Academic and Engineering Walkthrough
The errors in this chapter create the illusion of a completed argument through a hidden assumption, an incorrectly chosen standard, or ambiguous language. They cannot be diagnosed with a single truth table: it is necessary to restore the context, values, and obligations of the proponent.
Chapter Rule
Any controversial premise must have its own justification.
Reformulation, repetition, and shifting the burden are not justifications.
1. Begging the question: circularity in proof
Direct circle
Документ достоверен, потому что в нём всё написано верно.
Circle through a chain
A истинно, потому что B.
B истинно, потому что C.
C истинно, потому что A.
Circle in the Metric
Мы считаем модель качественной, если она совпадает с разметкой.
Разметку считаем качественной, потому что модель уверенно её воспроизводит.
Without an independent gold set the system confirms itself by itself.
Checking the graph of grounds
Draw a directed graph утверждение ← основание. A cycle is acceptable as
a model of mutual consistency, but does not provide an independent foundational basis.
2. Loaded question: complex question
Почему вы продолжаете скрывать инциденты?
The question implies:
- incidents were hidden;
- action continues;
- subject is exactly this.
The correct answer is not required to be почему:
First, check the premise. Which events do you consider hidden and by what
rule should they have been published?
Насколько вам понравилась улучшенная навигация?
Implies improvement and biases the evaluation. Neutral:
Как изменилась возможность найти нужный раздел?
3. Burden of Proof
The primary burden lies with the one who puts forward a substantive thesis:
В этой библиотеке есть критическая уязвимость.
The answer “prove that it doesn’t exist” shifts a practically impossible task.
Burden may transfer
If sufficient evidence is presented, the opposing party should:
- Challenge the data;
- Present an alternative;
- Limit the conclusion;
- Provide a counterexample.
The burden is not permanently attached to one person; it moves along the course of the dialogue.
Risk Asymmetry
In safety/security the absence of complete proof of harm does not always mean
inaction. The decision considers the cost of type I and type II errors. But
caution does not make the fact of a threat proven.
4. Appeal to Ignorance
Никто не доказал, что система уязвима.
Следовательно, она безопасна.
or:
Никто не доказал, что система безопасна.
Следовательно, она уязвима.
Absence of evidence becomes evidence of absence only when:
- the appropriate method was used;
- the method would with high probability detect the object;
- the search area is sufficient;
- the result is reproducible.
После полного сканирования образа известным набором сигнатур у нас нет
свидетельств этих конкретных CVE.
This is a bounded inference, not a “safe image in general”.
5. Appeal to Personal Distrust
Я не понимаю, как система может выдержать такую нагрузку.
Следовательно, результаты неверны.
Imagination complexity — knowledge information about the speaker. Useful reaction:
What mechanism ensures the result and what measurements confirm it?
6. False dilemma
Либо микросервисы, либо не масштабируемся.
Checks:
- do the options exhaust the space?
- does the division use a single basis?
- are there gradations and hybrids?
- can the solution be postponed or changed in stages?
Not every dilemma is false
Токен либо валиден в момент t, либо не валиден по заданному предикату.
An exact binary predicate provides exhaustive partitioning.
7. Continuum and sorites
From the absence of a sharp boundary, they conclude that there is no difference:
Нельзя назвать точный запрос, после которого система стала перегруженной.
Следовательно, перегрузки не существует.
Ambiguous boundaries do not destroy useful classes. Operational thresholds are needed,
intervals of uncertainty and decision sensitivity.
Reverse fallacy — to consider an artificial threshold a natural gap:
79 мс — быстро, 80 мс — медленно.
The threshold may be managerial, rather than ontological.
8. Slippery slope: slippery slope
Если разрешим один ручной deploy, скоро все контроли исчезнут и произойдёт
катастрофа.
The error is not in the consequence chain as such, but in the missed mechanisms and
probabilities.
For each edge ask:
- transition mechanism;
- probability;
- time;
- reversibility;
- barrier;
- event independence.
When the slope is justified
If there is:
- positive feedback;
- path dependence;
- erosion of norms through precedents;
- real data on transitions;
- absence of stopping rules.
Example: accumulation of unlimited exceptions from policy can really
break control. A term, owner, and automatic expiration are needed.
9. Special pleading: special exception
General rule applies to all, except a convenient case, without an independent
criterion:
Все команды обязаны проходить security review.
Наш проект экспериментальный, поэтому нам не нужно.
Experimentality may be relevant if the policy sets exceptions in advance
for data, access, and audience. Then it is not special pleading, but
a rule with a scope.
Check:
Would this criterion be adopted before knowing who benefits from it?
10. Moving goalposts: moving the goalposts
After completing the proof standard, the standard is changed:
— Покажите benchmark.
— Вот воспроизводимый benchmark.
— Но он не на production.
— Вот production canary.
— Всё равно нужен год наблюдений.
New questions may be reasonable. An error occurs if:
- the initial criterion was declared sufficient;
- the change is not explained by new information;
- no possible outcome allows the acceptance of the thesis.
Fix the decision rule before the check.
11. No true Scotsman: protecting a generalization by redefinition
Настоящие микросервисы не создают distributed monolith.
Вот система, называемая микросервисной, которая создала.
Значит, это были не настоящие микросервисы.
It is possible that the definition actually excludes the object. Question: was the criterion
set in advance and independently or added after the counterexample?
12. Suppressed testimony and cherry-picking
The conclusion is built on the convenient part of available data:
Показываем conversion после запуска, но скрываем рост возвратов и период
сезонной акции.
This is not only a “bad selection,” but also a violation of the obligation to present
significant counterarguments.
Check:
- time window;
- metric definition;
- missing segments;
- negative results;
- methodology changes;
- preselected hypothesis.
13. Fallacy of the Middle Position
Один говорит, что latency 20 мс, другой — 500 мс.
Истина, вероятно, 260 мс.
Truth need not lie in the middle. Compromise is useful for distributing
resources or values, but a measurement question is resolved by data.
14. Nirvana fallacy: comparison with the ideal
Canary не предотвращает все инциденты.
Следовательно, canary бесполезен.
Compare with a real alternative:
ожидаемый ущерб без контроля
ожидаемый ущерб с контролем
стоимость и побочные эффекты контроля
Reverse fallacy — accepting any improvement without comparing it to the best available
alternative.
15. Equivocation: changing the meaning
Open source означает открытый исходный код.
Открытое безопаснее скрытого.
Следовательно, любой open-source компонент безопаснее закрытого.
In the second shipment, “open” may mean checkable, not actually
checked and accompanied.
Indexing Technique
open₁ = доступен исходный код
open₂ = прошёл широкую независимую проверку
After indexing the transition disappears.
16. Amphiboly
Grammar ambiguity:
Мы удаляем данные пользователей с истёкшим согласием.
Has the user’s consent expired or the data retention period? Rewrite:
Когда срок действия согласия пользователя заканчивается, система удаляет
перечисленные категории данных в течение N дней, если нет иного основания.
17. Emphasis and Quoting
Исследование: «эффект обнаружен в exploratory analysis, но не подтвердился
после коррекции».
Заголовок: «Исследование обнаружило эффект».
A part of the quotation is true, the final thesis is distorted.
Check:
- complete sentence;
- constraints;
- negation;
- units;
- absolute values;
- baseline.
18. Composition
Properties of parts are transferred to the whole:
Каждый сервис выдерживает 1000 RPS.
Следовательно, цепочка из десяти сервисов выдерживает 1000 RPS.
The narrowest point, fan-out, correlated peaks, and common
dependencies emerge.
When a move is permissible
If a property is additive or the composition law is proven:
масса целого = сумма масс частей
But the latency of a sequential chain adds up, and the availability multiplies.
Need to name the composition operator.
19. Division
A property of the whole is transferred to a part:
Компания прибыльна.
Следовательно, каждый продукт прибыльный.
A loss-making product can be subsidized by others.
Engineering example:
Система доступна 99,9%.
Следовательно, каждый компонент доступен 99,9%.
Components may have different SLO, redundancy, and critical paths.
20. Category Error and Reification
Где физически находится permission?
Что хочет алгоритм?
Архитектура решила разделить команды.
Abstractions are useful, but causal explanation requires a mechanism:
- permission — relation of subject, resource, and action;
- algorithm optimizes a given function, and does not have desire;
- decision was made by people under constraints.
Materialization is especially dangerous when discussing AI: the words “understands”, “lies”, “remembers” should be expanded into observable behavior and system design.
21. Etymological fallacy
Modern meaning is derived from the origin of the word:
«Данные» значит «данное», поэтому данные объективны.
Value is defined by usage and the model, while data quality — by the retrieval process.
Etymology helps understand the history of a concept, but does not establish its current
essence.
22. Fallacy of defining as belief
The definition is formulated so that the evaluation is already embedded:
Микроменеджмент — необходимый контроль безответственных сотрудников.
Neutral:
Practice of detailed control of decisions and execution with low employee autonomy.
Causes and effects are proven separately.
23. Diagnostic Map
flowchart TD
A["Спорная реплика"] --> B{"Тезис уже содержится
в основании?"}
B -- да --> B1["Круг / begging the question"]
B -- нет --> C{"Вопрос содержит
непринятые предпосылки?"}
C -- да --> C1["Loaded question"]
C -- нет --> D{"Варианты искусственно
ограничены?"}
D -- да --> D1["False dilemma / continuum / slope"]
D -- нет --> E{"Стандарт меняется
после ответа?"}
E -- да --> E1["Moving goalposts / special pleading"]
E -- нет --> F{"Термин или синтаксис
сменил смысл?"}
F -- да --> F1["Equivocation / amphiboly / accent"]
F -- нет --> G{"Свойство перенесено
между частью и целым?"}
G -- да --> G1["Composition / division"]
G -- нет --> H["Проверить данные, причинность и контекст"]
24. Checklist
- Does the thesis not belong to its own basis?
- Does the question not impose a controversial premise?
- Is the burden of proof distributed among the advanced theses?
- Has the absence of data been obtained through a sensitive search?
- Are the alternatives exhaustive?
- Has the threshold not been turned into a natural gap?
- Does the chain of consequences have a mechanism and probabilities?
- Is the exception defined in advance and symmetrically?
- Does the acceptance criterion not move after the result?
- Is a counterexample not excluded by redefinition post hoc?
- Are essential data not suppressed?
- Is the term and grammatical structure unambiguous?
- Does the property indeed transfer between part and whole?
- Is abstraction revealed through a mechanism?
25. Tasks
- “No one has proven that the new model discriminates, therefore it is neutral.”
- “Either we adopt an AI-first strategy or we lose the market.”
- “One feature flag is safe; if we allow it, in a year the entire product will become a set of exceptions.”
- “A true DevOps does not require a separate ops team.”
- “Each module is covered at 90%, therefore the system’s behavior is covered at 90%.”
- “The service is critical because it is in the list of critical ones; it is in the list because it is critical.”
- “Why did you again choose an unsafe protocol?”
For each: extract the premise, specify the missing criterion and suggest
a testable formulation.
Sources
What’s next
The next layer is conclusions that formally do not promise necessity, but
assess probability and causes. It is precisely there that base rate neglect,
prosecutor’s fallacy, survivorship bias, regression to the mean, and
correlation-causation appear.
Causal and statistical fallacies
