The six sessions 🎯 · before any model exists
The running situation is Daybreak, the coffee-subscription company from the sibling courses. They ran an autumn promotion, and three people want to know whether it worked. They want three different things, only two of those questions can be answered with the data that exists, and the answer to one of them reverses the moment you look properly.
The question behind the question
"Can you look at the numbers" is not a question. What turns a request into something answerable: an outcome, a population, a timeframe, and the decision waiting on the other side of it.
🧾 Session 2 · easyWhat you would need to know
Work out the ideal data before you look at the data you have, because doing it the other way round quietly rewrites the question into whatever the spreadsheet can answer.
🔍 Session 3 · easyLook before you model
The first twenty minutes with any dataset: how many rows, what is missing, what is duplicated, what is impossible. The problems that break every later step are visible here and nowhere else.
📊 Session 4 · bench nightThe four answers bench
One dataset, three questions, four levels of answer each. One headline survives every level, one reverses completely, and one question cannot be answered at all. You cannot tell which is which by looking at the headline.
➗ Session 5 · mediumWhen a count is the answer
Most questions worth asking are answered by a count, a rate or a comparison. Knowing when to stop is a skill, and reaching for a model when a count would do is the commonest expensive mistake in the job.
🧐 Session 6 · mediumReading somebody else's analysis
The questions that find the weak spot in any chart or deck put in front of you, in the order worth asking them. Final scorecard, and where to go next on the shelf.
Your passport 🎫
Clear all three quiz questions in a session and it gets stamped. Six stamps means you can take a vague request and say what would answer it. Stamps live in this browser only.
What this is not, so you land in the right place 🧭
| If your question is | Go here instead |
|---|---|
| How do I train a model, and how do I read its accuracy without being fooled? | Intro to ML ↗ |
| What is a confidence interval, a p-value, a distribution? | Statistics ↗ |
| Which chart does this question need, and is the chart honest? | Data Visualization ↗ |
| How do I design and run an A/B test properly? | Experimentation ↗ |
| How do I write the SQL to pull this? | SQL ↗ |
| Somebody asked me to look at the numbers. What am I actually being asked, and can it be answered? | you are in the right place |
This is the door into the data science shelf. It touches no model and no hypothesis test. Everything here is counting, comparing and thinking, and every session ends by naming which course to take when you want the next step.
What is honest here 🧾
The bench in session 4 computes every number from its own rows. The dataset behind it is constructed, and the pages say so, because the real-world anchor for what it demonstrates is a published case rather than this one.
| Claim | Status |
|---|---|
| Every figure in the four-answer table | Really computed from the rows: a real filter, a real group-by, a real weighted average |
| The dataset | Constructed. Written so the three questions behave differently, not sampled from a real business. Every number describes this dataset only |
| That a headline can reverse under grouping | Real, and anchored to a published case, not to the constructed one. The bench shows the mechanism; the citation shows it happening to somebody |
| How often headlines reverse in practice | Not claimed. No published figure was found, so no rate is stated anywhere |
| The verdict line | A heuristic, badged as one on the widget |
| Daybreak | A teaching situation. The arithmetic is real |
The knowledge map 🧠
The whole course at a glance - hover a session to spotlight its concepts, click any node to jump in.