(/\r\n/g, '\n').trim();
875
+ if (raw.length <= max) return raw;
876
+ return `${raw.slice(0, max - 1)}…`;
877
+ }
878
+
879
+ /**
880
+ * 渲染证据包(隐藏注入;模型可见,聊天不露脸)。
881
+ * @param {object} result from retrieveEvidence
882
+ * @param {{question?:string,maxChars?:number,gate?:boolean}} [options]
883
+ * @returns {string}
884
+ */
885
+ export function renderEvidencePacket(result, { question = '', maxChars = 4000, gate = false } = {}) {
886
+ if (!result || !Array.isArray(result.items) || result.items.length === 0) return '';
887
+ const gateOn = gate === true;
888
+ const whitelist = result.items.map((item) => String(item.turn)).join(',');
889
+ const budget = Math.max(400, Math.min(4000, Number(maxChars) || 1200));
890
+ const lines = [
891
+ EVIDENCE_MARK,
892
+ `本轮问题需要回头查会话历史:${clip(question, 160)}`,
893
+ '检索到的真实轮次(只许引用下列轮号,引用格式「第 N 轮」):',
894
+ ];
895
+ // M1.5:gate 开启时先给机器可读白名单行(默认关不输出,保持逐字节兼容)。
896
+ if (gateOn) lines.push(`允许引用的轮号白名单:${whitelist}`);
897
+ for (const item of result.items) {
898
+ // R1:excerpt 窗口 300→800,避免 gold 落在 300 字符外时「注入了但看不到答案」。
899
+ const excerpt = clip(item.assistant || item.user || '', 800);
900
+ lines.push(`- 第 ${item.turn} 轮`);
901
+ if (item.user) lines.push(` 用户:${clip(item.user, 150)}`);
902
+ if (excerpt) lines.push(` 助手:${excerpt}`);
903
+ }
904
+ lines.push('引用规则:回答第一句先写「第 N 轮:」;证据里没有的内容不许补;证据不足就弃答。');
905
+ // M1.5:gate 开启时把硬规则写死在注入文本里;默认关时以下整段不出现。
906
+ if (gateOn) {
907
+ lines.push(
908
+ '【信心闸·硬规则】',
909
+ `1. 只能引用白名单内轮号:${whitelist};引用其他轮号=幻觉。`,
910
+ '2. 每个事实句必须带一个引用(「第 N 轮」)。',
911
+ '3. 任何数字/日期必须能在上面的证据摘录里逐字找到,找不到就不许写。',
912
+ '4. 不得声称做过未实际发生的工具调用;本轮工具结果才是证据。',
913
+ '5. 推断句子必须逐条标「【推断】」;未标注的推断=幻觉。',
914
+ '6. 证据已提供;请先引用再回答,不要回答「不知道」。',
915
+ );
916
+ }
917
+ let text = lines.join('\n');
918
+ if (text.length > budget) text = `${text.slice(0, Math.max(0, budget - 1))}…`;
919
+ return text;
920
+ }
921
+
922
+ /**
923
+ * 渲染「无证据」标记。
924
+ * @param {{question?:string,queryTerms?:string[],gate?:boolean}} input
925
+ * @returns {string}
926
+ */
927
+ export function renderNoEvidence({ question = '', queryTerms = [], gate = false } = {}) {
928
+ const terms = Array.isArray(queryTerms) && queryTerms.length > 0
929
+ ? queryTerms.slice(0, 8).join('、')
930
+ : '(无可用检索词)';
931
+ const lines = [
932
+ NO_EVIDENCE_MARK,
933
+ `本轮问题涉及会话历史,但检索层零命中:没有找到可核实的轮次(检索词:${terms})。`,
934
+ '规则:不许凭印象回答,也不许猜轮号。明确说「没找到相关记录」,或先问清是哪一轮/哪次对话。',
935
+ '弃答不扣分,编造轮号/数字/执行过程才算错。',
936
+ ];
937
+ // M1.5:gate 开启时从「提醒」升为「判定」;默认关时原数组逐字节不变。
938
+ if (gate === true) {
939
+ lines.splice(1, 0,
940
+ '【判定·必须弃答】本轮无证据:没有相关记录/无法核实。',
941
+ '禁止给出任何事实断言、数字、日期或轮号;禁止声称做过任何工具调用;不要猜测、不要补全。直接说「没有找到相关记录」,或问清是哪一轮/哪次对话。',
942
+ );
943
+ }
944
+ return lines.join('\n');
945
+ }
946
+
947
+ /** 系统提示不可用时的兜底规则消息;gate=true 时追加 M1.5 硬规则。 */
948
+ export function renderRuleBlock({ gate = false } = {}) {
949
+ return gate === true
950
+ ? [RULE_MARK, RULE_TEXT, GATE_RULE_TEXT].join('\n')
951
+ : [RULE_MARK, RULE_TEXT].join('\n');
952
+ }
953
+
954
+ /**
955
+ * 检测「声称执行了工具」。
956
+ * @param {string} text
957
+ * @param {string[]} knownTools
958
+ * @returns {Array<{kind:'tool'|'action',tool:string|null,phrase:string}>}
959
+ */
960
+ export function detectToolClaims(text, knownTools = []) {
961
+ const raw = stripCodeBlocks(text);
962
+ const known = new Set((Array.isArray(knownTools) ? knownTools : []).map((name) => String(name).toLowerCase()));
963
+ const claims = [];
964
+ for (const sentence of raw.split(SENTENCE_SPLIT_RE)) {
965
+ const line = sentence.trim();
966
+ if (line === '') continue;
967
+ if (NON_ASSERTION_RE.test(line)) continue;
968
+ TOOL_CLAIM_RE.lastIndex = 0;
969
+ let m;
970
+ while ((m = TOOL_CLAIM_RE.exec(line)) !== null) {
971
+ const tool = m[1].toLowerCase();
972
+ // 只认已知工具名,或明显的工具名形态(含下划线/x_ 前缀)。避免把普通英文单词当工具。
973
+ if (!known.has(tool) && !/^[a-z]+_[a-z0-9_]+$/u.test(tool)) continue;
974
+ claims.push({ kind: 'tool', tool, phrase: line.slice(0, 160) });
975
+ }
976
+ if (ACTION_CLAIM_RE.test(line) || ACTION_CLAIM_EN_RE.test(line)) {
977
+ ACTION_CLAIM_RE.lastIndex = 0;
978
+ ACTION_CLAIM_EN_RE.lastIndex = 0;
979
+ claims.push({ kind: 'action', tool: null, phrase: line.slice(0, 160) });
980
+ } else {
981
+ ACTION_CLAIM_RE.lastIndex = 0;
982
+ ACTION_CLAIM_EN_RE.lastIndex = 0;
983
+ }
984
+ }
985
+ return claims;
986
+ }
987
+
988
+ /**
989
+ * 生成后审计:只返回 flags,不改任何正文。
990
+ *
991
+ * @param {{
992
+ * text:string, turn:number, question?:string,
993
+ * historyQuestion?:boolean, zeroHit?:boolean, evidenceTurns?:number[],
994
+ * toolNames?:string[], sessionToolNames?:string[], sessionTurns?:number[],
995
+ * contextText?:string, knownTools?:string[],
996
+ * cardExists?:(ref:string)=>boolean,
997
+ * evidenceText?:string, gateCfg?:object, conflictCount?:number,
998
+ * }} input
999
+ * @returns {{verdict:'clean'|'flagged',flags:Array<object>,checks:object}}
1000
+ */
1001
+ export function auditTurn(input = {}) {
1002
+ const text = String(input.text ?? '');
1003
+ const turn = Number(input.turn);
1004
+ const question = String(input.question ?? '');
1005
+ const historyQuestion = input.historyQuestion === true;
1006
+ const zeroHit = input.zeroHit === true;
1007
+ const evidenceTurns = Array.isArray(input.evidenceTurns) ? input.evidenceTurns.map(Number).filter(Number.isInteger) : [];
1008
+ // FIX-05:evidenceUnits 主口径 = evidenceTurns 题级计数(live 恒可用,不依赖 meta)。
1009
+ // 显式传入仅作调用方覆盖(eval 映射没有该字段时绝不据此改 0)。
1010
+ const derivedEvidenceUnits = evidenceTurns.length;
1011
+ const effectiveEvidenceUnits = Number.isInteger(input.evidenceUnits) && input.evidenceUnits >= 0
1012
+ ? input.evidenceUnits
1013
+ : derivedEvidenceUnits;
1014
+ const toolNames = (Array.isArray(input.toolNames) ? input.toolNames : []).map(String);
1015
+ const sessionToolNames = (Array.isArray(input.sessionToolNames) ? input.sessionToolNames : []).map(String);
1016
+ const sessionTurns = (Array.isArray(input.sessionTurns) ? input.sessionTurns : []).map(Number).filter(Number.isInteger);
1017
+ const knownTools = Array.isArray(input.knownTools) ? input.knownTools : [];
1018
+ const cardExists = typeof input.cardExists === 'function' ? input.cardExists : () => true;
1019
+ const contextText = String(input.contextText ?? '');
1020
+ const contextTruncated = input.contextTruncated === true;
1021
+ // FIX-05:B 题标记 / 分层元数据 / 证据单元显式值(可选;缺省从 evidenceTurns/evidenceText 推导)。
1022
+ const canAbstain = input.canAbstain === true;
1023
+ const difficulty = String(input.difficulty ?? '');
1024
+ const questionType = String(input.questionType ?? '');
1025
+ const flags = [];
1026
+ const add = (code, detail, severity = 'fabrication', evidence = '') => {
1027
+ flags.push({ code, detail, severity, evidence: clip(evidence, 300) });
1028
+ };
1029
+ const checks = {
1030
+ abstained: isAbstention(text),
1031
+ askedClarification: asksClarification(text),
1032
+ turnRefs: extractTurnRefs(text),
1033
+ cardRefs: extractCardRefs(text),
1034
+ numbers: [],
1035
+ toolClaims: [],
1036
+ historyQuestion,
1037
+ zeroHit,
1038
+ evidenceTurns,
1039
+ // 2026-11-07: 取证上下文截断标记;评分器可对无出处数字走复核豁免。
1040
+ contextTruncated,
1041
+ };
1042
+
1043
+ // 信心闸(M1):纯函数三态判定;gate.enabled=false 时恒 answer(默认关=现状)。
1044
+ // fail-open:判定异常只置空、不改原 flags/decision;abstain 判定由下方 flags 兜底。
1045
+ let gate = null;
1046
+ try {
1047
+ gate = decideAnswerMode({
1048
+ question,
1049
+ draft: text,
1050
+ evidencePacket: String(input.evidenceText ?? ''),
1051
+ evidenceTurns,
1052
+ contextText,
1053
+ historyQuestion,
1054
+ zeroHit,
1055
+ // FIX-05 C3:B 题 + 0 证据单元时 gate 也判 abstain(与 answerMode 同口径)。
1056
+ canAbstain: input.canAbstain === true,
1057
+ evidenceUnits: effectiveEvidenceUnits,
1058
+ conflictCount: input.conflictCount,
1059
+ cfg: input.gateCfg,
1060
+ });
1061
+ } catch {
1062
+ gate = null;
1063
+ }
1064
+ checks.gate = gate;
1065
+ checks.citations = gate?.citations ?? extractTurnRefs(text).map((ref) => ({ turn: ref, readable: evidenceTurns.includes(ref) }));
1066
+ checks.assertions = gate?.assertions ?? { total: 0, cited: 0, inferred: 0, ungrounded: 0 };
1067
+
1068
+ // FIX-05:answerMode/evidenceUnits/sourceGate/tier 是纯输出字段;早退路径也必须带上。
1069
+ const applyAnswerMode = () => {
1070
+ Object.assign(checks, deriveAnswerModeFields({
1071
+ text,
1072
+ flags,
1073
+ gate,
1074
+ evidenceText: String(input.evidenceText ?? ''),
1075
+ evidenceTurns,
1076
+ evidenceUnits: effectiveEvidenceUnits,
1077
+ canAbstain,
1078
+ difficulty,
1079
+ questionType,
1080
+ }));
1081
+ };
1082
+
1083
+ // 弃答/反问:不要求轮号,也不做历史断言检查。
1084
+ if (text.trim() === '' || checks.abstained || checks.askedClarification) {
1085
+ applyAnswerMode();
1086
+ return { verdict: 'clean', flags, checks };
1087
+ }
1088
+
1089
+ // (a) 历史断言但无轮号引用。
1090
+ const historyClaim = historyQuestion || HISTORY_CLAIM_RE.test(text);
1091
+ checks.historyClaim = historyClaim;
1092
+ if (historyClaim && checks.turnRefs.length === 0) {
1093
+ add(
1094
+ 'HISTORY_CLAIM_NO_CITATION',
1095
+ '问题属于会话历史类,回答没有引用任何「第 N 轮」。',
1096
+ 'unsupported',
1097
+ text,
1098
+ );
1099
+ }
1100
+
1101
+ // 零命中仍然作答:明确违反无证据规则。
1102
+ if (zeroHit && !checks.abstained) {
1103
+ add(
1104
+ 'ANSWERED_WITHOUT_EVIDENCE',
1105
+ '本轮召回零命中并注入了「无证据」标记,回答仍未弃答。',
1106
+ 'unsupported',
1107
+ text,
1108
+ );
1109
+ }
1110
+
1111
+ // (d) 编轮号:引用的轮号必须存在,且不能超过当前轮。
1112
+ for (const ref of checks.turnRefs) {
1113
+ if (Number.isInteger(turn) && ref > turn) {
1114
+ add('FAKE_TURN_REF', `引用了第 ${ref} 轮,但当前才进行到第 ${turn} 轮。`, 'fabrication', `第 ${ref} 轮`);
1115
+ continue;
1116
+ }
1117
+ if (sessionTurns.length > 0 && !sessionTurns.includes(ref)) {
1118
+ add('FAKE_TURN_REF', `引用了第 ${ref} 轮,但该会话不存在这一轮。`, 'fabrication', `第 ${ref} 轮`);
1119
+ }
1120
+ }
1121
+
1122
+ // (d) 编卡/文件引用:路径必须真实存在(调用方给 exists;默认不拦)。
1123
+ for (const ref of checks.cardRefs) {
1124
+ let exists = true;
1125
+ try {
1126
+ exists = cardExists(ref) !== false;
1127
+ } catch {
1128
+ exists = true; // fail-open:存在性检查失败不冤枉模型
1129
+ }
1130
+ if (!exists) add('FAKE_CARD_REF', `引用了不存在的卡/文件:${ref}`, 'fabrication', ref);
1131
+ }
1132
+
1133
+ // (c) 编数字/日期:上下文与工具结果里都找不到。
1134
+ const corpus = numberCorpus([contextText, question].join('\n'));
1135
+ const flaggedValues = new Set();
1136
+ for (const claim of extractNumberClaims(text)) {
1137
+ const grounded = isGroundedNumber(claim, corpus);
1138
+ checks.numbers.push({ value: claim.value, kind: claim.kind, grounded, raw: claim.raw });
1139
+ if (!grounded && !flaggedValues.has(claim.value)) {
1140
+ flaggedValues.add(claim.value);
1141
+ add('UNGROUNDED_NUMBER', `数字/日期 ${JSON.stringify(claim.raw)} 在上下文与工具结果里都找不到。`, 'fabrication', claim.raw);
1142
+ }
1143
+ }
1144
+ checks.numbers = checks.numbers.slice(0, 20);
1145
+
1146
+ // (b) 假执行:被点名的工具本轮/本会话没有 tool call。
1147
+ const claims = detectToolClaims(text, knownTools);
1148
+ checks.toolClaims = claims.slice(0, 10);
1149
+ const turnSet = new Set(toolNames.map((name) => name.toLowerCase()));
1150
+ const sessionSet = new Set(sessionToolNames.map((name) => name.toLowerCase()));
1151
+ const seenClaim = new Set();
1152
+ let genericClaimed = false;
1153
+ let namedClaimFlagged = false;
1154
+ for (const claim of claims) {
1155
+ if (claim.kind === 'tool') {
1156
+ const tool = claim.tool.toLowerCase();
1157
+ const key = `tool:${tool}`;
1158
+ if (seenClaim.has(key)) continue;
1159
+ seenClaim.add(key);
1160
+ if (!turnSet.has(tool) && !sessionSet.has(tool)) {
1161
+ namedClaimFlagged = true;
1162
+ add('FAKE_TOOL_CLAIM', `声称用 ${claim.tool} 执行过操作,但本轮和该会话都没有对应 tool call。`, 'fabrication', claim.phrase);
1163
+ }
1164
+ } else {
1165
+ genericClaimed = true;
1166
+ }
1167
+ }
1168
+ if (genericClaimed && !namedClaimFlagged && turnSet.size === 0 && sessionSet.size === 0) {
1169
+ add('FAKE_TOOL_CLAIM', '声称已执行过工具操作,但本轮和该会话都没有任何 tool call。', 'fabrication', text);
1170
+ }
1171
+
1172
+ // 信心闸闭环:判定应弃答却输出事实断言 / 应降级却没有首句标注。
1173
+ // R3 评分安全:gate 违规只算 unsupported(评分器只把 fabrication 计幻觉),
1174
+ // 否则「本该弃答却答」会被评分污染成幻觉(评审模拟 12→31)。
1175
+ if (gate && gate.mode === 'abstain' && !checks.abstained) {
1176
+ add(
1177
+ 'GATE_ABSTAIN_VIOLATION',
1178
+ `信心闸判定应弃答(${gate.reasons.join(',')}),回答仍输出事实断言。`,
1179
+ 'unsupported',
1180
+ text,
1181
+ );
1182
+ }
1183
+ if (gate && gate.mode === 'degraded' && !GATE_DEGRADED_RE.test(text.trim())) {
1184
+ add(
1185
+ 'GATE_DEGRADED_MISSING',
1186
+ `信心闸判定应降级(${gate.reasons.join(',')}),回答首句缺少「不确定」标注。`,
1187
+ 'unsupported',
1188
+ text,
1189
+ );
1190
+ }
1191
+
1192
+ // FIX-05:所有 flag 产生后再定 answerMode(hard flag 必须能把答案改弃答)。
1193
+ applyAnswerMode();
1194
+
1195
+ return {
1196
+ verdict: flags.length > 0 ? 'flagged' : 'clean',
1197
+ flags,
1198
+ checks,
1199
+ };
1200
+ }
1201
+
1202
+ /** 把审计结果压成一行 JSONL;文本截断,避免审计文件无限膨胀。 */
1203
+ export function buildAuditRow({
1204
+ sessionId,
1205
+ turn,
1206
+ question,
1207
+ text,
1208
+ audit,
1209
+ historyQuestion,
1210
+ zeroHit,
1211
+ evidenceTurns,
1212
+ toolNames,
1213
+ mode = 'live',
1214
+ now = Date.now(),
1215
+ contextInfo = null,
1216
+ answerMode: answerModeOverride = null,
1217
+ evidenceUnits: evidenceUnitsOverride = null,
1218
+ sourceGate: sourceGateOverride = null,
1219
+ tier: tierOverride = null,
1220
+ }) {
1221
+ const contextTruncated = contextInfo?.truncated === true;
1222
+ // FIX-05:五个稳定字段。优先调用方显式传入(query/audit 路径补 fixture 元数据),
1223
+ // 其次取 auditTurn 写入的 checks.*,最后按旧字段回退(老行向后兼容)。
1224
+ const checks = audit?.checks ?? {};
1225
+ const answerMode = ANSWER_MODE_VALUES.includes(answerModeOverride)
1226
+ ? answerModeOverride
1227
+ : (ANSWER_MODE_VALUES.includes(checks.answerMode)
1228
+ ? checks.answerMode
1229
+ : (checks.abstained ? 'abstained' : 'answered'));
1230
+ // FIX-05:evidenceUnits 主口径 = evidenceTurns 计数(live 恒可用);显式覆盖优先,
1231
+ // 其次从 evidenceTurns 派生,最后退回 auditTurn 的 checks。绝不让缺 meta 变成 0。
1232
+ const derivedEvidenceUnits = Array.isArray(evidenceTurns) ? evidenceTurns.length : 0;
1233
+ const resolvedEvidenceUnits = Number.isInteger(evidenceUnitsOverride) && evidenceUnitsOverride >= 0
1234
+ ? evidenceUnitsOverride
1235
+ : (derivedEvidenceUnits > 0
1236
+ ? derivedEvidenceUnits
1237
+ : (Number.isInteger(checks.evidenceUnits) && checks.evidenceUnits >= 0 ? checks.evidenceUnits : 0));
1238
+ const resolvedSourceGate = sourceGateOverride && typeof sourceGateOverride === 'object'
1239
+ ? sourceGateOverride
1240
+ : (checks.sourceGate && typeof checks.sourceGate === 'object' ? checks.sourceGate : { hard: [], soft: [] });
1241
+ const resolvedTier = typeof tierOverride === 'string' && tierOverride !== ''
1242
+ ? tierOverride
1243
+ : (typeof checks.tier === 'string' && checks.tier !== '' ? checks.tier : 'unknown/unknown');
1244
+ return {
1245
+ v: ANSWER_GUARD_VERSION,
1246
+ at: new Date(now).toISOString(),
1247
+ mode,
1248
+ sessionId: String(sessionId ?? 'unknown'),
1249
+ turn: Number.isInteger(Number(turn)) ? Number(turn) : null,
1250
+ question: clip(question, 240),
1251
+ historyQuestion: historyQuestion === true,
1252
+ zeroHit: zeroHit === true,
1253
+ evidenceTurns: Array.isArray(evidenceTurns) ? evidenceTurns.slice(0, 10) : [],
1254
+ toolCalls: Array.isArray(toolNames) ? [...new Set(toolNames)].slice(0, 20) : [],
1255
+ verdict: audit?.verdict ?? 'clean',
1256
+ flags: Array.isArray(audit?.flags) ? audit.flags.slice(0, 12) : [],
1257
+ checks: audit?.checks ?? {},
1258
+ // 信心闸(M1)机器可读字段;旧行没有这些字段仍可正常解析(向后兼容)。
1259
+ gate: audit?.checks?.gate ?? null,
1260
+ citations: Array.isArray(audit?.checks?.citations) ? audit.checks.citations.slice(0, 10) : [],
1261
+ assertions: audit?.checks?.assertions ?? { total: 0, cited: 0, inferred: 0, ungrounded: 0 },
1262
+ gateVersion: GATE_VERSION,
1263
+ // FIX-05 稳定字段(1.8.0):answerMode 三态互斥 + 证据单元 + 双通道来源闸 + 分层。
1264
+ // answered/abstained/clarified 为互斥布尔,供评分器直接读,修 0461 同题状态冲突。
1265
+ answerMode,
1266
+ answered: answerMode === 'answered',
1267
+ abstained: answerMode === 'abstained',
1268
+ clarified: answerMode === 'clarified',
1269
+ evidenceUnits: resolvedEvidenceUnits,
1270
+ sourceGate: resolvedSourceGate,
1271
+ tier: resolvedTier,
1272
+ fix05Version: 1,
1273
+ // 2026-11-07 run C 修复:显式记录「审计上下文是否被截断」。
1274
+ // 评分器可据此对 UNGROUNDED_NUMBER 走「context-truncated」复核豁免,
1275
+ // 不再把「取证上下文被截断」误算成模型编数字(旧 run C:159→复核 10)。
1276
+ contextTruncated,
1277
+ context: contextInfo
1278
+ ? {
1279
+ chars: Math.max(0, Number(contextInfo.chars) || 0),
1280
+ keptChars: Math.max(0, Number(contextInfo.keptChars ?? contextInfo.chars) || 0),
1281
+ keyChars: Math.max(0, Number(contextInfo.keyChars) || 0),
1282
+ truncated: contextTruncated,
1283
+ }
1284
+ : null,
1285
+ assistantText: clip(text, 2000),
1286
+ };
1287
+ }
1288
+
1289
+ /** 审计上下文被截断时插在开头的显式标记(评分器/人眼都能看到)。 */
1290
+ export const AUDIT_CONTEXT_MARKER = '…[审计上下文已截断:前文省略,关键证据保留在末尾]…';
1291
+
1292
+ /**
1293
+ * 取文本数组的尾部,总长 ≤ maxChars;被截断时加显式标记。
1294
+ * 为什么取尾部:run C 的 40k 截断是「从前往后」,把后段历史/证据包切掉了,
1295
+ * 导致模型引用后段内容里的数字仍被判无出处(159 条误报)。审计取证应优先保留
1296
+ * 最近/最相关的文本,并显式标记截断,让评分器可以豁免。
1297
+ *
1298
+ * @param {string[]} parts 按时间顺序的文本块
1299
+ * @param {number} [maxChars]
1300
+ * @returns {{text:string, truncated:boolean, chars:number, keptChars:number}}
1301
+ */
1302
+ export function tailTextParts(parts, maxChars = 40000) {
1303
+ const list = (Array.isArray(parts) ? parts : []).map((part) => String(part ?? '')).filter((part) => part !== '');
1304
+ const max = Number.isFinite(maxChars) && maxChars > 0 ? Math.floor(maxChars) : 40000;
1305
+ const joined = list.join('\n');
1306
+ if (joined.length <= max) {
1307
+ return { text: joined, truncated: false, chars: joined.length, keptChars: joined.length };
1308
+ }
1309
+ const budget = Math.max(0, max - AUDIT_CONTEXT_MARKER.length - 1);
1310
+ const kept = [];
1311
+ let left = budget;
1312
+ for (let index = list.length - 1; index >= 0 && left > 0; index -= 1) {
1313
+ const part = list[index];
1314
+ if (part.length <= left) {
1315
+ kept.unshift(part);
1316
+ left -= part.length + 1;
1317
+ } else {
1318
+ kept.unshift(part.slice(part.length - left));
1319
+ left = 0;
1320
+ }
1321
+ }
1322
+ const text = `${AUDIT_CONTEXT_MARKER}\n${kept.join('\n')}`;
1323
+ return { text, truncated: true, chars: joined.length, keptChars: Math.max(0, text.length - AUDIT_CONTEXT_MARKER.length - 1) };
1324
+ }
1325
+
1326
+ /**
1327
+ * 组装审计取证上下文:关键证据(问题 / 注入证据包 / 工具结果 / 本轮先前正文)
1328
+ * 一律不截断;只有「泛上下文」按剩余预算取尾部。返回结果带截断标记,
1329
+ * 供 buildAuditRow 落盘、评分器复核豁免。
1330
+ *
1331
+ * @param {{question?:string, evidenceText?:string, toolResultText?:string,
1332
+ * priorText?:string, genericText?:string, maxChars?:number}} input
1333
+ * @returns {{text:string, truncated:boolean, keyChars:number, chars:number,
1334
+ * genericChars:number, genericKeptChars:number}}
1335
+ */
1336
+ export function composeAuditContext(input = {}) {
1337
+ const question = String(input.question ?? '').trim();
1338
+ const evidenceText = String(input.evidenceText ?? '').trim();
1339
+ const toolResultText = String(input.toolResultText ?? '').trim();
1340
+ const priorText = String(input.priorText ?? '').trim();
1341
+ const genericText = String(input.genericText ?? '');
1342
+ const maxChars = Number.isFinite(input.maxChars) && input.maxChars > 0 ? Math.floor(input.maxChars) : 40000;
1343
+ // 关键证据顺序固定:问题 → 注入证据包 → 工具结果 → 本轮先前正文。
1344
+ const keyParts = [question, evidenceText, toolResultText, priorText].filter((part) => part !== '');
1345
+ const keyText = keyParts.join('\n');
1346
+ const room = Math.max(0, maxChars - keyText.length - (keyText === '' ? 0 : 1));
1347
+ let genericKept = genericText;
1348
+ let truncated = false;
1349
+ if (genericText.length > room) {
1350
+ truncated = true;
1351
+ genericKept = room > 0 ? genericText.slice(genericText.length - room) : '';
1352
+ }
1353
+ const sections = [keyText];
1354
+ if (truncated) sections.push(AUDIT_CONTEXT_MARKER);
1355
+ if (genericKept !== '') sections.push(genericKept);
1356
+ const text = sections.filter((part) => part !== '').join('\n');
1357
+ return {
1358
+ text,
1359
+ truncated,
1360
+ keyChars: keyText.length,
1361
+ chars: keyText.length + genericText.length + (truncated ? AUDIT_CONTEXT_MARKER.length + 1 : 0) + (keyText === '' ? 0 : 1),
1362
+ genericChars: genericText.length,
1363
+ genericKeptChars: genericKept.length,
1364
+ };
1365
+ }
1366
+
1367
+ /** 从审计行里挑出「幻觉」类(编引用/编数字/编执行)。 */
1368
+ export function isFabricationRow(row) {
1369
+ return (Array.isArray(row?.flags) ? row.flags : []).some((flag) => flag?.severity === 'fabrication');
1370
+ }