← Principles
第 4 章 · 22 分钟
Root causes are not proximate causes
Verbs are proximate, adjectives are root — a sharp test whose cost is that it points at people.
概念地图
Root cause
Verbs are proximate, adjectives are root; a sharp test whose cost is that it points at people.
深色为本章概念，浅色为其他章
陪练模式
不想只是读？让它带你走一遍。
导师把这一章拆成小步，每一步先问你，再讲；你用自己的话答，它先认下对的部分再纠偏。用你自己的 AI，对话只存在这台设备上。
Locus01
延伸阅读
Industrial root-cause questioning · mid twentieth century
Asks why repeatedly until the answer stops pointing at another shallower cause. It argued against stopping once the surface is repaired; the cost is that the questioning has no stopping rule — how many layers count as deep enough rests entirely on the asker's judgement at the time.
Foundational
Dalio · the diagnosis section · proximate versus root
Proximate causes are typically actions or inactions described with verbs; root causes are the deeper recurring reasons, and he holds they are usually adjectives describing what a person is like. He argued against booking problems against actions; the cost is that this test puts the end of diagnosis almost inevitably on a person.
Turn
The paired principle · mistakes are good if they produce learning
Ties error to learning, offsetting the side effect of the previous item: if root causes point at people, then naming an error must not mean punishing one, or nobody will say the root cause out loud. This is the interface to chapters 5 and 6.
Turn
The systems-safety line · human error is a consequence, not a cause
Holds that attributing an accident to a person's traits hides the design and conditions that made the action look reasonable; an investigation should ask what made it seem sensible at the time. This is the strongest opposition here, and the debate block lands on it.
Counterexample
机制02
A test made of parts of speech
How deep a diagnosis has to go is vague in most accounts. This book gives an unusually operable test: look at the part of speech of the answer you wrote down. A proximate cause comes out as a verb and is still at the level of the action; only when it comes out as an adjective has it reached the layer that keeps operating. What this buys in the argument is a stopping condition for 'keep asking why' — no more judging by feel whether you have asked enough, just look at the part of speech.
机制A part-of-speech test, through requiring the answer to move from verb to adjective, makes 'ask to the bottom' checkable on the spot.
可迁移性测试
Move it to a bug report: 'a null check is missing here' is the action layer; 'this module's null convention was never written down' is the layer that keeps operating. The isomorphism breaks in that a codebase can be searched for how many instances of the same omission exist, while a person's traits admit no such search.
机制03
The test necessarily points at people
The book says plainly that a root cause usually describes what a person is like. That is not incidental wording; it fixes where the whole diagnosis lands. Once the end point is a person, naming a root cause is assessing a person, so two further things must be arranged: making people willing to be named, and making the namer qualified. The first is handled by pairing error with learning, the second is left to chapter 6's believability. See this dependency and it is clear why the practice needs a whole culture to run at all.
机制A test that ends on a person, through turning diagnosis into assessment, forces the practice to carry protection and authorisation with it.
可迁移性测试
Move it to a post-mortem: once the conclusion reads 'X is careless', the meeting turns from finding problems into assigning responsibility, unless it was agreed in advance that this line does not enter performance review. The isomorphism breaks in that meeting rules can be written down, while the promise not to count it has no way of being verified.
Derivation04
Why fixing only the proximate cause lets it recur
What is being reconstructed is the thought behind the proximate/root split, and where this test puts its cost.
One root cause yields different proximate causes on different occasions
A premise the author chose. Forgetfulness yields not checking the timetable, leaving papers behind, missing a reply.
Fixing one proximate cause removes that one action only
A premise the author chose. Setting one alarm does not change forgetfulness.
Root causes are usually traits of a person, and so can be written as adjectives
A premise the author chose, and the one bearing the weight. The systems-safety line objects exactly here: it holds root causes often live in design and conditions, and those come out as nouns rather than adjectives.
推导 · 0 / 4
Worked derivation05
Taking the missed-train example one layer down
'I missed the train because I did not check the timetable' — by the part-of-speech test, what comes next? Put two candidates side by side.
Write two adjective candidates, then see what each of them predicts.
裂缝06
Cracks in this reading
争议地形07
本书主张
The root cause of a problem is usually a recurring trait of a person; only an answer written as an adjective has reached the bottom, and fixing the action layer lets the problem recur structurally.
另一种看法
The systems-safety line holds that attributing an accident to personal traits conceals the design and conditions that made the action reasonable at the time; the investigation should ask what made it look sensible, and the remediation should move process, interfaces and constraints.
分歧扎在
The disagreement is rooted in whether the subject of a root cause should be a person or a design, not in whether to ask deeper. Both sides grant that stopping at the action layer lets problems recur.
什么证据能裁决
What would settle it is the recurrence rate after remediation: split the same class of incidents in two, remediate one by personal attribution (training, reminders, replacement) and the other by system attribution (process, constraints, interfaces), then compare later incidents of the same class. Aviation and medicine already hold public records of exactly this.
In high-risk domains the balance clearly favours system attribution: those fields moved 'human error' from the cause column to the consequence column decades ago, with documented improvements in recurrence. But keep the range in view — the actions there are highly standardised and the design is easy to change; the judgement work this book addresses has far fewer interfaces to alter, so the personal column may not be emptied entirely. What is missing is the same comparison inside judgement work.
Falsification08
Turning 'roots usually sit in people' into a testable specification
The claim under test: for the same class of problem, remediating by personal traits lowers recurrence more than remediating by design and conditions.
Data source: one organisation's remediation records for a class of problem, with each remediation's type (personal / design) and the timing of later problems of that class
The type is tagged at the time of remediation; retrospective tagging follows the result.
Sample range: at least 50 cases of the class, with no fewer than 20 of each remediation type
Both types need enough cases or no difference can be seen.
Time window: occurrences of the same class within 6 months after each remediation
Fixed in advance and not to be adjusted for the result.
Decision threshold: recurrence in the personal group at least 30% lower counts as support for the book
The reverse, 30% lower in the design group, counts as support for the systems line; in between is recorded as undiscriminated.
Failure condition: the personal group's recurrence is not lower than the design group's
That counts as the book's claim not holding in this domain.
推导 · 0 / 3
接口09
Which slot it hangs on
挂在哪个槽位The conclusion you already hold is probably 'find the underlying cause when something goes wrong' — a demand about depth. This chapter goes at the two things it never supplies: when you have reached the bottom, and who the subject should be.
Does this chapter (a) replace 'find the underlying cause', or (b) constrain it: after writing a root cause you must write one sentence saying what it predicts, then go and check whoever the prediction falls on?
慢变量Register two slow variables: how many root-cause conclusions you have written recently, and for how many of them you can say what they predict. Look at the second first.
小结10
本章小结
01The part-of-speech test is the most operable thing in the book: verbs are proximate, adjectives reach the layer that keeps operating.
02'Fixing only the proximate cause recurs' unpacks a one-to-many relation; it is not a lesson from experience.
03The test puts the end of diagnosis on a person, so it must come with error-tied-to-learning and with believability.
04It guarantees only that you reached the adjective layer, not that the subject is right — this course's added check is 'write what it predicts'.
05How many layers count as the bottom has no stopping rule, so this course books it as unadjudicable.
提取练习 · 合上书，先自己答一遍。
?Is 'one root yields several proximates, so fixing one lets it recur' (a) a one-to-many relation unpacked or (b) an empirical claim observation could overturn?
?Does the systems-safety objection attack (a) whether to ask deeper or (b) whether the subject of a root cause should be a person or a design?
Forced choice11
B · Identify the tag
'A diagnosis that stops at the proximate cause recurs structurally' — what kind of step is that?
二选一
Forced choice12
D · Judge the interface
Which one hangs on a slot in your existing structure?
二选一
读到这里，把它记为已读。