Scientific testability

What makes a claim genuinely testable rather than merely compatible with any possible result?

We praise a theory for how much it explains. Notice the assumption: that fitting the facts is evidence of truth. But consider a claim that fits every fact — markets rose because sentiment was bullish, markets fell because sentiment turned. Whatever happened, the claim was waiting for it. Did it ever risk anything? And if it risked nothing, what did the world's cooperation tell us?

So the interesting property may not be how much a claim explains, but how much it forbids.

Try measuring a claim by its prohibitions. "Something will happen tomorrow" forbids nothing and is worthless. "The star's apparent position will shift by about 1.75 arcseconds during the eclipse" forbids nearly every other number — and had the plates shown no shift, the claim would have been in serious trouble. That asymmetry is Popper's insight: a theory earns its keep by sticking its neck out, and confirmations only count when they were the sort of thing that could have gone badly wrong.

Now push on it, because this is where the tidy version breaks. Suppose the prediction fails. Have you refuted the hypothesis? Ask what you actually deployed to generate the prediction. Not the hypothesis alone — also that the instrument reads true, that the sample was uncontaminated, that no unaccounted body or field intervened, that the mathematics was applied correctly. The prediction came from the bundle. So a failure tells you the bundle contains an error. Does it tell you where? It cannot. This is the Duhem-Quine problem, and it is not a technicality; it is the ordinary condition of every experiment ever run.

Which means the falsifying blow can always be absorbed somewhere else. Is that cheating? Consider two cases from the same century. Uranus wandered from its predicted orbit; rather than abandon Newtonian gravitation, astronomers blamed an auxiliary — there must be an unseen planet. Neptune was found in 1846, roughly where the rescue said to look. Encouraged, the same move was made for Mercury's stubborn perihelion: another unseen planet, Vulcan. Nobody ever found it, and the anomaly was eventually settled by general relativity instead. Identical logical move, opposite verdicts. So what distinguished them? Not the form of the rescue — the content. One rescue made a fresh, risky, independently checkable prediction. The other, as the searches failed, had to keep shrinking and hiding its planet, buying survival with no new commitments.

Two corrections follow. First, naive falsificationism — one failed prediction, one dead theory — is not the settled view in philosophy of science, and it never described real practice; scientists live with anomalies for decades and are usually right to. Second, the criterion is not thereby empty. The surviving line is not "can this be refuted by a single result?" but something closer to Lakatos's question: when the programme repairs itself, does it generate new testable content, or merely re-describe what already happened?

So testability is best treated as a matter of degree, and as a property of a research programme over time rather than of an isolated sentence. That weakens Popper's sharp demarcation line, and it is contested — some think it dissolves the criterion, others that it makes it usable for the first time.

• Whether Kuhn's paradigms make demarcation the wrong question to ask.
• How Bayesian confirmation distributes blame across the bundle by prior probability.
• Why pre-registration is, in effect, an institutional device for restoring risk.

Every term the collection defines is gathered in the glossary.