The six
1. Strike-through
Take the paragraph below and mark every clause that is a conclusion rather than a recording. Then rewrite the ones you can, and mark the ones you cannot recover as unrecoverable.
He was nervous the whole time. His body language screamed insecurity, and when the partner walked in he got visibly intimidated — you could see him shrink. He was trying way too hard to seem confident, which just made it worse. The room could feel it.
Checking your work. You should have found seven or more. If you found three, you are still reading visibly, you could see, and the room could feel it as evidence rather than as rhetorical insistence — those three phrases are the tell that a writer knows they are asserting and is bracing the assertion instead of supporting it. Every one of your rewrites must pass the camera test: could a camera, with no knowledge of anyone involved, have recorded it. He shrank fails. He moved his chair back about a foot when the door opened passes, and only if that is what happened.
2. Sixty seconds, ten lines
One minute, one real human being — a café, a train, a recorded interview if you would rather not stare. Ten separate lines of description. Timer on. No line may contain a state word.
Checking your work. The metric is not quality, it is quota. If you wrote six good lines instead of ten, you failed the exercise, and the reason is worth sitting with: to get six good ones you were filtering for significance, and significance-filtering is interpretation running upstream of the description, choosing what gets described. Ten forces you past the interesting details into the boring ones. The boring ones are where calibration lives, because they are the ones you did not already have a theory about. Second check: at least three lines should be about timing or change — onset, duration, rhythm, what moved and in what order. A ten-line still life means you were describing a photograph of a person rather than a person.
3. Your own chest
Inward, now. Twice a day for six days, catch yourself mid-something and describe the sensation without naming what it is. Six axes, and you must hit at least four: location (where, precisely, and does it have edges), extent (how big, and does the boundary hold still), quality (pressure, buzz, heat, tug, hollow, grain), intensity, movement (still, spreading, pulsing, travelling — in what direction), timing (when it began, whether it arrived fast or accrued, how it relates to what just happened).
Checking your work. Use the transfer test, since the camera cannot help you here: could you read your entry to someone who does not share your emotional vocabulary and have them locate roughly the same region and quality in their own body. Anxious fails. A band across the top of the ribs, thumb-width, tighter on the left, arrived over about ten seconds after the phone buzzed passes.
And a floor you should know about: interoceptive accuracy varies enormously between people, and it is not a moral quality. If you look inward and find nothing locatable, write down that you found nothing. An honest null is data. A manufactured sensation, invented to fill the box, is worse than a blank — it teaches you to trust a signal you fabricated, and everything downstream of it in this book will steer on that fabrication.
4. Twin predictions
Return to the ten reads you have made about people this week — you will have to reconstruct them, which is itself the lesson. For each, write the forecast if true, the forecast if false, and the single observable that separates them. Then go and find out.
Checking your work. The discriminating observable must be something that could plausibly have gone either way. If your if false column contains things that would never happen regardless — she would have laughed and hugged me — you have written an unfalsifiable read and dressed it as a test. The sharper diagnostic: who introduces the topic. Predictions that hinge on what the other person raises unprompted are real tests. Predictions that hinge on how they respond to your prompt are contaminated, because you supplied the stimulus and will read the answer through the same lens that generated the question.
Expect to be right about half the time. If you are scoring above eighty percent, your predictions are too safe to be worth making.
5. The body you already have a theory about
Someone you love, or someone who reliably irritates you. Pick the one where you would say you know them. Fifteen minutes of pure description, no state words, written during or immediately after. Then — and this is the exercise — have a second person who was also present describe the same fifteen minutes, independently, before either of you reads the other.
Checking your work. You are looking for the places you disagree, and specifically for disagreements at the level of description, not interpretation. Two people disagreeing about whether he was angry is uninteresting. Two people disagreeing about whether he stood up before or after the sentence about the money is enormously interesting, because one of you has back-filled the sequence to fit a story, and sequence is the load-bearing evidence for every causal claim you will want to make later.
The edge on this one, and you should feel it while you work: description used on an intimate is not neutral. There is a mode of this practice that curdles into surveillance — coding a partner's jaw while they talk to you, having a clinical relationship with someone who thinks they are having a human one. The tell is that you have stopped responding. A body across from you registers being coded with high accuracy and will read it, correctly, as withdrawal. Refusing to conclude is not the absence of a stance; it is a stance, and one that lands as coldness. The discipline belongs in the notebook and in the half-second before you speak. It does not belong in your eyes for an hour.
This is also the edge at scale. The TSA's behaviour-detection programme trained officers to spot deception from bodily indicators, and the Government Accountability Office's review of it found no sound evidence the indicators worked — while the officers using them remained entirely confident. A whole apparatus of interpretation, operating with the felt certainty of observation, on people who could not object. Confidence is not calibration. It never was, and at institutional scale the gap becomes policy.
6. The log
Not finishable today. Say six weeks, and the shape of it is this.
Every read you make about a person's state, logged as it happens: timestamp, person, the read, the discriminating observable, and — left blank — the outcome. You fill the outcome column later, from evidence, not from memory of how it felt. Then, and only once you have enough entries, you score.
The scoring is where it stops being a journaling exercise and becomes an instrument. Three columns you compute at the end:
Hit rate. Straightforward, and nearly meaningless alone.
Base rate. How often was the read true regardless of what you observed. If you called Rana tired eleven times and were right nine, but she is tired seventy percent of the time, your nine-from-eleven is worth almost nothing — you were reading the population, not the person. Skill is the distance between your hit rate and the base rate. Most people have never computed this about themselves and are startled by the result, because a great deal of what passes for reading people is confident recitation of what is usually true.
Slice by person. This is the finding that takes the six weeks. Your accuracy is not a trait you have; it is a relationship you have with each specific body. You will find you read one colleague at genuine skill and another at pure base rate, and the second one is almost never the person you would have guessed. It is often someone you are close to — because closeness gives you a dense, confident, and old model, and an old model stops sampling.
Checking your work. You will know the log is honest if it hurts somewhere. Specifically: if the outcome column was filled in from evidence rather than recollection, you will have at least one entry where you were confident and wrong about someone you would have sworn you knew. If six weeks produced no such entry, the outcome column was scored by the same faculty that made the prediction, and you have built a very careful machine for confirming yourself.
Forty entries before the per-person slices mean anything. Sixty before you would bet on them. There is no version of this that goes faster, and the reason is not that you are slow — it is that you are estimating a rate, and rates need samples. Start it today. Do not expect it to tell you anything for a month.
Ekman and Friesen's Facial Action Coding System is the fullest working version of this discipline anyone has built: a scheme that codes lip corner puller and inner brow raiser as numbered units and pointedly does not code happy or worried, so that the measurement stays separable from the inference. It was built that way on purpose, and the purpose was vindicated — because when later researchers challenged the field, the challenge landed on the inference layer, on whether facial configurations map onto emotion categories at all reliably across people and contexts. The coding layer survived intact. That is what a clean split buys you: when your theory turns out to be wrong, your observations remain worth having.
Which is the whole case for the week, and it is not the case you were expecting. A description is not more true than a conclusion. Often it is less useful — you cannot act on thirty degrees away from me. What a description is, is durable. It holds its shape while your theory about it collapses and gets rebuilt, and it is still there afterward, still worth something, ready for the better theory.
Every pattern in the second half of this book is a control loop, and a control loop needs an error signal. The conclusion is the prediction. The description is the measurement. Collapse them into one thing — which is what your perceptual system does for free, all day, at no felt cost — and you have a system that outputs confidence and receives no correction, forever.