研究不是生成，是持续修正。Research is not generated. It is continuously revised.
一座会留下证据、保存分歧、承认修订并持续运行的研究机构，如何在一百天里长出来。How an institution that preserves evidence, keeps dissent, admits revision, and continues operating grew across one hundred days.
Evidence, dissent, revision, recap, and portfolio mapping now share one public timeline. This edition becomes a fixed historical snapshot while live research continues beyond Day 100.
DAY 100 / DECISION MEMO
2026-07-30 | 7月30日：AI瓶颈从方向判断进入兑现筛选，油价冲击尚未获得信用确认
2026-07-30: AI scarcity shifts from aggregate capex to powered conversion and return validation
The Day 100 filter shifts from broad AI capex growth to secured power, locked equipment slots, verifiable revenue, and lower crowding. Grid equipment ranks first, HBM stays selective, while unpowered projects, high-beta equipment, and capex without revenue anchors deserve a wider discount.
100 天不是为了证明 AI 能写很多报告，而是证明模型可以被组织成一座会留下证据、承认错误并持续运行的研究机构。
The point is not that AI can write many reports. It is that models can be organized into a research institution that preserves evidence, admits errors, and keeps operating.
01
模型不是机构
A model is not an institution
角色、节奏、记忆、协作、QA 和读者契约共同构成研究院。
Roles, cadence, memory, collaboration, QA, and a reader contract make the institution.
02
数量不是质量
Volume is not quality
真正的价值来自观点如何变化、为何变化，以及何时应被证伪。
Value comes from how views change, why they change, and when they should be falsified.
03
研究必须可读
Research must be readable
图结构材料只有进入人类判断顺序，才真正成为投资研究产品。
Graph-shaped evidence becomes useful only after entering a human decision sequence.
01B / THE FULL STORY
一座研究工厂，如何学会怀疑自己的绿灯
How a research factory learned to doubt its own green lights
This is not a smooth curve from manual work to autonomy. It is three rewrites of capability: establish institutional boundaries, govern repetition and contamination, then place market judgment, failure signals, and human review on one operating chain.
It began as one person's next-step memory, then became shared institutional state, a resource-allocation thesis, and a candidate-action ranker. The ranker proposes priorities; it does not gain authority to rewrite the system.
01MAY 17
个人看板Personal board PERSONAL BOARD
把操作子系统的下一步写成可见队列。
Make the operating subsystem's next actions visible.
02MAY 18
全院路线图Institute roadmap INSTITUTE MAP
跨角色、研究线和基础能力共享未决项。
Share open work across roles, research lines, and core capabilities.
03JUN 03
结构化状态Structured state JOINED STATE
路线图由单一文本转向文档状态与结构化快照的联接。
Move from one text file to joined document state and structured snapshots.
04JUN 11
资源分配原则Allocation thesis FUND → SPEND
“每件事只说一次”，把节省的容量投向更少、更深的问题。
Say each thing once, then fund fewer and deeper questions.
05JUN 13
候选动作排序Candidate ranking RANK, DON'T RULE
按偏离、纵深、陈旧度和积压排序；最终动作仍经操作角色与人类门禁。
Rank by drift, depth, staleness, and backlog; action still passes operator and human gates.
Day 100 路线图记录分类Day 100 roadmap record classification
这是从文档状态头得到的机制账本，不是功能数量、完成率或结果评分。
This is a mechanism ledger derived from document status headers, not a feature count, completion rate, or outcome score.
6 月 10 日不是 AI 基础设施主线退潮，而是研究密度从广谱 AI / 宏观 / A 股讨论收缩到更硬的物理约束：变压器、关键电网部件、并网容量、能源成本、核电交付时滞和重资产 GPU 云的资本回报。
READER MEMO2026-06-09
2026-06-09 AI Institute 每日读者报告
6 月 9 日的主线不是 AI 基础设施降温，而是研究注意力从泛 AI capex 进一步收敛到电网并网、变压器/开关设备、GOES、铜、电价、PPA 和公用事业资本开支；A 股风险从主题选择转向科创 50 低位、融资盘、情绪低位和防守修复的执行窗口。
READER MEMO2026-06-08
2026-06-08 AI Institute 每日读者报告
6 月 8 日的核心不是 AI 热度消失，而是 AI 资本开支的定价轴从模型和芯片继续下沉到工业金属、GOES、变压器、电价、绿证和并网约束；AI 基础设施正在被重新翻译成一条可通胀、可拥挤、也可被生产率抵消的物理供应链。
READER MEMO2026-06-05
2026-06-05 AI Institute 每日读者报告
6 月 5 日的主线不是 AI 交易结束，而是 AI 基础设施进入业绩、供电和风险预算三重检验期；科创 50、USD/JPY 160、美债 4.5% 和 AVGO 后续指引需要放在同一张事件窗口表里读。
READER MEMO2026-06-04
2026-06-04 AI Institute 每日读者报告
研究院进入事件窗口管理状态：风险偏好尚未崩，但 AI 基础设施、宏观通胀、人民币防线和 A 股 AI 硬件去杠杆已经被同一张图谱连到一起。
READER MEMO2026-06-03
2026-06-03 AI Institute 每日读者报告
AI 基础设施热度和宏观通胀传导风险同时位居前列，市场仍在奖励 AI 硬件与算力叙事，但研究图谱要求优先核对电网、设备交付、并网队列、HBM、信用再融资与利率敏感性。
AI CAPEX / MACRO2026-06-03
AI CapEx 通胀与去通胀拉锯复盘
能源、设备、利率先涨；生产率、自动化和单位任务成本等待验证
PRODUCTIVITY / DISINFLATION2026-06-03
AI生产率去通胀复盘
单位任务成本、采用率、边缘 AI、ASIC 与 Jevons 反证
AI POWER / SECOND ORDER2026-06-03
AI电力瓶颈二阶复盘
从“电不够”到并网、变压器、GOES、铜铝、开关柜、液冷与监管摩擦
A-SHARE / AI HARDWARE2026-06-03
A股算力硬件与科创50延伸复盘
从指数点位恐慌转向 CPO、存储、先进封装、国产算力和融资盘质量筛选
HBM / PACKAGING2026-06-03
HBM、存储与先进封装复盘
把 HBM、CoWoS、ABF、CPO 和测试排产放进 AI capex 兑现链
CHINA GRID / EXPORT2026-06-03
中国电力设备出海复盘
海外缺口、订单质量、本地化产能、原产地规则与 GOES/铜成本
AI INFRA2026-06-02
AI 基础设施瓶颈复盘
电力、CPO、存储、先进封装与电力设备
MAG72026-06-02
Mag7 下一时代输家复盘
入口、基础设施、现金流桥梁与物理供应链
STAR50 / KC502026-06-01
科创50研究观点与市场表现复盘
A-Share 衍生品与杠杆出清复盘
MSCI / MOC2026-05-30
MSCI 调仓复盘与口径统一
MOC 被动买盘、主动派发与数据披露陷阱
DEEP RESEARCH2026-05-24
下一代周期中，Mag7 谁会输？
AI Institute 语料指向的输家不是简单的股价下跌名单，而是三类失败：没有被本轮 AI 基础设施、能效、推理商业化和电力约束直接验证的战略相对输家；有清晰资本开支、折旧、SBC 与自由现金流压力的利润表输家；以及仍能增长但估值从稀缺性重定价为资本强度的倍数输家。若必须给出一个名字，苹果是最清晰的战略相对输家；若只看语料中直接证据，Meta 是最清晰的 P&L/FCF 压力对象；NVIDIA 更像倍数风险而非基本面失败。
DEEP RESEARCH2026-05-23
AI 生产率去通胀：采用率、工作流重构与滞后风险
AI 的去通胀叙事需要真实采用率、流程再设计和可计量产出改进共同兑现；在此之前，资本开支、能源与人才成本可能先表现为再通胀。
DEEP RESEARCH2026-05-23
变压器、GOES 与铜：AI 电力硬件瓶颈的价格传导
AI 电力交易的核心不只是发电量，而是高压变压器、取向硅钢、铜铝与开关设备的交付约束；这些约束决定设备商利润率、客户交期与终端算力释放速度。
DEEP RESEARCH2026-05-19
消费降级、银发经济与大健康结构性机会
消费链条正在从总量弹性转向结构分化：银发消费、医疗健康、糖替代和 B 端餐饮渗透可能形成防御性需求，但需要验证支付能力和渠道真实动销。
DEEP RESEARCH2026-05-18
AI 算力资本开支、电力瓶颈与通胀再定价
AI 基础设施需求正在先通过电力、电网、变压器与材料供给形成成本脉冲，再通过生产率与架构效率形成中期缓释；投资结论取决于这两个方向谁先兑现。
DEEP RESEARCH2026-05-18
中国电力设备出海：AI 电网需求、贸易壁垒与利润率分化
全球 AI 电网投资为中国电力设备提供需求窗口，但关税、认证、交付能力和原材料成本会决定哪些企业真正把订单转化为利润。
06 / INSTITUTION LOOP
模型如何变成研究同事
How models become research colleagues
这不是一条单向生产线。复盘和证伪会回到下一轮信号选择，形成可观察的机构记忆。
This is not a one-way pipeline. Recap and falsification return to signal selection as observable institutional memory.
01 / SIGNAL信号进入研究议程Signals enter the research agenda
市场变化、研究缺口和既有 thesis 的证伪条件决定下一步由谁研究，而不是让模型随机生成内容。
Market changes, evidence gaps, and thesis falsifiers determine what is researched next.
07 / OPERATING PROOF
一座研究机构，不只要会研究，还要知道自己何时没有产生价值
A research institution must know not only how to research, but when it produced no value.
来自研究机构、执行基础设施与公开阅读层的跨系统证据对账。机制、运行指标与结果验证被明确分开。
A cross-system evidence review spanning the research institution, execution substrate, and public reader layer. Mechanisms, operating metrics, and demonstrated outcomes are kept separate.
机制可验证Mechanism verified指标为自报Metric self-reported结果待积累Outcome still early故障已复盘Incident reviewed
WHAT / WHENAI INSTITUTE
决定研究什么、何时运行，并提交执行意图。
Decides what to research, when to run, and submits execution intent.
→
HOWAGENT ROUTE
负责排队、路由、执行、产物与恢复。
Owns queues, routing, execution, artifacts, and recovery.
→
FOR WHOMVIBE
把脱敏研究重组为人类可检查的判断。
Recomposes sanitized research into inspectable human judgment.
01 / RESEARCH INSTITUTION
AI Institute：把模型组织成会修正的研究机构
AI Institute: organizing models into a self-correcting research institution
负责研究议程、专业分工、证据与反证、质量复核、论点结算和机构记忆。
Owns the research agenda, specialist roles, evidence and counterevidence, quality review, thesis settlement, and institutional memory.
知识库从事实仓库变成带信任门的研究消费者The knowledge base becomes a trust-gated research consumer机制可验证Mechanism verified+
An extraction-integrity audit found that shared outputs could contaminate the fact table; suspect facts were quarantined and consumption paths were restricted to trusted, current records.
诚实边界Honest limit
历史污染被隔离而非删除，KB 驱动研究仍主要是机制证明；当时深度研究跟踪到的事实数量仍为 0。
Historical contamination was quarantined rather than deleted, and KB-directed research remained mostly mechanism proof; tracked deep-research facts were still zero at the review point.
如何推翻What would refute it
审计任一知识读取路径；若可疑或已被取代的事实仍能进入研究上下文，信任门并未成立。
Audit every knowledge read path. Any suspect or superseded fact entering research context refutes the trust gate.
Matured forecasts can become hit-rate, Brier, and calibration records; research quality moves from a binary flag to scored dimensions with before-and-after revision deltas.
诚实边界Honest limit
机制已经闭合，但首批可结算样本极少；“有反馈回路”不等于“预测已经变准”。
The loops exist, but the first settled sample was tiny. A feedback mechanism is not evidence that forecasts have improved.
如何推翻What would refute it
若结算样本长期不增长，或复核后评分没有被记录和反馈，闭环只是界面结构。
If settled samples do not grow, or post-revision scores are not recorded and fed back, the loop is decorative.
Thesis 从形容词变成可结算的赔率合约Theses move from adjectives to settleable odds contracts结果待积累Outcome still early+
可证明What is supported
动态论点和赛道被映射到多期限、带失效条件的影子赔率，迫使观点同时表达方向、时间、阈值和证伪条件。
Living theses and lanes map into multi-horizon shadow odds with invalidation conditions, forcing views to express direction, time, thresholds, and falsifiers.
系统开始监测“没有产出”而不只监测“仍在运行”The institution starts monitoring missing value, not only liveness故障已复盘Incident reviewed+
可证明What is supported
空白结果与知识事实骤降曾在任务显示成功时悄然发生；随后加入空输出守卫、产出脉搏与有界补偿。
Blank results and collapsing knowledge yield once passed through apparently green task states; empty-output guards, value-accrual pulses, and bounded recovery followed.
诚实边界Honest limit
输出守卫只能识别空白或占位符，不能识别“有文字但结论错误”；工作区同步问题当时仍依赖绕行方案。
Shape guards catch blanks and placeholders, not fluent but wrong output; the workspace-sync failure still relied on a workaround.
如何推翻What would refute it
若成功记录中再次出现空白占位输出，或事实产出归零却没有触发异常，这一防线仍未闭合。
A completed blank placeholder, or zero knowledge yield without an alert, would refute the guardrail.
Tasks reported completion without files reaching their session workspace. The research layer contained the failure with an echo fallback and fact-yield pulse; the execution layer confirmed that the root fix and completed-task workspace-yield probe remain outstanding.
The roughly 270 green daily tasks and six-day zero-fact interval are point-in-time internal counts; restored yield through a workaround does not prove the transport contract is repaired.
如何推翻What would refute it
派发一个必须写文件的测试任务并读取其工作区；若任务显示完成而文件仍为空，跨层缺口仍然存在。
Dispatch a task required to write a file, then read its session workspace. A completed task with an empty workspace confirms the gap remains open.
A roughly six-week blank-output incident led to boundary guards: empty placeholders now retry within bounds and fail explicitly, followed by daily hand-by-node failure statistics.
The seven-day zero-sentinel close condition was not yet met; some edge hands still had a 30-minute ceiling, so two-hour end-to-end support was not true.
如何推翻What would refute it
任何新的空白占位任务仍被记录为成功，或超时继续被折叠为普通失败，都会推翻修复与可见性声明。
Any new blank placeholder recorded as success, or timeout folded back into generic failure, refutes the fix and visibility claim.
人类可读报告成为独立成果而不是摘要附件The reader report becomes a first-class output机制可验证Mechanism verified+
可证明What is supported
每日综合按一句话结论、优先判断、变化、证据与反证、投资映射、监测和证伪重新编排，并保留来源账本。
Daily synthesis is reordered into a conclusion, priorities, changes, evidence and challenges, investment mapping, monitoring, falsifiers, and a source ledger.
诚实边界Honest limit
可读性不是正确性的替代品；模型综合必须被允许承认报价为空、口径冲突和无法验证的市场表现。
Readability is not correctness; synthesis must disclose empty quotes, conflicting counts, and unverifiable market performance.
如何推翻What would refute it
若读者仍必须打开多条内部来源才能知道结论、风险与证伪条件，重组工作没有完成。
If a reader still needs multiple internal sources to learn the conclusion, risks, and falsifiers, recomposition is incomplete.
How a blank report passed through three green states
这次事故最能说明为什么 100 天的核心不是生成量，而是系统能否发现自己没有产生价值。
This incident best explains why the 100-day claim is not output volume, but whether the system can detect that it produced no value.
01边缘执行器返回 0The edge hand returns zero
地域限制导致零字节输出，但进程仍以成功状态退出。
A regional restriction produces zero-byte output while the process exits successfully.
02运行层记录完成Runtime records completion
退出状态被机械映射为 completed，空白被包装成占位文本。
Exit state maps mechanically to completed and the blank becomes placeholder text.
03研究层继续消费Research keeps consuming it
卡片和研究链有记录，因此传统存活监测仍然显示绿色。
Cards and chains exist, so conventional liveness monitoring remains green.
04边界守卫被加入A boundary guard is added
空白与纯占位输出改走有限重试，并在失败后显式标记。
Empty placeholders enter bounded retry and fail explicitly when recovery is exhausted.
05从存活转向价值脉搏From liveness to value pulse
系统开始监测产出是否累积；但“有文字却错误”仍需要证据与 QA。
The system monitors whether value accrues, while fluent but wrong output still requires evidence and QA.
OPERATING HANDBOOK / ADOPT + REJECT + TEST
真正经受过运行压力的做法
Practices that survived operating pressure
以下不是架构口号，而是从污染、空白输出、状态分裂、重复研究和不可结算观点中留下的最小操作原则。
These are not architecture slogans. They are the smallest operating practices left by contamination, blank output, split state, repeated research, and un-settleable views.
01RESEARCH CAPACITY
先测语义冗余，再增加角色与算力
Measure semantic redundancy before adding roles or compute
覆盖度不是人数和报告数。先在真实样本上测“观点是否重复”，再把容量从横向复述转向垂直纵深。
Coverage is not headcount or report count. Measure repeated meaning on real artifacts, then redirect capacity from horizontal repetition toward vertical depth.
REJECT
拒绝：把更多输出自动解释为更多研究能力。
Reject: treating more output as more research capability.
TEST
验证：变更后必须复测冗余；没有复测，就只能证明组织变化，不能证明价值提升。
Test: rerun the redundancy audit after the change; without it, only organizational change is proven.
02TRACK RECORD
从可结算的起点向前积累，不补造历史
Accrue forward from a settleable inception; do not manufacture history
Forecasts, odds, and portfolios need frozen inception, benchmark, horizon, and invalidation rules; an honestly sparse forward sample is better than a backfilled curve with selection bias.
REJECT
拒绝：对旧观点追溯打分并把它包装成真实业绩。
Reject: retro-scoring old views and presenting the result as a live track record.
TEST
验证：结算样本应持续增长，方法和起始日固定，结果可按时间顺序重算。
Test: settlements grow, methodology and inception remain fixed, and outcomes can be replayed chronologically.
03PROVENANCE
证据来源由管道注入，模型只能解释不能发明
The pipeline injects provenance; the model may interpret but not invent it
事实需要来源标识和短证据摘录；缺少来源的抽取应被明确拒绝，而不是用流畅文字补齐。
Facts require a source identifier and a short evidence quote; an extraction without provenance should be rejected rather than completed with fluent prose.
REJECT
拒绝：把引用完整性留给模型自觉。
Reject: leaving citation integrity to model discretion.
TEST
验证：缺少来源或证据摘录的记录不能进入可信知识层。
Test: records without source identity or an evidence quote cannot enter the trusted knowledge layer.
04VALUE PULSE
监测价值是否累积，而不只监测任务是否存活
Monitor whether value accrues, not only whether tasks stay alive
Completion state, queue depth, and error rate can all be green while facts, prose, or reader value remain zero. Every critical pipeline needs a value-bearing measure.
REJECT
拒绝：把 exit 0、completed 或心跳正常等同于研究成功。
Reject: equating exit zero, completed, or a healthy heartbeat with research success.
TEST
验证：注入空白输出或零产出场景，系统必须在一个监测周期内转红。
Test: inject a blank or zero-yield condition; the system must turn red within one monitoring interval.
05FAILURE POLICY
按原因处理失败：有界重试、快速失败、独立时钟
Treat failures by cause: bounded retry, fail-fast, separate clocks
拥塞、永久错误和单次查询成本超限不是同一种失败；排队时间、执行时间与回收窗口也不能共用一个时钟。
Congestion, permanent errors, and per-query cost failures need different policies; queue time, execution time, and reap windows also need separate clocks.
REJECT
拒绝：所有错误统一重试，所有执行器统一超时。
Reject: retrying every error and applying one timeout to every hand.
TEST
验证：暂时错误有限重试，永久或成本错误立即停止，长任务不会因排队时间被提前回收。
Test: transients retry within bounds, permanent or cost failures stop immediately, and queued long work is not reaped early.
06RECOVERY
推送是加速器，对账和幂等才是正确性底座
Push is an accelerator; reconciliation and idempotency carry correctness
结果推送必须可重复、可延迟、可丢失；轮询合约、条件更新和分裂状态对账让系统最终收敛。
Result delivery must tolerate duplicates, delay, and loss; a polling contract, guarded updates, and split-state reconciliation let the system converge.
REJECT
拒绝：只依赖回调，或把恢复机制当成根因已经修复。
Reject: callback-only correctness or treating recovery as root-cause repair.
TEST
验证：重复、延迟或遗漏一次回传后，任务和产物状态仍能在宽限期后收敛。
Test: after a duplicated, delayed, or lost delivery, task and artifact state still converge after the grace window.
07VERIFY EDGE
每个闭环都要命名验证边，并区分机制与结果
Every loop names a verify edge and separates mechanism from outcome
文档中的箭头和代码中的挂载都不代表闭环有效。验证边必须有吞吐、有时间序列，并允许明确写下尚未证明。
An arrow in a document or a mounted hook does not make a working loop. Its verify edge needs throughput, a time series, and an explicit unproven state.
REJECT
拒绝：用“已接通”代替“已改善”。
Reject: substituting wired for improved.
TEST
验证：样本长期不增长或结果趋势不改善时，页面必须降级为机制证明，而不是继续宣称飞轮。
Test: if samples stall or outcomes do not improve, the claim must downgrade to mechanism proof rather than remain a flywheel claim.
08HUMAN GATE
人工保留异常判断和不可逆决策，机器承担持续执行
Humans retain anomaly judgment and irreversible decisions; machines sustain execution
Human counter-observations repeatedly corrected model consensus, redirected root causes, and set role, permission, and risk boundaries; gates should be earned with evidence, not removed by default.
REJECT
拒绝：把零人工介入当作成熟度，也拒绝为了协调而增加持续心跳。
Reject: treating zero human intervention as maturity, or adding standing heartbeats merely to feel coordinated.
TEST
验证：不可逆动作有明确人工确认；例行流程只有在连续、可复现的验证后才减少门禁。
Test: irreversible actions require explicit human approval, while recurring paths lose gates only after repeatable verification.
HUMAN ROLE / OUTSIDE-SYSTEM SENSOR
人类不是轮询监控员，而是系统之外的异常传感器
The human is not a polling monitor, but an anomaly sensor outside the system
The most consequential human work was not writing code for the models, but noticing counterexamples consensus missed: why a latest fact returned an old version, why blank output was not an authentication issue, and what reversibility a fleet rebalance required. Machines can sustain a given abstraction; humans remain strongest at noticing that the abstraction itself is wrong.
这也是商业可持续性的边界：研究院要减少人工轮询，但不能伪装成不需要人类判断。 This is also the business-sustainability boundary: the institution should reduce human polling without pretending it no longer needs human judgment.
DAY 101 / MEASUREMENT CONTRACT
下一阶段不再问“还可以加什么”，而问“什么结果会推翻我们”
The next phase asks not what else to add, but what outcome would prove us wrong.
01
把“已接通”的闭环变成可测量结果
Turn wired loops into measured outcomes
MEASURE
预测结算样本跨分析师持续增长、Brier 形成时间趋势，KB 需求研究出现真实解决记录。
Settled forecasts grow across analysts, Brier forms a time trend, and KB-need research records real resolutions.
REFUTE
30 天后样本仍接近个位数、校准无趋势、KB 解决数仍为 0，则闭环仍是装饰。
If samples remain near single digits after 30 days, calibration has no trend, and KB resolutions remain zero, the loops are decorative.
02
关闭端到端价值完整性缺口
Close the end-to-end value-integrity gap
MEASURE
空白占位成功为 0 并连续保持 7 天；模拟零产出或节点冻结能在一个监测周期内触发异常。
Blank-placeholder successes remain zero for seven days, and simulated zero-yield or frozen-execution conditions trip within one monitoring interval.
REFUTE
强制故障后状态仍为绿色，或空白正文仍能进入研究与发布层，则只完成了局部 containment。
If a forced fault stays green or blank prose still reaches research and publishing, only local containment has been achieved.
03
建立带基准、时间戳和读者价值的决策记录
Build a benchmarked, timestamped record of decision value
MEASURE
组合与赔率从固定起始日逐日结算并对照基准；人类可读交付的时延、引用覆盖和修订原因形成稳定口径。
Portfolios and odds settle daily from a fixed inception against benchmarks, while reader-delivery latency, citation coverage, and revision reasons gain stable definitions.
REFUTE
若业绩仍只有无法复算的时点数字，或读者仍需打开内部链路才能理解结论与证伪条件，决策产品尚未成立。
If performance remains an unreplayable point estimate, or readers still need internal chains to find conclusions and falsifiers, the decision product is not established.
WHAT 100 DAYS DO NOT PROVE
把不能证明的部分也放进品牌承诺
Put what cannot be proven inside the brand promise.
01尚不能证明预测质量已经提高Forecast quality improvement is not yet proven
结算、校准和反馈回路已经存在，但样本仍短；应以持续增长的结算样本和下降的校准误差验证。
Settlement, calibration, and feedback exist, but the sample is short; growing settlements and declining calibration error must carry the claim.
02尚不能把产出规模等同于研究质量Output scale is not research quality
报告、关系和页面数量只能证明覆盖；引用质量、修订率、证伪率与遗漏召回才是下一阶段指标。
Reports, relations, and pages prove coverage only; citation quality, revision rate, falsification rate, and omission recall are the next measures.
03尚不能宣称组合产生了可验证超额收益Verified portfolio alpha is not yet demonstrated
公开组合和 odds 仍处于早期或 shadow 阶段，缺少足够长、带基准、可结算的完整样本。
Public portfolios and odds remain early or shadow-only, without a sufficiently long, benchmarked, settleable sample.
04分布式基础设施仍有边缘缺口The distributed substrate still has edge gaps
Node freezes, workspace sync, edge time limits, and database capacity can still affect research completeness; the workspace workaround restored yield, but the root fix and completed-task yield probe remain open.
08 / PUBLIC SCORECARD
规模、结构、可读性与复盘
Scale, structure, readability, and recap
以下只展示当前公开数据能够证明的指标。尚未形成可靠口径的质量指标不会被伪造成完成率。
Only metrics supported by current public contracts are shown. Unmeasured quality is not presented as a fabricated completion rate.