← Thinking, Fast and Slow
第 2 章 · 20 分钟
Coherence before completeness
What you see is all there is, and the confidence it produces.
概念地图
Coherence
A story is built from what is at hand, with no slot for what is missing.
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
Asch · 1940s · order effects in impression formation
Found that the same list of traits in a different order produces a different overall impression, against the view that impressions are the sum of separate traits. The cost is that 'overall impression' can only be defined by the participant's own rating, with no independent measure.
Foundational
Tversky and Kahneman · 1983 · the conjunction fallacy
Used the Linda problem to show that adding a specific detail makes an event seem more likely, though a conjunction cannot be more probable. It made 'coherence beating probability' measurable.
Turn
Hertwig and Gigerenzer · 1990s
Argued that 'probable' in ordinary usage does not mean mathematical probability, so participants are answering a different question than the one intended. The strongest reply, and rewording the item does reduce the effect.
Counterexample
Kahneman · 2011 · the chapters on WYSIATI
Distils coherence-first into a general rule and derives the confidence illusion from it. The book does not separate 'concluding while information is missing' from 'overrating your certainty while information is missing'; the first crack lands there.
Subject of this course
机制02
Coherence first
The goal of the judging system is not to weigh all the evidence but to assemble a story that hangs together, quickly. Missing information is not registered as missing; it simply does not participate. Kahneman calls this WYSIATI. It explains something counterintuitive: with less information the story is easier to make coherent, so confidence can actually be higher.
机制Missing information, through never being registered, makes the story more coherent and so raises confidence.
可迁移性测试
Move it to accident investigation: with one witness statement the sequence of events looks perfectly clear; by the fifth it no longer does. The isomorphism breaks in that an investigation is institutionally required to collect every statement, while everyday judgement has nothing that forces you to count what is missing.
机制03
Confidence tracks coherence, not accuracy
The next step is the one with the largest practical consequence in the book: subjective confidence is driven mainly by how coherent the story is, not by the amount or quality of the evidence. This is the illusion of validity. The practical upshot: you cannot use 'I am sure' to tell whether you are right this time. A link between confidence and accuracy has to be built by outside calibration.
机制The coherence of a story, through standing in for evidential strength, becomes the source of confidence and decouples it from accuracy.
可迁移性测试
Move it to weather forecasting: if a forecaster's certainty comes from 'I have seen this chart before' rather than from a calibration table, certainty and hit rate have no necessary connection. The isomorphism breaks in that meteorological services publish hit rates, closing the loop, while most professional judgement has no scorecard at all.
Derivation04
From coherence first to 'less information, more confidence'
What is being derived is the least intuitive and most testable prediction in the chapter.
Confidence is increasing in the coherence of the story
Chosen, and the heaviest line here: it fixes the source of confidence on coherence.
The more evidence, the higher the chance that some of it conflicts
Chosen, and fairly safe, so long as the evidence is not from one identical source.
Conflicting evidence lowers coherence
Chosen: it assumes the judge does not fold the conflict into a more complex story. Training may change this.
推导 · 0 / 4
裂缝05
Cracks in this framework
争议地形06
本书主张
Judgement puts coherence before completeness; missing information is not registered, so less information makes a smoother story and possibly more confidence, and confidence does not indicate accuracy.
另一种看法
The pragmatics line holds that participants are not ignoring missing information but filling it in by ordinary conversational defaults: whatever the experimenter supplies is assumed relevant and sufficient. On that reading their inference is reasonable and only differs from the experimenter's norm.
分歧扎在
The disagreement is rooted in which question the participant is actually answering. The book assumes they answer the experimenter's; the other side assumes they answer one completed by context. Note this is the same class of dispute as chapter 1's substitution, with the stand-in coming from pragmatics rather than difficulty.
什么证据能裁决
What would settle it is manipulating the wording: rewrite every phrase that could be pragmatically completed (state explicitly that information is incomplete, replace 'probable' with a frequency) and see whether the effect survives. Survival under all rewrites defeats the pragmatic account; shrinking with each rewrite shows a large share comes from wording.
The evidence points to both: rewording weakens conjunction-type effects but does not flatten them, while the information-to-confidence relation is fairly stable across wordings. What is missing is an independent way to quantify how strong the pragmatic completion is.
Falsification07
Confidence tracks coherence as a preregistrable test
The mechanism under test: the coherence of a story, through standing in for evidential strength, becomes the source of confidence.
Data source: newly collected judgement data, with coherence rated blind by external coders unaware of the hypothesis
The central variable must be independent of self-report or the correlation shows nothing.
Sample: a pre-computed sample size analysed in full, with no exclusion on response time or consistency
Post hoc exclusion is the commonest degrees-of-freedom entry here.
Window: a single session, within-subject, with two information levels presented at random
Within-subject nets out individual differences in baseline confidence.
Threshold: controlling for accuracy, the partial coefficient of blind-rated coherence on confidence is significantly positive
The 'controlling for accuracy' is the point — otherwise coherence and accuracy are entangled.
Failure condition: with accuracy controlled the coherence coefficient is not significant, or inter-coder agreement is below the acceptable level
The second voids the test rather than supporting or refuting.
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 'when I feel sure, I am usually right' — confidence read as an accuracy indicator. This chapter goes straight at it.
Does this chapter (a) replace 'confidence indicates accuracy', or (b) open a new slot beside it about the coherence of stories?
慢变量Register two slow variables: how often you write a prediction down and check it, and how often you list what you do not know before deciding. Look at the first one in a few days.
小结09
本章小结
01Judgement puts coherence first: missing information is never registered as missing.
02'Less information, more confidence' is the arithmetic consequence of three assumptions, not a separate finding.
03Confidence tracks the coherence of the story, so it cannot certify its own accuracy.
04Training on conflicting evidence, and long-run calibration feedback, both break the regularity.
05Coherence has no measure independent of confidence; that part is unfalsifiable.
提取练习 · 合上书，先自己答一遍。
?Is 'evidence quantity is negatively related to coherence' (a) an identity or (b) an empirical claim?
?Does a weather forecaster whose certainty matches their hit rate weaken (a) the coherence-first mechanism or (b) the inference that confidence must be decoupled from accuracy?
Forced choice10
B · Identify the tag
'Confidence is negatively related to evidence quantity' — what kind of step is that?
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。