Appearance
10.11 — Building an Accurate Model of One Person
Everything in this Part so far has been about the general case, and the general case is where accuracy is poor.
This page is about the case where accuracy can be excellent: one specific person, observed over time.
The difference is not subtle. Your ability to tell whether a stranger is upset is barely above chance. Your ability to tell whether your sister is upset, from three words on the phone, is close to perfect. Same skill, different amount of data.
What a model of a person actually is
Not a personality label (9.3). A set of specific expectations, each of which could turn out to be wrong.
Their baseline. How they normally are. Talkative or quiet, warm or reserved, fast or considered. This is the reference point that makes every other reading possible (10.3).
Their tells. What their tired looks like, what their angry looks like, what their genuinely-fine looks like as opposed to their saying-fine. Learned by observing, and by later finding out you were right or wrong.
Their triggers. Which of the six threats from 5.2 is theirs. Almost everyone has one or two, and knowing somebody's is one of the most useful things you can know about them.
Their default under stress. Withdraw, attack, appease, take control, go quiet, make a joke. Highly consistent within a person (4.5), and predictable once identified.
What they need when upset. Listening, ideas, or practical help (4.7) — and this is remarkably stable per person, so you only have to learn it once.
Their values (9.7). What they will not trade.
And their reliability profile, by domain. Reliable with commitments and unreliable with time. Excellent in a crisis and hopeless with admin. Specific, not global (10.8).
How to build it
Four inputs, and the fourth is the one everybody skips.
Observe over varied conditions
A model built only from good conditions is not a model. You need to have seen them tired, under pressure, wrong about something, disappointed, and dealing with someone difficult.
Which means it takes time, and it means that after six months of pleasant interactions you have a model of how they are when things are fine, which is the least informative condition.
Notice the change points
From 10.9. What is different from their normal, and when did it start. The absolute behaviour carries little; the deviation carries the information.
Ask, directly, more than you think is normal
The most efficient input and the most neglected.
People will tell you things about themselves if you ask plainly and do not make it strange.
"What do you actually want when you're upset — do you want me to listen or to help?""What's the thing I do that annoys you most?""How do you like to be told when something's gone wrong?""What did you make of that meeting?"
These questions feel unusual to ask and are almost always welcome, because most people have never been asked and have an answer ready.
Get feedback on your predictions
The step that turns observation into skill (3.5), and the reason most people's models never improve despite decades of data.
Predict, then check. I think she'll be annoyed about the change of plan. Then find out. Not by inferring from her behaviour — by asking.
Without this you accumulate confidence rather than accuracy (10.1). Twenty checked predictions about one person will teach you more than five years of unchecked observation.
Why models go stale
The failure mode that causes the most damage in long relationships.
Once a model is good enough, you stop updating it. You already know how they are, so you predict rather than observe, and you attend to confirming evidence (1.4).
Meanwhile the person changes (9.1). Over ten years, substantially.
The result is a couple who know each other perfectly and are each responding to a version from six years ago. You always hate those things — said about something she began to enjoy in 2021.
Three signs your model has gone stale:
You are surprised by them and you dismiss the surprise as out of character rather than as data.
You finish their sentences and you are sometimes wrong, and you do not notice being wrong.
And you have stopped asking them questions you do not know the answer to. This is the reliable one, and it is measurable: when did you last ask this person something and genuinely not know what they would say?
How to keep a model current
Ask one question a month you do not know the answer to. Not about logistics. What are you actually enjoying at the moment? What's been hardest this year? Has anything changed about what you want?
Treat surprises as updates rather than exceptions. When somebody does something out of character, the honest first hypothesis is that your model is out of date, not that they were being unlike themselves.
Revisit the model deliberately once a year for the three or four people who matter most. What has changed about them in the last year? Most people cannot answer, which is itself the finding.
And notice the drift in yourself, because they are running a stale model of you too, and telling them what has changed is a kindness that saves both of you a category of misunderstanding.
The specific thing worth building for the people closest to you
Five items. Written down, and you will find you do not know two of them.
One: their trigger. Which of the six threats from 5.2.
Two: their default under stress. Pursue or withdraw (9.5), attack or appease.
Three: what they need when upset. Listening, ideas, or help. Asked, not guessed.
Four: what makes them feel valued. What lands, specifically — words, help, time, gifts, physical affection. Not what you would want; what they respond to (12.4).
Five: what they are carrying at the moment (10.9). What is going on in their life right now that you would need to know to interpret their behaviour this month.
The fifth one expires and needs replacing. Most people's model of their spouse contains items one to four, updated in 2018, and nothing for item five.
The limits
A good model predicts tendencies, not moments (9.1). Even with twenty years of data, you cannot say what they will do today.
And it can become a cage. Telling somebody who they are is one of the most limiting things you can do to a person, particularly a child or a partner. A model is for anticipating and helping — never for informing them of what they will do or feel. You're going to hate this is a prediction that also removes the possibility of being wrong.
Hold the model as your current best guess about a person who is still changing, and say so occasionally. The sentence "I might be out of date about this — is that still true of you?" is worth having in a long relationship, and almost nobody says it.
What to do with this
Write the five items for the two or three people who matter most. You will get stuck on at least one, and the sticking point is the useful part.
Then ask them. Directly, in a calm hour, without making it an occasion.
Start predicting and checking. One prediction a week about one person, checked by asking. Twenty of these will change how accurate you are, permanently.
Ask one question a month you do not know the answer to.
And update item five constantly, because whatever they are carrying right now explains more of their behaviour this month than everything else in the model combined.
Next: Part 11 widens from individuals to groups — why people conform, how situations overpower character, and the levers of influence, seen from both ends.