The Good Explanation Test
Judge any claim on four tests: explanatory, testable, hard to vary, risky predictions.
- Difficulty
- Moderate
- Time to result
- ~weeks to results
- Steps
- 7
- Confidence
- 92%
Deutsch's upgrade to the scientific method treats a good explanation as something you can score, not just accept or reject. First, it must be an explanation: a creative account of the underlying mechanism, not a summary of what was observed. Second, it must be testable — some experiment in the world could show it false. Third, and most importantly, it must be hard to vary: every component has to be doing load-bearing work, so that swapping a piece destroys the predictions. Persephone leaving Hades explains the seasons but Persephone could be swapped for any other god without changing anything; a 23-degree axial tilt cannot be varied at all without breaking its predictions. Fourth, its predictions should be narrow, precise and risky. Apply the four tests in order and most confident claims collapse at the third.
Origin
Naval Ravikant and Brett Hall unpack David Deutsch's criteria for a good explanation, an extension of Karl Popper's falsifiability, drawn from The Beginning of Infinity and The Fabric of Reality.
Core principles
- 01A theory that is merely testable tells you nothing about its quality.
- 02The best explanations are the ones you cannot patch without wrecking them.
- 03Explanation comes first; prediction is a by-product, not the point.
- 04Certainty is a feeling, not evidence — the strongest feeling of certainty has been wrong before.
- 05No explanation is final; the best one is only the best one for now.
How to run it
- 1
Demand an explanation, not a correlation
Ask what mechanism is being proposed and why it produces the observed effect. A claim that only reports a pattern is not yet an explanation of anything.
Pro tip Explanations are creative leaps, so expect the good one to look less obvious than the naive reading of the data — the sun appearing to move is explained by the earth rotating.
- 2
Find the test that could kill it
Identify a concrete experiment or observation whose outcome would show the claim false. If you cannot name one, stop here.
Watch out Testability alone is a very low bar — falsifiable theories are a dime a dozen and passing this step means almost nothing on its own.
- 3
Run the hard-to-vary check
Take each component of the explanation and try replacing it with something else. If the predictions survive the substitution, the component was decorative and the explanation is weak.
Pro tip Do this before you look at any evidence — it filters out most stories without needing a single experiment.
Watch out An explanation with interchangeable parts can absorb any result, which is exactly why it never makes progress.
- 4
Check the predictions are risky and narrow
Good explanations stick their neck out: exact lengths of seasons at given latitudes, a specific amount of starlight bending during an eclipse. Vague or safe predictions signal a weak explanation.
Pro tip The riskier the prediction that survives, the more you have learned from the test.
- 5
Refuse post-hoc rescue
If the test fails, you may not quietly adjust the numbers or conditions to save the theory. Moving from one kilogram of grass to 1.1 kilograms is not a correction, it is an admission the explanation was empty.
Watch out Watch for this move in your own thinking — it feels like refinement and is actually retreat.
- 6
Only refute against a rival
A disagreeing test damages a theory but does not remove it unless a better explanation is available to move to. Where there is no alternative, suspect the test before the theory.
Pro tip The crucial test is the case where two viable rival explanations make different predictions and one experiment separates them.
- 7
Hold the winner provisionally
Treat the surviving explanation as the best available for now, not as settled truth. Keep looking for the error that will produce the next, better explanation.
In the wild
Someone tells you that eating 1.0 kilograms of grass cures the common cold. The claim is perfectly testable, so a naive falsifiability filter waves it through. But no mechanism is offered for how grass would act on a cold, and when the test fails the claimant can say the dose should have been 1.1 kilograms, or a different kind of grass, or a different day. Every part of the theory is freely variable, so no experiment can ever move it forward. The four tests reject it at step three without anyone having to eat any grass.
→ A testable claim is correctly discarded as worthless because it has no mechanism and is trivially easy to vary.
The ancient Greek account of the seasons has Persephone permitted to leave Hades for part of the year. Persephone could be swapped for Nike and Hades for Zeus and the story would predict exactly the same seasons — every element is interchangeable. The axial tilt explanation says the earth is angled at roughly 23 degrees relative to its orbit, and from that single hard-to-vary fact you get the precise length of summer and winter at any latitude and where on the horizon the sun rises in each season. Change the angle and every prediction breaks.
→ Two theories that both fit the observed seasons are separated cleanly by the hard-to-vary criterion.
Einstein's general relativity predicted that starlight would bend by a specific amount as it passed the sun. That was a risky, narrow prediction that took years to confirm. When the eclipse measurement came in, the correct reading was not that relativity had been proven finally true, but that Newton's theory of universal gravitation had been ruled out because it was inconsistent with the result. Two viable rival theories existed, one experiment separated them, and science moved to the better explanation while leaving the door open to replacing it later.
→ A single experiment eliminated a rival theory rather than establishing a final truth, which is what a crucial test actually does.
Common mistakes
Treating testability as the whole test
Falsifiable theories are abundant and cheap; the criterion that does the real filtering is whether the explanation is hard to vary. Stopping at 'is it testable' lets almost anything through.
Patching a theory after it fails
Adjusting the dose, the conditions or the definitions after a disconfirming result makes the theory unfalsifiable in practice. The adjustment is evidence the explanation had no load-bearing parts.
Discarding a theory with nothing to replace it
If a test conflicts with the only theory you have, you have nowhere to jump to — and historically the faulty element has usually been the test. Refutation is a move between rival explanations, not an act performed on one theory in isolation.
Is it for you?
Best for
Anyone who has to judge competing claims — investors, founders, researchers, or readers trying to tell real explanation from confident storytelling.
Not ideal for
Situations that need an immediate decision under time pressure with no room to interrogate the underlying mechanism.
From the transcript
“a good explanation first and foremost is testifiable falsifiable you can run some experiment in the real world to see if it's true or not”
“so i think the second piece of good explanation is hard to vary it has to be very precise and there's a good reason for…”
“finally the predictions that it makes should be very narrow and precise and they should be risky”
From the episode
The Beginning of Infinity, Part 1