The whole thing, briefly
整体,简述
An organization buys a capability. Months later, someone asks whether it worked. The honest answer is usually unavailable — not because nobody looked, but because nobody agreed in advance what "worked" would mean, or where the evidence would come from.
一个组织买下一项能力。几个月后,有人问它有没有用。诚实的回答通常给不出来——不是因为没人去看,而是因为没有人在事前约定「有用」是什么意思,证据又该从哪里来。
DI-Std treats that as a measurement problem with a specific shape. Capability arrives installed. It becomes useful only when it is realized — actually reaching real decisions, on time, at quality, reliably. The distance between those two states is not mysterious. It is structural lag plus execution friction, and both are observable if you instrument the right thing.
DI-Std 把这件事当作一个有特定形状的测量问题。能力到达时是已安装的。它只有在已实现时才开始有用——真正抵达真实的决策,准时、达标、稳定。两种状态之间的距离并不神秘:它是结构滞后加上执行摩擦,而只要仪器架在对的地方,两者都可观测。
Version 2.0 adds the dimension 1.0 lacked: time. It is not enough to know how much of an installed capability has been realized. You need to know how fast that share is changing, whether it has stopped changing, and — when it has — exactly where in the decision path the conversion is being lost.
2.0 补上了 1.0 缺的那个维度:时间。只知道已安装能力中有多少被实现,是不够的。你需要知道这个比例变化得多快、有没有停止变化,以及——如果停了——转化究竟在决策路径的哪一步被消耗掉。
This guide explains the concepts. It is not the specification text, it states no conformance requirements, and it is not an implementation manual. Examples throughout follow one worked case — Meridian Distribution — which is illustrative, constructed to show what the constructs mean rather than how to operate them.
本指南解释概念。它不是规范正文,不设任何合规要求,也不是实施手册。全文贯穿一个示例案例——Meridian Distribution——为示意性质,用于说明各构念的含义,而非操作方式。
Installed is not realized
已安装不等于已实现
An organization licenses a replenishment model. It passes acceptance testing. It is live, callable, and on the architecture diagram. By every procurement measure, the capability exists.
一个组织采购了一套补货模型。它通过了验收测试。它上线了、可调用、出现在架构图上。按采购口径的每一项衡量,这项能力都存在。
Six months later, buyers are still rebuilding its output by hand before anyone acts on it, half the categories it was scoped to cover are not routed through it, and the recommendation arrives after the ordering window twice a week. The capability exists. It is not doing anything.
六个月后,采购员仍然要在有人行动之前手工重建它的输出;它设计覆盖的品类有一半没有走它;建议每周有两次在下单窗口关闭之后才到。能力存在。它没有在做任何事。
That distance has a name and an arithmetic:
这段距离有名字,也有算式:
This is an accounting identity, not a predictive model. It says only that the difference between what you bought and what you got decomposes into exactly two things, and that the two behave differently:
这是一个会计恒等式,不是预测模型。它只说明一件事:你买到的和你得到的之间的差额,恰好分解为两样东西,而这两样的行为方式不同。
Capability the structure has already given up while standing still. Ownership that was never assigned. A decision right that sits three levels above where the consequence lands. Scope that was designed out.
结构在静止状态下就已经放弃掉的能力。从未被指派的归属。停在后果落点之上三层的决策权。在设计阶段就被砍掉的范围。
Fixed by redesign. Running faster does not touch it.
靠重设计解决。跑得更快碰不到它。
Capability lost in motion. Queue time waiting for a weekly cadence. Duplicate verification. Re-authorization of something already authorized. Manual translation between systems.
在运动中损失掉的能力。等待周度节奏的排队时间。重复核验。对已经授权过的事情再授权一次。系统之间的人工搬运。
Fixed by clearing the path. Redesigning the structure does not touch it.
靠打通路径解决。重设计结构碰不到它。
Confusing the two is not a harmless imprecision — it sends the entire remediation effort to the wrong place. An organization that responds to structural lag with process acceleration gets faster at doing the wrong thing. One that responds to execution friction with reorganization disrupts everything and changes nothing.
把两者混淆不是无害的不精确——它会把整个修复动作送到错误的地方。用流程提速来应对结构滞后的组织,只会更快地做错的事。用重组来应对执行摩擦的组织,会扰动一切而什么也没改变。
Misdiagnosing stock as flow, or flow as stock, is itself a measurable failure mode — and the standard exists in large part to catch it.
把存量误诊为流量,或把流量误诊为存量,本身就是一种可测量的失败模式——而标准存在的一大理由,正是抓住它。
Why AI made this urgent
为什么是 AI 让它变得紧迫
The gap is not new and is not a property of AI. Market signals, regulatory notices, simulation output and human expert judgment all enter an organization as intelligence that must be converted into consequence, and all of them are subject to the same lag and friction.
这个落差不是新的,也不是 AI 的属性。市场信号、监管通知、仿真输出、人类专家判断,都是进入组织、必须被转化为后果的智能,全都受同样的滞后与摩擦支配。
What changed is the ratio. Organizations acquired the ability to generate recommendations far faster than they acquired the ability to act on them. The backlog that produced was the first version of this gap large enough that nobody could look past it.
变的是比例。组织获得生成建议的能力,远快于它获得执行建议的能力。由此堆积出来的积压,是这个落差第一次大到没人能够视而不见。
What version 2.0 is organized around
2.0 围绕什么组织起来
Version 1.0 asked a diagnostic question: why does installed intelligence fail to become realized capability? That question is well posed, and the 1.0 measurement model answers a defined part of it.
1.0 提的是一个诊断性问题:为什么已安装的智能没能变成已实现的能力?这个问题提得好,而 1.0 的测量模型回答了其中界定清楚的那一部分。
But it is a question about failure, and a standard organized around failure accumulates failure taxonomies — a metric for each observed pathology, then another for each newly observed one, until the specification is a collection rather than a system.
但这是一个关于失败的问题。围绕失败组织起来的标准,会不断累积失败分类法——为观察到的每一种病症加一个指标,再为每一种新观察到的加一个,直到规范变成一个集合,而不是一个系统。
Version 2.0 is organized around a question that subsumes it:
2.0 围绕一个把它包含在内的问题组织起来:
How should an organization be designed so that intelligence continuously converts into action and realized capability?
一个组织应该如何设计,才能让智能持续地转化为行动和已实现能力?
This changes nothing about what the existing constructs measure. It changes what they are for. Each now occupies a defined position in a single answer:
这不改变既有构念测量的内容。它改变的是它们为何存在。每一个现在都在同一个答案里占据一个确定的位置:
| Construct构念 | Role in the conversion system在转化系统中的角色 | Section章 |
|---|---|---|
| Decision Ontology决策本体论 | describes conversion structure描述转化结构 | 03 |
| Decision Event Protocol决策事件协议 | makes conversion observable让转化可观测 | 04 |
| ICI / RCI / CCRICI / RCI / CCR | measures conversion magnitude测量转化的量 | 05 |
| Absorption Curve / AAR吸收曲线 / AAR | measures conversion velocity测量转化的速度 | 06 |
| OAFC / EFIOAFC / EFI | locates conversion loss定位转化损耗 | 07 |
| EDI redesignEDI 重设计 | changes the conversion architecture改变转化架构 | 08 |
| Verification验证 | establishes whether the change worked确认改动是否奏效 | 09 |
Read top to bottom, that table is also the rest of this guide.
从上往下读,这张表也就是本指南余下的部分。
What counts as a decision
什么才算一个决策
Everything downstream depends on a unit of analysis. DI-Std's unit is the decision — not the task, not the workflow step, not the user session. A candidate qualifies as a decision when three questions have answers:
下游的一切都取决于分析单元。DI-Std 的单元是决策——不是任务,不是工作流步骤,不是用户会话。一个候选对象只有在三个问题都有答案时,才算一个决策:
- Does it have its own moment of commitment? A point at which the organization is bound to a course of action.
- Does it have an owner? A named party accountable for the outcome, not merely for the completion.
- Does it have a consequence? Something observable that differs depending on which way it went.
- 它有自己的承诺时刻吗?一个组织被绑定到某条行动路线上的时点。
- 它有归属人吗?一个对结果负责的具名方,而不只是对完成负责。
- 它有后果吗?某种可观测的东西,会因为它往哪边走而不同。
This is deliberately decision-centred rather than interaction-centred. Who talked to whom is a projection of the decision structure, not the structure itself — which is why the standard can be strict about decisions while remaining almost entirely uninterested in surveillance of people.
这是刻意以决策为中心,而不是以交互为中心。谁跟谁说了话,是决策结构的一个投影,不是结构本身——这也是为什么标准可以对决策极其严格,同时几乎完全不关心对人的监控。
The seven-stage loop
七阶段决策环
A decision that has been instrumented passes through seven stages. The order is normative — a loop observed running out of order is not a variant, it is a defect.
一个装上了仪器的决策会经过七个阶段。顺序具有规范性——观察到乱序运行的环不是一种变体,而是一个缺陷。
Four operating models
四种运营模型
Not every decision should be run the same way. The ontology distinguishes four human–machine operating models, and a great deal of observed dysfunction turns out to be a decision running under the wrong one.
不是每个决策都该用同样的方式运行。本体论区分四种人机运营模型,而大量被观察到的功能失调,最后都归结为一个决策跑在了错误的模型下。
| Model模型 | Machine role机器角色 | Human role人的角色 |
|---|---|---|
| M1 | Acts within a declared bound在已声明的边界内行动 | Owns the bound and the exceptions拥有边界与例外 |
| M2 | Recommends, with reasoning给出建议,并附推理 | Commits, and is accountable for the outcome作出承诺,并对结果负责 |
| M3 | Assists and assembles evidence辅助并汇集证据 | Judges and commits判断并承诺 |
| M4 | Informs only仅提供信息 | Owns the decision entirely完全拥有该决策 |
Two kinds of rule
两类规则
This is the architectural distinction that most shapes how the standard behaves in practice, and it is worth understanding before anything else.
这是最能塑造标准实际行为的架构分野,值得先于其他一切来理解。
States that can never be authorized. An instance in one of these is not non-conformant-but-tolerated; it is rejected by the schema.
永远不可能被授权的状态。落在其中的实例不是「不合规但可容忍」,而是被 schema 直接拒绝。
- Loop stages out of order
- Decide never instantiated
- M1 running with no declared bound
- Operating model mismatched to decision class
- 环的阶段乱序
- Decide 从未被实例化
- M1 在没有声明边界的情况下运行
- 运营模型与决策类别错配
Pathological states the standard exists to find. Accepting them is the point; concealing them is not permitted.
标准正是为了发现而存在的病态状态。接受它们是本意;隐瞒它们不被允许。
simulation-absentconsequence-openowner-vacant/owner-diffuseunbounded-machine-commit
Machine commitment is prohibited as a design state, and observable as a diagnosed current state.
机器承诺作为设计状态被禁止,作为被诊断出的当前状态则可观察。
A standard that rejected every sick organization at the door would describe only organizations that did not need it.
一个把所有生病的组织都挡在门外的标准,最后只能描述那些本来就不需要它的组织。
Record state changes, not activity
记录状态改变,而不是活动
The obvious way to instrument decisions is to capture the work: meetings, messages, calls, screen time. It is also the way that fails — on privacy, on data volume, on cross-platform integration, and on the simple fact that most of the relevant conversation happens where no system is listening.
给决策装仪器的直觉做法,是捕捉工作本身:会议、消息、通话、屏幕时间。这也正是会失败的做法——败在隐私上,败在数据量上,败在跨平台集成上,也败在一个朴素的事实上:大部分相关的对话,发生在没有系统在听的地方。
The Decision Event Protocol takes a different cut. It does not ask what people did. It asks a narrower question with a far better answer rate:
决策事件协议换了一个切口。它不问人做了什么。它问一个更窄、但答案率高得多的问题:
What changed the state of this decision?
是什么改变了这个决策的状态?
Two people have a forty-minute call. Afterwards, the decision moved from pending to approved. What must be recorded is the transition and its timestamp. Not the recording, not the transcript, not who said what.
两个人打了四十分钟电话。之后,决策从 pending 变成了 approved。必须被记录的是这次状态转移和它的时间戳。不是录音,不是转写,不是谁说了什么。
This single choice resolves four problems at once — privacy exposure, surveillance objections, data volume, and the integration burden of reaching into every communication tool an enterprise owns. And it changes what DEP is actually specifying: not how all activity is logged, but what constitutes a meaningful decision state transition.
这一个选择同时解决四个问题——隐私暴露、监控异议、数据量,以及伸手进企业拥有的每一个沟通工具所需的集成负担。它也改变了 DEP 真正在规定的东西:不是所有活动如何被记录,而是什么构成一次有意义的决策状态转移。
Evidence classes
证据等级
Not all evidence of a transition is equal, and a measurement system that pretends otherwise can be inflated at will. Every recorded transition carries a class:
并非所有状态转移的证据都等价,而一个假装它们等价的测量系统,可以被随意注水。每一次被记录的转移都带一个等级:
| Class等级 | What it is含义 | Admissible?可否采信 |
|---|---|---|
| A | Verified system record — pulled from a system of record, timestamped已验证的系统记录——取自记录系统,带时间戳 | Yes可以 |
| B | Reported with supporting data — a person reports it, backed by an exportable artifact有佐证的报告——由人报告,并有可导出的材料支撑 | Yes可以 |
| C | Estimated or self-reported, no supporting artifact估计或自报,没有佐证材料 | Directional only仅作方向性参考 |
| D | Anecdotal, secondhand, or inferred轶事、二手或推断 | Directional only仅作方向性参考 |
Only classes A and B feed normative calculations. This closes an otherwise obvious gaming route: if self-reported evidence counted, any organization could raise its scores by loosening what it accepted from itself.
只有 A、B 两级进入规范性计算。这堵住了一条本来显而易见的博弈路径:如果自报证据算数,任何组织都可以靠放宽自己对自己的采信标准来抬高分数。
Decision Observability Ratio
决策可观测率(DOR)
Because evidence is graded, the instrument can report on its own blind spots — a property most measurement frameworks lack.
因为证据分了等级,这台仪器可以报告自己的盲区——这是大多数测量框架不具备的性质。
A low DOR is a finding about the data source, not about the organization. This distinction matters more than it looks: an organization with poor decision observability will appear frictionless, because the friction is simply invisible. Reporting DOR alongside every diagnosis is what stops that from becoming a flattering conclusion.
低 DOR 是关于数据源的发现,不是关于组织的发现。这个区分比看上去重要:一个决策可观测性差的组织,会显得毫无摩擦,因为摩擦根本看不见。在每一份诊断旁边一起报告 DOR,正是为了不让这变成一个讨人喜欢的结论。
How much has actually converted
究竟转化了多少
Three indices, and one discipline that makes them mean anything.
三个指数,以及一条让它们真正产生意义的纪律。
What the capability can do under a defined reference condition: the controlled environment in which it was accepted. Scored 0–100.
在既定参考条件下——也就是它被验收时的那个受控环境——这项能力能做到什么。0–100 计分。
What the capability is doing under actual operating conditions, in real decisions, with real people and real deadlines. Scored 0–100, on the same scale.
在真实运营条件下,在真实决策中,面对真实的人和真实的截止时间,这项能力正在做到什么。0–100 计分,同一把尺子。
CCR = RCI / ICI. The share of installed potential that has become real.
CCR = RCI / ICI。已安装潜力中,已经变成现实的那一份额。
The same-yardstick rule
同一把尺子规则
CCR is a ratio of two numbers, and a ratio is only meaningful if both numbers measure the same thing. This is where most capability measurement quietly breaks: the "installed" figure comes from a vendor benchmark and the "realized" figure comes from a user satisfaction survey. Those are two different instruments, and their quotient is arithmetic without content.
CCR 是两个数字的比值,而比值只有在两个数字测的是同一件事时才有意义。大多数能力测量正是在这里悄悄崩掉的:「已安装」的数字来自供应商基准,「已实现」的数字来自用户满意度调查。这是两台不同的仪器,它们的商是没有内容的算术。
DI-Std forbids it structurally. Capability is decomposed into four dimensions — Quality, Timeliness, Coverage, Reliability — and for each dimension, the installed side and the realized side must use the identical metric definition and unit. Only the measurement condition differs.
DI-Std 在结构上禁止这件事。能力被分解为四个维度——质量、及时性、覆盖度、可靠性——而对每一个维度,已安装侧与已实现侧必须使用完全相同的指标定义和单位。唯一允许不同的,是测量条件。
Defining what "good" means, once
把「好」定义一次
Meridian runs a replenishment model across roughly 2,000 stocking points. Before measuring anything, they fix four definitions and four weights, and freeze them for the life of the cohort.
Meridian 在约 2,000 个备货点上运行一套补货模型。在测量任何东西之前,他们先固定四个定义和四个权重,并在该队列的整个生命周期内冻结它们。
| Dimension维度 | Metric definition (binds both sides)指标定义(约束两侧) | Weight权重 | Reference参考值 |
|---|---|---|---|
| Quality质量 | service level attained on recommended orders建议订单达成的服务水平 | 0.40 | 90 |
| Timeliness及时性 | % available before the ordering window closes下单窗口关闭前可用的比例 | 0.20 | 95 |
| Coverage覆盖度 | reached / eligible replenishment decisions已触达 / 合格补货决策 | 0.25 | 90 |
| Reliability可靠性 | % uptime during ordering windows下单窗口期间的可用率 | 0.15 | 90 |
ICI = 0.40(90) + 0.20(95) + 0.25(90) + 0.15(90) = 91.0 — frozen at cohort admission, and not touched again.
ICI = 0.40(90) + 0.20(95) + 0.25(90) + 0.15(90) = 91.0 —— 在队列准入时冻结,此后不再改动。
The values are illustrative. What matters is the shape: four definitions, agreed once, binding on both sides of the ratio.
数值是示意的。重要的是形状:四个定义,约定一次,对比值的两侧同时具有约束力。
How fast, and has it stopped?
有多快,以及停了没有
CCR is a snapshot. It tells you where conversion stands today and nothing about where it is going. That is a serious limitation, because the two organizations below report the same number and are in completely different situations: one is three months into a healthy ramp, the other stalled nine months ago and has been flat since.
CCR 是一张快照。它告诉你转化今天站在哪里,却不说它正往哪里去。这是个严重的局限:两个组织可以报出同一个数字,处境却完全不同——一个在健康爬坡的第三个月,另一个九个月前就停住了,此后一直是平的。
Adding time seems trivial — differentiate CCR. It is not, and the reason is worth understanding, because it determines the entire structure of this layer.
加上时间看似平凡——对 CCR 求导就好。并非如此,而原因值得理解,因为它决定了这一整层的结构。
Why the denominator has to be frozen
为什么分母必须冻结
| ICI | RCI | CCR | |
|---|---|---|---|
| January一月 | 60 | 30 | 50% |
| March三月 | 90 | 45 | 50% |
CCR says nothing happened. In fact realized capability rose by half, from 30 to 45 — the organization bought more capability in February, and the denominator moved underneath the measurement. The reverse is just as easy to produce: a burst of procurement can make CCR fall in a quarter when the existing capability was being absorbed beautifully.
CCR 说什么都没发生。事实上已实现能力涨了一半,从 30 到 45——组织在二月又买了更多能力,分母在测量底下移动了。反过来一样容易发生:一波采购可以让 CCR 在某个季度下降,而存量能力其实正被出色地吸收。
Velocity cannot be computed against a denominator that is free to move. It can only be computed inside a unit where the denominator is frozen.
速度无法对着一个可以自由移动的分母计算。它只能在一个分母被冻结的单元内计算。
The smallest measurable unit, with its ICI baseline frozen at admission. Fields: cohort_id, capability_id, capability_version, decision_class, installed_at, measurement_start, ICI_baseline, DCT_version.
最小可测量单元,其 ICI 基线在准入时被冻结。字段:cohort_id、capability_id、capability_version、decision_class、installed_at、measurement_start、ICI_baseline、DCT_version。
Any material change to the capability — a version upgrade, a change of decision scope, a revision to a DCT definition — opens a new cohort, linked to the old one by supersedes / superseded_by. The old cohort keeps its history and stops accepting data. It is never edited.
任何对能力的实质性改动——版本升级、决策范围变更、DCT 定义修订——都会开启一个新队列,并通过 supersedes / superseded_by 与旧队列相连。旧队列保留自己的历史,停止接收新数据。它永远不被编辑。
This is a design rule rather than a convention, for a blunt reason: permitting a silent change of denominator is permitting a measurement that can be tuned by whoever reports it.
这是一条设计规则,不是惯例,理由很直白:允许分母被悄悄改动,等于允许一个可以由报数者自行调校的测量。
When the clock starts
时钟从哪里开始
The absorption clock starts at ICI admission — the moment the capability becomes demonstrably callable under the defined reference condition. Explicitly not the licence purchase date, and not the pilot kickoff. Both are administrative events; neither is a state of the world that can be evidenced.
吸收时钟从 ICI 准入开始——能力在既定参考条件下变得可被证明地调用的那一刻。明确不是 licence 采购日,也不是 pilot 启动日。两者都是行政事件,都不是可以取证的世界状态。
The curve, and its slope
曲线,以及它的斜率
Two summary points on the curve turn out to be what executives actually ask for:
曲线上有两个摘要点,恰好是高管真正会问的:
- T50 — months from clock start until half the installed capability is realized.
- T80 — the same at 80%. Where T80 is projected rather than observed, it must be labelled as projected and carry a confidence grade.
- T50 —— 从时钟起点到已安装能力实现一半所需的月数。
- T80 —— 同理,到 80%。若 T80 是外推而非实测,必须标注为外推,并附置信等级。
The state worth catching
值得抓住的那个状态
When the curve flattens across consecutive periods, the system reports Absorption Plateau. It does not report that adoption is complete. The difference is the entire value of the construct: a plateau is an unexplained state that triggers diagnosis, and a completion is a conclusion that ends inquiry.
当曲线在连续几个周期内走平,系统报告 Absorption Plateau(吸收平台期)。它不报告「采纳已完成」。这个差别就是整个构念的价值所在:平台期是一个触发诊断的待解释状态,而完成是一个终止追问的结论。
Fast to half, stuck at two-thirds
冲到一半很快,卡在三分之二
Meridian reached half of installed capability in 1.6 months — genuinely fast, and the number the programme reported upward. Then it stalled at roughly 65% and stayed there for three consecutive periods. The plateau flag fired at M6; without it, the next milestone review would have recorded a successful deployment.
Meridian 在 1.6 个月内实现了已安装能力的一半——确实快,这也是项目往上汇报的数字。然后它停在大约 65%,连续三个周期没有动。平台期标记在 M6 触发;没有它,下一次里程碑评审记录下来的会是一次成功的部署。
What AAR is not
AAR 不是什么
Why it stalled
它为什么停了
A plateau is a question. Answering it requires a different act from measuring — not how much was lost, but where. The claim this layer makes is that conversion loss has a location, and that the location is a specific step on a specific decision path rather than a diffuse cultural condition.
平台期是一个问题。回答它需要一个和测量不同的动作——不是损失了多少,而是在哪里。这一层提出的主张是:转化损耗有位置,而这个位置是某条决策路径上的某个具体步骤,不是一种弥散的文化状况。
What makes that claim testable is a single question, put to each step:
让这个主张可检验的,是一个问题,对每一个步骤提出:
What explicit risk, right, obligation or accountability does this step protect?
这个步骤保护了哪一项明确的风险、权利、义务或问责?
A step with an answer is Required. A step without one is Excess — an excess node, in the standard's vocabulary. Note what the question does not ask. It does not ask whether the step is slow, whether people resent it, or whether it seems bureaucratic. Those are judgements about experience; this is a judgement about function, and it is deliberately unflattering to both sides — it disqualifies steps that merely feel necessary, and it protects steps that are genuinely load-bearing however tedious they are.
有答案的步骤是必需的。没有答案的是多余的——按标准的术语,是一个 excess node。注意这个问题没有问什么。它没问这个步骤慢不慢、有没有人讨厌它、看起来是否官僚。那些是关于体验的判断;这是关于功能的判断,而且它对两边都刻意不留情面——它会取消那些仅仅「感觉」必要的步骤的资格,也会保护那些真正承重、无论多么繁琐的步骤。
Two quantities come out of this. Excess intervention load — how many unprotecting steps sit on a typical path — and F_T, the ratio of actual elapsed time to the reference path's legitimate minimum. Neither is a productivity measure. Both are statements about where a specific capability's output is being consumed before it reaches a decision.
从中得出两个量。多余干预负荷——一条典型路径上有多少个不保护任何东西的步骤——以及 F_T,实际耗时与参考路径正当最小值之比。两者都不是生产力度量。两者说的都是:一项具体能力的产出,在抵达决策之前被消耗在了哪里。
The honesty requirement on T_ref
T_ref 的诚实性要求
F_T = T_actual / T_ref looks like a measurement and is not entirely one. T_actual is observed. T_ref — the legitimate minimum elapsed time — is a judgement, and presenting a judgement as an observation is the most common way a measurement system quietly becomes unfalsifiable.
F_T = T_actual / T_ref 看起来像一次测量,但它并不完全是。T_actual 是观测到的。T_ref ——正当的最小耗时——是一次裁定,而把裁定当作观测来呈现,是测量系统悄悄变得不可证伪的最常见方式。
So T_ref carries a declared basis: who set it, on what grounds, and which role approved it. Without one, the instrument reports elapsed time and issues no F_T at all. The same governance applies everywhere in the standard that a normative judgement enters a calculation — it is the same rule that decides which evidence may feed RCI, and the same rule that forbids an unstratified percentile.
所以 T_ref 必须带一份声明依据:谁设定的、依据什么、哪个角色批准的。没有它,仪器只报告耗时,完全不出具 F_T。标准里凡是规范性裁定进入计算的地方,都适用同一套治理——它与决定哪些证据可以进入 RCI 的是同一条规则,与禁止未分层百分位的也是同一条。
A number the standard cannot justify, it declines to produce.
标准无法为之辩护的数字,它拒绝产出。
Where the missing third went
少掉的那三分之一去了哪里
The diagnosis found a typical path carrying two required steps and three excess ones, running at F_T 4.5× against a reference path whose basis was on file. The excess was not approval layering, which is what the programme had assumed. It was translation and rationale loss — the recommendation was re-keyed into the ordering system by hand, its reasoning did not survive the crossing and had to be reconstructed downstream, and a finance signature was being collected on orders already inside an agreed limit.
诊断发现一条典型路径上有两个必需步骤和三个多余步骤,相对一份依据在册的参考路径,运行在 F_T 4.5×。多余的部分并不是审批层叠——那是项目组原本的假设。它是翻译与理由丢失:建议被手工重新录入下单系统,它的推理没能跨过这次交接、只好在下游重建,而已经落在既定额度内的订单还在收取一个财务签字。
None of this appears in usage telemetry. The model was in constant use. It was being used into a path that consumed most of its timeliness, and much of its coverage, before anything was ordered.
这些在使用遥测里全都看不到。模型一直在被使用。它是被使用进了一条路径——这条路径在任何东西被订购之前,就消耗掉了它大部分的及时性和相当一部分覆盖度。
Deletion, not optimization
删除,而不是优化
Process improvement asks how to make an approval faster. It is the wrong question, or rather it is the third question. The first two are better:
流程改进问的是怎样让审批更快。这是错的问题,或者说,这是第三个问题。前两个更好:
- Why does this approval exist? — what does it protect that nothing else protects?
- Why does this decision enter an approval process at all? — given its risk, reversibility, value and exposure, is a human commitment event warranted?
- 这个审批为什么存在?—— 它保护了什么别的东西保护不了的?
- 这个决策为什么要进入审批流程?—— 考虑它的风险、可逆性、价值和暴露,一次人类承诺事件有必要吗?
Where a decision's properties permit, the target state is not a faster approval. It is no approval event: machine action inside a declared autonomy bound, with a named human owner engaged only when the decision falls outside it.
在决策的属性允许的地方,目标状态不是更快的审批。是没有审批事件:机器在已声明的自治边界内行动,只有当决策落在边界之外时,才由具名的人类归属人介入。
This is not an extension to the standard; it is already latent in it. Stage 06 of the loop is Act within bounds, and the design rule prohibiting unbounded machine commitment is precisely what makes acting within bounds safe rather than reckless. A declared bound is the thing that distinguishes delegation from abdication.
这不是对标准的扩展;它本来就潜藏在标准里。环的阶段 06 就是 Act within bounds,而禁止无边界机器承诺的那条设计规则,正是让「在边界内行动」变得安全而非鲁莽的原因。已声明的边界,是把授权和弃权区分开的那样东西。
An organization with no declared autonomy bounds has not avoided machine autonomy. It has made its machine autonomy undeclared.
一个没有声明任何自治边界的组织,并没有避开机器自治。它只是让自己的机器自治变成了未声明的。
Three changes, two of them deletions
三处改动,其中两处是删除
- Declared an autonomy bound. Orders within ±10% of the standing quantity and under a stated value ceiling are placed without a human commitment event. A named owner stays accountable for what happens inside the bound and reviews it in aggregate, not case by case.
- Deleted the finance signature inside that bound — it had been authorizing what the bound already authorized, protecting nothing the limit did not already protect.
- Carried the reasoning across the handover so it no longer had to be rebuilt downstream. The reconstruction work existed because the interface dropped something, not because anyone needed to think again.
- 声明了一条自治边界。在常备订货量 ±10% 以内、且低于既定价值上限的订单,不经人类承诺事件直接下达。一位具名归属人仍对边界内发生的事情负责,并以汇总方式而非逐单方式复核。
- 删掉了该边界内的财务签字——它一直在授权边界已经授权过的东西,没有保护任何额度本身没有保护的内容。
- 让推理跟着交接一起过去,于是它不必在下游被重建。那些重建工作之所以存在,是因为接口丢了东西,不是因为谁需要再想一遍。
Note what was not done: nobody was asked to work faster, no meeting was shortened, and no additional dashboard was built.
注意没有做的事:没有人被要求干得更快,没有会议被缩短,也没有新增任何看板。
Did the slope change?
斜率变了吗
A redesign that is not measured afterwards is an opinion with a project plan attached. The loop closes by re-measuring the same cohort, on the same frozen baseline, with the same definitions — which is exactly what the cohort discipline was for.
一次事后不做测量的重设计,是一个附了项目计划的意见。环的闭合方式是:在同一个队列上、同一条冻结基线上、用同一套定义重新测量——这正是队列纪律当初的用途。
The test is not whether CCR improved. CCR improves for many reasons, including a good quarter. The test is whether the rate changed: a redesign that addressed real conversion loss should show up as a change in slope, in the same cohort, shortly after the intervention.
检验的不是 CCR 有没有改善。CCR 因为很多原因都会改善,包括一个好季度。检验的是速率有没有变:一次真正处理了转化损耗的重设计,应该在同一个队列里、在干预之后不久,表现为斜率的改变。
Before and after, same cohort, same yardstick
前后对比,同一队列,同一把尺子
| M3 → M6 | M6 → M8 | |
|---|---|---|
| AAR | 0.8 pp/mo | 8.0 pp/mo |
| CCR | 63% → 66% | 66% → 81% |
| Excess steps / path每条路径的多余步骤 | 3.0 | 1.0 |
| F_T | 4.5× | 1.8× |
Coverage and timeliness carried most of the gain, which is what the diagnosis predicted: those were the dimensions the excess path was consuming. Quality moved least — the model was always sound, it simply was not reaching enough decisions in time to matter.
覆盖度和及时性贡献了大部分增益,这正是诊断预测的:它们是那条多余路径一直在消耗的维度。质量变化最小——模型本来就是可靠的,只是没能及时抵达足够多的决策。
One discipline holds here as everywhere: this is a before-and-after on a single cohort, not a controlled experiment. It is evidence that the intervention and the change in slope coincided in a cohort whose measurement basis did not move. That is a real claim, and it is a narrower claim than causation.
有一条纪律在这里和在别处一样成立:这是单个队列上的前后对比,不是一次受控实验。它构成的证据是:干预与斜率的改变,在一个测量基线未曾移动的队列里同时发生。这是一个真实的主张,而它比因果关系要窄。
What may be compared with what
什么可以和什么比较
The commercially attractive output is a ranking: your absorption is 1.8, the industry runs at 2.5. It is also the fastest available way to destroy the instrument's credibility, because absorption rates are not comparable across dissimilar capability types. A demand-planning capability and a coding assistant have different natural absorption profiles for reasons that have nothing to do with how well either organization is run.
商业上最有吸引力的输出是排名:你们的吸收是 1.8,行业跑在 2.5。这也是摧毁这台仪器可信度最快的方式,因为吸收速率在不相似的能力类型之间不可比。一项需求计划能力和一个编码助手有不同的天然吸收曲线,原因与两个组织各自经营得好不好毫无关系。
So percentile and ranking claims are permitted only after stratification on four axes:
所以百分位和排名类主张,只有在按四个轴分层之后才被允许:
Which licenses statements of this shape:
它许可这种形状的陈述:
For M2 demand-planning capabilities at cohort month 4, this cohort's AAR sits in the 38th percentile.
对于处于队列第 4 个月的 M2 需求计划类能力,本队列的 AAR 位于第 38 百分位。
And forbids statements of this shape:
并禁止这种形状的陈述:
Company A · AI absorption score: 72
A 公司 · AI 吸收评分:72
The same logic already governs the rest of the standard: a boundary must be declared before a number is issued inside it. A single organizational score is not a simplification of this rule, it is an abandonment of it.
同样的逻辑早已支配着标准的其余部分:边界必须先被声明,数字才能在其中出具。单一的组织评分不是这条规则的简化,而是对它的放弃。
Four things this is often mistaken for
它常被误认成的四样东西
The underlying position
底下的立场
Every discipline in this guide — the frozen denominator, the single binding metric definition, evidence classes, the declared basis on T_ref, stratification before percentiles — costs something. Each one narrows what the standard is willing to claim.
本指南中的每一条纪律——冻结的分母、唯一具有约束力的指标定义、证据等级、T_ref 的声明依据、百分位之前的分层——都有代价。每一条都在收窄标准愿意主张的范围。
That narrowing is the product. In a category where every vendor publishes a maturity framework and every platform publishes an adoption trend, the scarce thing is not another number. It is a number whose limits are stated as clearly as its value.
这种收窄就是产品本身。在一个每家供应商都发布成熟度框架、每个平台都发布采纳趋势的品类里,稀缺的不是又一个数字。是一个把自己的界限讲得和自己的价值一样清楚的数字。