第 7 章 · 20 分钟

Reverse audit: checking the modelling commitments of chapters 1–6

The audit covers chapters 1 to 6, including right conclusions reached wrongly.

概念地图

Two ways

Count one set in two ways and an identity falls out.

本章不引入新概念，图为全书脉络

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。

对话只存在本机

Locus01

延伸阅读

Pólya · mid twentieth century · How to Solve It and plausible reasoning

Wrote the process of guessing then proving into teachable steps and insisted on asking afterwards where else the method applies. He argued against teaching only the finished proof; the cost is that the looking-back step is the one most often skipped in practice, because it produces no new answer.

Foundational

White · 2000 · data snooping and the reality check

Showed that repeated selection on one sample systematically inflates in-sample performance by an estimable amount. It turned 'right for the wrong reason' from a warning into a quantity, and this course's bookkeeping follows that line.

Turn

Chapters 1 to 6 of this course

The subject of the audit. The third block (derivation) and fifth block (boundary) of each chapter are the entry points, with the focus on whether that chapter's one modelling claim stands.

Under audit

机制02

Computed right is not applied right

Combinatorial formulas hardly ever compute wrongly — they are definitions unpacked and arithmetic, and they hold at all times. What goes wrong is always the step before applying them: the equivalence was never fixed, the classification overlapped, the decomposition was not unique, the discarded information mattered. This course books that step separately as a modelling commitment precisely so that the certainty of the formula does not vouch for it.

机制A formula beyond doubt, through making the whole computation look sound, spares the step before it from inspection.

可迁移性测试

Move it to units: the conversions are all correct and the instrument was read on the wrong range. Rechecking the arithmetic will never find it, because the arithmetic is not wrong. The isomorphism breaks in that the range is at least printed on the instrument.

机制03

Integrality is the free check

Counting problems have a convenience other fields lack: the answer must be a non-negative integer, and small cases can usually be enumerated outright. Chapter 2's non-integer quotient, chapter 3's class counts exceeding the total, chapter 5's recurrence overshooting enumeration — all three failures surface immediately under those two checks. This course books it as cheap falsification: you do not have to wait for the conclusion to fail in use.

机制A counting result, through integrality and small-scale enumeration, exposes a failed hypothesis before it is applied.

可迁移性测试

Move it to bookkeeping: debits must equal credits, a check that costs nothing and surfaces any mis-entry on the spot. The isomorphism breaks in that balanced books guarantee balance and not correctness, while small-scale enumeration is complete verification at small scale.

Derivation04

Counting the categories in chapters 1–6

This aggregates the tags across the first six chapters and derives one conclusion: the observational risk sits on six modelling claims, five of which have a cheap check available.

Chapters 1 to 6 have 4 derivation steps each, 24 in total
Counting rule: only the steps inside each chapter's third block, not the boundary or debate sections.
Twelve are tagged [definition] and six are tagged [assumption]
Definitions and arithmetic carry no observational risk; assumptions are hypotheses, and dropping one makes the claim false.
Six are tagged [modelling], one per chapter
In a formal subject this is the only class observation can refute.

推导 · 0 / 4

裂缝05

The boundary of this chapter's claims

争议地形06

本书主张

This course's position: in combinatorics the only thing worth checking repeatedly is the modelling decision made before a formula is applied; the rigour of the formula guarantees nothing about the applicability of the conclusion.

另一种看法

Another line holds that for most users, fluency with the formulas repays more than interrogating modelling decisions — real counting problems are mostly of standard shape, and the chance of applying the wrong formula is far below the chance of miscalculating.

分歧扎在

The disagreement is rooted in the actual distribution of errors. If most errors are arithmetic, the fluency claim holds; if most come from readings and classifications, this course's position does. Neither disputes the mathematics.

什么证据能裁决

What would settle it is a survey of error sources: collect real counting errors and classify them as 'arithmetic' or 'modelling decision', and see which dominates. The classification rule must be fixed in advance and executed by someone blind to the hypothesis.

The evidence points to the modelling side and is weak: existing cases come mostly from teaching settings, whose distribution need not match professional work. What is missing is a sample drawn from real working output.

Boundary and counterexample07

The 'cheap falsification' commitment as a testable specification

The modelling commitment under test: integrality and small-scale enumeration are sensitive to all five kinds of hypothesis failure this course lists.

Hypothesis dropped: every kind of failure changes either integrality or the small-scale count
The prerequisite for the cheap checks to suffice; a failure invisible at small scale slips through.
The counterexample: chapter 6's kind (the discarded information mattering) leaves the small-scale enumeration entirely unchanged
Isomorphic graphs give identical counts at every size.
The observable: the firing rate of the two checks after each kind of failure is deliberately injected
Failures must be injected by the definitions in each chapter's boundary block, not waited for.
What counts against it: the firing rate for some kind falls below a pre-registered threshold
The threshold is fixed in advance on how low stops counting as sensitive.
Pre-registered failure condition: any kind's firing rate below threshold, or the checks firing often with no failure present
The second points to excessive false positives, which also destroys their value.

推导 · 0 / 3

接口08

Which slot it hangs on

挂在哪个槽位The conclusion you already hold is probably 'if the arithmetic is right there is no problem' — a judgement about where reliability comes from. This chapter goes at which layer it applies to.

Does this chapter (a) replace 'right arithmetic means no problem', or (b) constrain it: the formula is right and the risk sits entirely in the step before it?

慢变量Register two slow variables: how many of your recent counts had the equivalence written down, and how many were checked against small-scale enumeration. Look at the second first.

小结09

本章小结

01Of 24 derivation steps in chapters 1–6, twelve are definitions, six are assumptions, and six modelling claims carry the observational risk.

02Combinatorial formulas hardly ever miscompute; what fails is always the modelling decision before them.

03Integrality and small-scale enumeration are two free checks sensitive to five of those kinds.

04Chapter 6's kind (the discarded information mattering) has no cheap check and needs separate verification.

05Which step counts as a modelling claim is the analyst's call, so this chapter is bookkeeping and not evaluation.

提取练习 · 合上书，先自己答一遍。

?Is 'risky steps equal the total minus definitions minus assumptions' (a) classification arithmetic or (b) a modelling commitment about reality?

?Does a non-integer quotient expose (a) an arithmetic slip or (b) the failure of 'uniform multiplicity'?

Forced choice10

A · Identify the mechanism

Which of these is a mechanism?

二选一

Forced choice11

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。