Twenty-two lessons. One dataset: yours.

Small samples don't just hide patterns — they invent them. Your best and worst early stretches are usually the same phenomenon.

A model earns its keep only by beating "always guess your average" — scored on nights it never saw.

Differences smaller than your model's typical miss carry no information — the statistical root of tracker anxiety.

Redraw your own history and watch the estimate wobble — stability across two hundred replays, not one strong showing, is the bar. Live resampling, on your actual nights.

interactive · resample your history

Ask twenty questions of one dataset and some answers are flukes by arithmetic alone — the jelly-bean problem, and why volume of claims signals a low bar.

Prediction and causation are different jobs. Only changing one thing on purpose — the n-of-1 trial — graduates one into the other.

Your model runs like a committee with a hard budget: effects pay rent by predicting, or get exactly nothing. Regularization, felt before it's named.

When two features carry the same information, apportioning credit is a question your data cannot answer — and the model says so.

Rough days cluster. Autocorrelation gets its own slot so momentum stops masquerading as whatever you did last night.

"Is my sleep good?" can't be answered by any amount of data. A real question names its label, its features, and what being wrong would look like.

Temperature is measured; sleep stages are inferred by someone else's model. Provenance — and why raw records stay separate from derived ones.

A rating recalled a day later isn't noisy data — it's contaminated data. Why the app refuses backfill, and why honest deletions are decided by rule, in advance.

The distribution is the object; the average is one summary of it. Your histogram won't draw until you commit to a guess about its shape.

interactive · guess the shape first

Your own HRV history drawn twice — anchored at zero, then cropped into a crisis. Axis truncation deceives even when you notice it.

interactive · toggle the axis

Confounding: a shared cause dressing up as a direct link between its effects — and why "control for everything" is a fantasy.

A claim stated with its scope and uncertainty intact can be checked, argued with, and built on. Hedges aren't weakness; they're coordinates.

Two healthy instruments disagree systematically, not randomly. Calibration over shared nights — and the end of "machines are objective."

Coefficients are weights; a neuron is a weighted sum. Not an analogy — an equivalence, with your own weights under your fingers.

interactive · be the neuron

Stacked linear layers collapse back into your lasso. One bend stops the collapse — and lets networks invent their own features.

Your rating-transition table is the gentlest language model there is. Same job, different table — and no database of answers anywhere inside.

interactive · your own bigram table

Attention as context updating meaning, then count the knobs: thirteen coefficients here, hundreds of billions there. Same atoms; emergent chemistry.

Preference training, fluency mistaken for truth, and the closing argument: your model's coefficients are its whole story — no one can say that of a frontier model.