HEL-860 handoff: deliverable/acceptance/deferred/verification JSON; audit860_postfix regenerated at final commit 6595153 (zero-embed); dual-target requirement checks pass vs origin/main and origin/agent/agent/de8874a3803b with handoff (4/4)

Co-authored-by: multica-agent <github@multica.ai>
This commit is contained in:
zonggong
2026-10-26 00:50:04 +08:00
co-authored by multica-agent
parent 6595153473
commit 22472e4b92
2 changed files with 51 additions and 1 deletions
+50
View File
@@ -0,0 +1,50 @@
{
"baseline_version": "1.11",
"read_commit": "9f3e6a33c7c03071abacec4a05141bebce977212",
"task": "HEL-860 让隐私检查守住问答出口并核准承诺状态(HEL-859 总管独立复核未通过的剩余纠错,零真实模型请求)",
"requirement_ids": [
"CG-008",
"CG-009",
"CG-011",
"CG-013",
"CG-014",
"CG-015",
"CG-016",
"CG-021",
"AM-001",
"UX-004",
"UX-006",
"GOV-001",
"GOV-002",
"GOV-003",
"GOV-004",
"GOV-005",
"GOV-006",
"DEL-001"
],
"deliverable": "总管独立复核《HEL-859-独立复核.md》四条新增反例的补正(原hel-857-mem0-spike分支,起点9f3e6a3,提交6595153):(1)全部问答出口收口到来源门——无门assemble_block删除,S2/S3/S4/S6/S7的Mem0侧最终装配一律经s7_gated_assembly逐条链校验,S6状态标注走同一门(annotator不开旁路),未结对照块剔除与状态矛盾的结果类条目保持自洽(剔除记status_conflict_excluded);真实A.run编排+假提取/假搜索/假问答捕获全部8个问答出口,mem0侧5个全部门内装配,unknown-origin毒条目修前进入S2实际发送上下文(门调用0)修后不出现在任何发送上下文。(2)锚点互证——turn_ids与turn_roles轮次清单矛盾拒anchor_turn_mismatch,命中文本逐字可识别为某已知原句而锚点指错行拒role_text_conflict(真实t1伴侣原句挂(t1,user)锚点t1 told=0、turn_ids=[t2]互矛均拒),同轮真实用户正例保留(row_verbatim入块),batch批次边界如实分层。(3)S6判据分别肯定验证——两侧请求成功AND已结侧肯定陈述承诺对象且肯定已告知AND未结侧肯定陈述且肯定尚未告知(预声明等价词集离线先行证明);'不知道'/两侧同答/已结侧答未结/未结侧同义宣称'结果已收到'均判失败;未结上下文剔除矛盾结果条目后自洽。(4)S2总判据以该次实际发送上下文含细节+请求成功+答案用上细节为准,答案碰词而上下文无细节判失败;三层分列、带门回读、gist子测保留。附:audit859.py修正(status_in_block→status_annotation_in_block假证据、退出码改为正负例全决、AUDIT859_REUSE_STORE复用store快照模式2次嵌入复跑全过)。",
"acceptance": [
"四条反例修前在9f3e6a3可复现:audit860_prefix.json(零嵌入)——毒条目进入S2实际发送上下文且s7_gated_assembly调用0(8出口仅1次门调用于S7)、错挂角色/互矛轮次resolved=true且伴侣原句入最终块、S6三组错误答案与S2碰词无细节均pass=true",
"修后统一真实出口全部拒绝/判失败:audit860_postfix.json@6595153(零嵌入)——8个问答出口全捕获、mem0侧5个全为带门装配(门调用6)、毒条目不在任何发送上下文;anchor_turn_mismatch/role_text_conflict拒绝且同轮用户正例row_verbatim入块;S6四组负例全判失败;S2碰词无细节判失败",
"完整离线套件一次通过:rework860_offline.json 26/26(R1夹具/重判、R2门禁2全套、R3R4编排与链+S6带门双块自洽与结果条目剔除实证、R6来源门收口+锚点互证负例+判据负例含碰词/请求失败、R5资源账+本卡嵌入预算≤70门、G0假传输),退出码0,零网络(socket硬拒绝+HF离线+本地缓存嵌入)",
"audit859修正后复跑通过:audit859_reuse860.json(AUDIT859_REUSE_STORE复用总管store快照,2次本地嵌入)——全部负例拒绝、正例保留(含原status_in_block误键导致的S6_correct_answers_pass=false假证据,现true)、退出码0且显式依赖all_negatives_rejected AND all_positives_kept",
"正例不回归:S7链正例恢复t1用户原句(row_verbatim)、同轮user正例入块、判据正确答案代入通过、既有R1-R4全部检查保持通过;run1/run2、rework858/859证据与原26行真实账本只读未动,新证据独立命名(audit860_prefix/postfix、rework860_offline、audit859_reuse860)",
"资源账逐笔可核验:唯一权威账hel-859-1f079e50f551/workdir/audit-data/embed_ledger.sqlite3续账390→458/500(套件66+audit859复用复跑2,另audit859定稿前预验2已含在458内),本卡新增嵌入68≤70、保留42≥40供13次复测计划;词面136→160另列;不补种不回退不双扣,旧作者账停止并行记账;真实出站0,26/60与余34冻结未动",
"check_requirements 对 origin/main 与 origin/agent/agent/de8874a3803b 双基准(含本handoff共四项)通过"
],
"deferred": [
"真实A/B复测仍冻结:本卡真实新增上限0未执行;13次复测方案(rework860_offline.json _retest_plan,5批提取+8次问答逐笔)待总管独立核对解除隐私停止门后在主卡'同意测试'授权与60次额度内执行,不重复索取相同范围授权;任一隐私负例失败继续冻结",
"复测前置核验:infer=True真实提取条目须随add携带来源锚点(ADD1-ADD5适配层责任),否则带门装配fail-closed拒绝、复测mem0侧如实判失败——已知边界如实列明,不伪装通过",
"正式采用Mem0或迁移产品接线:待白栖知拍板,本卡结果不构成采用决定",
"完整真实体验、八字情绪/生活连续/主动联系:继续按R1-R7承接,不以实验替代完整产品验收",
"中文提取提示词定制、真实语料验收、M5/M6产品接线:按原承接卡继续"
],
"requirement_changes": [],
"verification": [
"spike/hel857/results/audit860_prefix.json(修前四反例全复现@9f3e6a3)与 audit860_postfix.json(修后全拒/判失败、正例保留@6595153),同一脚本spike/hel857/audit860.py,零嵌入零网络(判据纯函数+假后端payload注入+总管store快照只读+真实A.run假外部操作全流程捕获8出口)",
"spike/hel857/results/rework860_offline.json(完整离线套件26/26+S6带门双块自洽实证+本卡嵌入预算门)",
"spike/hel857/results/audit859_reuse860.json(修正后audit859复跑:全负例拒绝/正例保留/退出码0,复用store快照2次嵌入)",
"spike/hel857/README.md《HEL-860 纠错修订记录》(四项旧/新对照表、资源账、已知边界:infer=True锚点传播复测前须核验)",
"本轮零网络出站、零真实模型(stub+本地fastembed缓存+假传输+socket硬拒绝);未验证现网/真实档案/真实模型(本卡范围禁止触及);真实复测与真实体验验收仍未运行,交总管独立核对"
]
}
+1 -1
View File
@@ -3,7 +3,7 @@
"real_model_requests": 0,
"local_embeddings": 0,
"phase": "postfix",
"commit": "9f3e6a3",
"commit": "6595153",
"criteria_negatives": {
"S6_resolved_answer_says_open": {
"layers": {