fix(organization): 解除结束探索对穷尽信息的隐含要求

- 根据具体线索的信息价值继续，结束不代表证明不存在

- 保持行为合同与预算，记录根因讨论和复测授权

Author: xiaoland

tasks/knowledge-lifecycle-capabilities/decisions/D551-D560.md

@@ -1,5 +1,14 @@
 # Decisions D-551–D-560
 
+### D-555 — 结束探索不承担排除遗漏的证明责任
+
+- **确认与授权**：Sir 同意将当前有依据的组织改进作为任务目标，而不是穷尽所有可能关系；要求实施并复测。
+  继续探索依据具体线索的预期信息价值；无有希望的下一步时可以结束，即使仍承认有未发现的信息。
+  结束不构成“相关信息不存在”的断言。
+- **实现范围**：只替换六种探索型 Agent 的共享目标/结束指导，移除被替代的重复句。保留语义合同、工具、
+  候选范围和预算；不要求输出逐步理由、固定检查或新增状态。Rumination 的独立三工具定义保持不变。
+- **验证**：复用既有 preview 初始世界，另存 discovery 轮；不新增测试，不改语义配置，不在运行中追加修复。
+
 ### D-554 — 恢复 rumination 能力组合，接受派生内容的 best-effort 残余
 
 - **接受**：Sir 接受部分派生内容不够准确是当前 best-effort 残余，本轮不为这类误差继续追加提示词修复。

tasks/knowledge-lifecycle-capabilities/decisions/index.md

@@ -77,7 +77,7 @@ so one stable ID has one predictable address；the shard boundary does not imply
 
 ## Current Edge
 
-- Latest confirmed decision by registered ID: [D-554](D551-D560.md)。MCP sink retains D-381–D-420；Telegram extension retains
+- Latest confirmed decision by registered ID: [D-555](D551-D560.md)。MCP sink retains D-381–D-420；Telegram extension retains
   D-421–D-460；Organization Nowledge study owns D-461–D-560。
 - MCP and Telegram are closed。The active Unit is
   [organization-nowledge-study](../units/organization-nowledge-study/packet.md) in Verify / Acceptance after implementation。

tasks/knowledge-lifecycle-capabilities/units/organization-nowledge-study/acceptance/preview-tool-repair.py

@@ -31,9 +31,11 @@
   "references",
   "guidance",
   "focal",
+  "discovery",
 ):
   raise ValueError(
-    "Choose baseline, repaired, prompt, batch, array, references, guidance or focal"
+    "Choose baseline, repaired, prompt, batch, array, references, guidance, "
+    "focal or discovery"
   )
 OUT = Path(__file__).with_name(f"tool-repair-{MODE}.json")
 RESUME = "--resume" in sys.argv
@@ -43,7 +45,7 @@
 DEFINITIONS_PATH = ROOT / "tests/organization/acceptance/agent_definitions.json"
 DEFINITIONS = (
   json.loads(DEFINITIONS_PATH.read_text())
-  if MODE in ("prompt", "batch", "array", "references", "guidance", "focal")
+  if MODE in ("prompt", "batch", "array", "references", "guidance", "focal", "discovery")
   else None
 )
 secret = subprocess.check_output(

tasks/knowledge-lifecycle-capabilities/units/organization-nowledge-study/acceptance/remaining-budget-diagnosis.md

@@ -5,6 +5,55 @@
 
 ## Sir 复核后的进一步定位
 
+### focal 轮 evidence stance / Job 84 根因复核
+
+**后续讨论：不完备信息下怎样结束（待评审）。** Sir 指出无法读完整图与担心遗漏之间的张力。
+当前方案不是让模型确信图已搜全，而是区分“本次结束探索”与“断言图中不存在相关证据”。no-op 只是不修改，
+不构成不存在性证明。因而继续探索应基于下一项获取动作的预期信息价值，而不是停止前必须消除一切遗漏可能。
+这是判断原则，不要求输出理由、固定步骤或额外报告，不暴露预算，也不限制 Agent 只能使用初始候选。
+
+初始 seed 足以支持适量探索；发现具体引用、可定位来源、相关命题或能够纠正已知检索错配的新路径，都可支持
+继续。负结果也可能使改换策略有价值，不要求每轮必须命中新 Block。仅仅“可能还有证据”、对同一空泛目标
+反复换词则不自动构成有价值的新线索。无具体有希望的下一步时可以结束，即使仍承认图中可能有遗漏。
+后续新信息或新线索可以重新触发组织；这不是本次必须解决全图的理由，也不新增永久 no-op 标记或状态机。
+
+拟将任务成功标准表达为“实现当前有依据的组织改进”，而不是“排除所有遗漏后给出结论”。这与已有允许 no-op
+不同：不止准许退出，还解除退出对充分检索的隐含证明责任。实际能否改善需复测，不能承诺换一句 prompt
+就消除模型行为问题。未实施。
+
+Sir 接受“有产出但继续工作而耗尽”可以有价值；本例没有任何写入，因此继续诊断，而非把预算终止一概当失败。
+原始证据为 tool-repair-focal.json。12 次模型返回的可见文本均为空，只有工具请求；不能把下述策略归因
+描述成模型亲口表达的想法，也不能声称已证明其内部停止条件。
+
+**已证实的输入条件。** Seed 305 是修订方案：批准状态、三种计划采用的机制，以及 rollout 以 replay passes
+为前提。它不是宣称修复已奏效或 replay 已通过的实验结果。邻接的 311 只是抽出的方案标题，304 是旧方案，
+这三者不能仅凭内容相关、版本先后或同源重述产生 supports/challenges。
+自动 seed 选择将 candidate、近期关系端点/Block、随机 Block 合并；本轮 305 进入时并非已验证的 stance pair。
+宽泛候选本身符合自动组织的职责，不应因此要求用户提供问题或让确定性筛选承担语义判断。
+
+**已证实的轨迹。** 调用 1/2 读取输入已有的 305 和邻域已完整返回的 311；调用 3/4 得到旧方案 304。
+调用 5 起持续搜 approval、replay test production、balancing/retry 等；调用 10 又搜索 approval decision、
+replay test passed。没有命中可用的批准记录或修复验收记录。调用 8 的 Nimbus 已扩大到事故材料；调用 11
+邻域才返回应用假说 301；调用 12 的 Reliability Lab replay load 为空，实验 302 正文没有 load。
+12 次检索中 6 次为空，其余多次返回已知方案；不存在工具错误或写入失败后的重试。
+
+**资料边界。** Fixture 的实验 302 复现了 routing rule 引发的故障模式，没有检验新的 remediation 机制，
+更没有报告它们通过了 rollout 验收。即便更早定位 302，也不能直接把它当成“方案有效/验证已通过”的证据。
+这不排除它对某个更窄命题有价值，也不意味着所有方案/决策类 Block 都不适合 evidence stance；但不能
+把条件“通过后才能 rollout”变成需要被证实已经发生的事实。
+
+**最有解释力的根因假设。** Agent 未把宽泛 seed 收敛成实际可比较的断言与证据问题，转而围绕方案寻找
+可能存在的证明材料；这些材料未出现时，继续换词寻找，而没有结束当前没有成立的比较。现有专用 SOP
+着重“判断一对完整 Block 的方向/scope/stance”与“已 settled pair 后停止”，共享 no-op 指导则很概括。
+它们允许退出，但在“尚未形成 pair”时没有在实际行为上阻止无依据的继续求证。故问题不是缺一条禁止循环
+的规则，而是目标形成与信息价值判断没有生效。此处是对轨迹和指令结构的推断，不是已隔离验证的单一因果。
+
+长词法查询和重复读取是可见放大因素；语义通道缺失进一步提高寻找材料的成本，却不是足以解释全部失败的
+根因。即使读得更快、召回更好，仍需承认没有合适关系可写；不能靠提高预算或放松 whole-assertion 语义掩盖。
+这也不是“有了很多产出还想多做一些”：整个执行没有 record_evidence_stance 或其它写入。
+
+本轮只诊断并更新 packet。未改工具、提示词、候选选择、预算或测试；后续方案仍须 Sir 复核。
+
 ### guidance 轮之后：能力组合与确认性复读复核
 
 D-554 已实施恢复专用 rumination definition，原先“待复核”提案的当前状态以本段为准。

tasks/knowledge-lifecycle-capabilities/units/organization-nowledge-study/packet.md

@@ -1,6 +1,16 @@
 # Organization Nowledge Study
 
 - **Unit ID**: `organization-nowledge-study`。
+- **Current implementation（D-555）**: Sir 已确认并授权实施“不以排除遗漏作为结束前提”的指导并复测。
+  已替换六种探索型 Agent 的共享目标/结束措辞，保留语义合同、工具与预算，rumination 不变。
+  准备提交部署并运行 discovery 轮，不新增测试。
+- **Current discussion**: Sir 提出无法穷尽图与担心遗漏的矛盾。待评审方向是区分“结束本次探索”与“证明不存在”，
+  以具体下一步的预期信息价值支持继续，而不以排除一切遗漏作为退出前提；不限定初始候选、不暴露预算、
+  不新增状态或报告。见 [诊断中的后续讨论](acceptance/remaining-budget-diagnosis.md)，尚未实施。
+- **Current diagnosis**: Sir 要求追查 focal 轮 evidence stance 无产出耗尽。已核对 seed 是修订方案而非已成立
+  的 stance pair；轨迹持续搜批准/验证材料，未将“没有形成可比较关系”收敛为结束。长词法查询与重读是
+  放大因素，不足以解释全部失败；目标形成/信息价值判断未生效是当前根因假设，未声称隔离因果已证明。
+  详见 [根因复核](acceptance/remaining-budget-diagnosis.md)，本轮未修改实现。
 - **Current verification**: `5fef0fd` 已完成 focal 轮验收及清理。Rumination 实际只绑定三工具，4/4/5 次
   均自然结束；整轮 5/7 Job、15/17 次执行自然结束，100 次模型/126 次工具调用，零工具错误。
   Evidence stance 检索耗尽、synthesis 最后一次写入后耗尽；写前重复读取仍存在。派生内容残余按 Sir

tests/organization/acceptance/agent_definitions.json

@@ -1,5 +1,5 @@
 {
-  "common_system_prompt": "You organize a neutral information base, not personal memory and not a user-facing conversation. Follow this definition's behavior and the semantic contracts of its tools. Initial seeds are starting points, not a boundary on exploration.\n\nBase judgments on sufficiently complete content already available; obtain more when information is missing. Full Blocks in the input or graph results are already available content; retrieval excerpts may be incomplete. Already available complete content needs no reread merely to prepare a write. Retrieval supplies candidates, not proof of relevance, identity or absence. Use get_entities for persisted records and Resolver for interpretation. Independent reads can share a model turn. Use graph queries for connections and paths; filter by actual or contract-defined relation contents, not guessed category names.\n\nOrganize exploration around a specific unresolved semantic question within this behavior's purpose. Use names, terms or relations from known material to find missing information, choosing retrieval or graph navigation as appropriate. Follow new evidence or promising candidates, including beyond the seed. Keep observation, testimony, hypothesis, inference and decision distinguishable, with their scope and attribution.\n\nWrite only when the specific behavior is justified. A write settles that candidate, not every nearby question: continue for a concrete useful candidate or necessary check, not an automatic new search cycle. When the available evidence supports no change, or an unresolved question has no promising next lead, end without proving global absence. Graph density is not a goal. For a concrete representation or prerequisite gap, cautiously mark a candidate for an appropriate registered behavior. A successful tool result confirms that operation. Do not reread returned Blocks, Relations, or neighborhoods merely to verify the write. Recording a candidate does not execute the behavior. Finish with a brief outcome, not a process report.",
+  "common_system_prompt": "You organize a neutral information base, not personal memory and not a user-facing conversation. Follow this definition's behavior and the semantic contracts of its tools. Initial seeds are starting points, not a boundary on exploration.\n\nBase judgments on sufficiently complete content already available; obtain more when information is missing. Full Blocks in the input or graph results are already available content; retrieval excerpts may be incomplete. Already available complete content needs no reread merely to prepare a write. Retrieval supplies candidates, not proof of relevance, identity or absence. Use get_entities for persisted records and Resolver for interpretation. Independent reads can share a model turn. Use graph queries for connections and paths; filter by actual or contract-defined relation contents, not guessed category names.\n\nYour goal is to realize currently justified organization improvements, not exhaust every possible relation. Use concrete leads to obtain information that can advance a judgment, including beyond the seed. Names, terms or relations in known material can guide retrieval or graph navigation. When no promising next step is apparent, finish even though undiscovered information may remain. Ending this exploration does not assert that relevant information is absent. Keep observation, testimony, hypothesis, inference and decision distinguishable, with their scope and attribution.\n\nWrite only when the specific behavior is justified. Graph density is not a goal. For a concrete representation or prerequisite gap, cautiously mark a candidate for an appropriate registered behavior. A successful tool result confirms that operation. Do not reread returned Blocks, Relations, or neighborhoods merely to verify the write. Recording a candidate does not execute the behavior. Finish with a brief outcome, not a process report.",
   "agents": {
     "rumination": {
       "system_prompt": "You organize a neutral information base, not personal memory and not a user-facing conversation. Reconsider the supplied focal Block and direct-relation context to create a useful, reusable distinction or representation; there is no prescribed transformation that every Block needs.\n\nIdentify what is hard to address, connect, understand or reuse in the focal Block. Preserve the difference between what a source observed, what its speaker believes, what an experiment reproduced and what you infer. A plausible explanation must not become an observed fact through rewriting or citations.\n\nDraft and submit the information and relations that realize the useful distinction. Keep extracted claims connected to their source and retain conditions, disagreement and uncertainty. Smaller or differently worded text is not automatically new information. Respect the meaning of any relation you author rather than using its label to imply an unestablished conclusion.\n\nOnly creating new Blocks requires the selected Resolver's draft input_schema; relation-only submissions do not. Pass draft arguments under draft_graph.input and keep temporary IDs disjoint when combining drafts.\n\nNo-op is appropriate when the supplied information offers no useful change. A successful tool result confirms the operation. Finish with a brief outcome.",