The observation is identical under all three. On its own it discriminates between none of them.

And this is n = 1. One episode cannot separate these stories. Science is the problem of finding comparisons that can.

We cannot run both worlds for the same Sofia. The causal effect is a difference we can never directly observe.

Causal effect = what happened minus what would have happened otherwise. Everything that follows is about finding a credible stand-in for that second line.

Each time we saw a pattern — a constant conjunction, in Hume’s phrase. The word causes is an inference — something we added.

Consequence: the pattern is real; the story we tell about it might not be.

Summer causes both: heat drives ice cream sales and swimming.

The same shape, in claims you might actually believe:

• “Countries with more McDonald’s have longer life expectancy.” — wealth produces both.
• “People who watch TV news vote more often.” — age produces both.
• “People who take vitamins live longer.” — the kind of person who takes vitamins also exercises and sees a doctor.

Test: ask “what third factor could produce both?”

Data works the same way:

• US cheese consumption tracks deaths by bedsheet entanglement.
• Nicolas Cage films per year track swimming-pool drownings.

Nobody noticed these. A computer trawled thousands of datasets and printed the pairs that happened to line up.

One country, fourteen years. The two lines move together almost perfectly.

Nobody thinks chocolate causes weddings. So where did this chart come from?

That chart was Country 13. The same two variables, in nineteen other countries, show nothing at all.

Look in enough places and one of them will line up. Test: ask “how many were compared before this one was shown to me?”

therefore — thus — this shows — this proves — this rules out — this is inconsistent with — this confirms

These words assert that the evidence supports the conclusion. They do not demonstrate it.

You already stop at “therefore”: conclusion? premises? unstated assumption? With evidence, that third question has a standing answer — “and no rival explanation fits.” The audit is that check, grown.

The Inference Audit: every time you meet one of these words, stop and test the logic. Six questions, one rule.

• Story 1: the pill worked.
• Story 2: the headache would have gone anyway.

The evidence fits both. On its own it cannot pick one.

Good evidence does not merely fit an explanation — it helps us choose between competing explanations. That is the whole job of the six questions that follow.

Both stories predict exactly what he saw. His evidence cannot choose between them.

The comparison that could: students who ate the chicken vs. those who ate there and did not.

He never asked what the other story would predict — confirmation bias, dressed as data analysis.

A student scores 50, then 80, then 90. The first jump is +30. The second is +10.

The mistake: “the new teaching method has stopped working.”

Slower improvement is not a weaker cause. Sometimes there is simply less left to gain.

Sofia reached for the pill when the headache was at its worst. Headaches at their worst tend to ease, whatever you do.

The mistake: “the pill worked.”

Measure anything at its extreme, and the next measurement will usually be less extreme. That looks exactly like a treatment working.

Students who sit in the front row get better grades.

The mistake: “sitting at the front raises your grade.”

Ask who ended up in the group, and whether they were already different. Nobody assigned those seats.

The mistake: “standards are slipping.”

Not one student got worse. The average fell because ten new students joined — before reading a moving average, ask whether it is still the same group.

Smartphone ownership rose steadily from 2000 to 2025. So did house prices.

The mistake: “phones drove up housing costs.”

Before reading a trend as a link, ask what else was moving the same way over the same years. Almost everything was.

So everything the course has taught about arguments — plus the Inference Audit — applies. Plus four questions for studies:

Who? What? Compared to what? How was it selected?

Ask: who is in the sample, who is missing, and does the conclusion quietly extend beyond the people actually studied?

The test: if the measure moved, would the thing we actually care about necessarily have moved?

Without a credible comparison, “after” is not “because of.”

A coin has no opinion about who deserves the treatment, so the two groups end up alike — including on the things you never thought to measure.

Randomisation is not a statistical nicety. It is the closest we get to the row we can never observe.

The two that found something get published. The other eighteen go in a drawer.

You read the literature, see two positive results, and conclude the effect is real. Nobody lied. Nobody even did anything unusual.

The question to ask: how many analyses were tried before this one appeared? Bias does not require fraud — the filter between what gets run and what gets read is simply not neutral.

And funding? A conflict of interest tells you to look harder at the design and the reporting — never on its own that the finding is false.

A researcher tests twenty outcomes. Then a few subgroups. Then a few statistical models.

Sooner or later one of them looks impressive. Every individual step felt reasonable.

The result is real in this dataset — which is not the same as evidence about the world.

This has a name: p-hacking — trying many analyses and reporting the ones that came out well. It is Part 2’s chance explanation, inside a single study.

Every answer has the same shape: name the rival, then name the comparison that would tell them apart.

Some claims have several rivals at once. That is normal — the question is never which fallacy is this, it is whether the evidence can choose.