This governs what you are willing to call understanding. Deutsch's claim is that the point of science was never prediction but explanation, and that a machine which predicted every outcome perfectly would still leave you unable to build anything or fix anything. Four strands (quantum physics, epistemology, computation, evolution) each look narrow and mechanistic alone and only make sense held together, which is why several of the items below are positions about reality rather than moves to make, and sit under Hold.

Do

Ask what a result explains before you ask what it predicts. Deutsch's oracle thought experiment: give someone a machine that correctly predicts the outcome of any experiment they can specify, and they still cannot design a spaceship with it, or work out why the one they built exploded on takeoff. A formula for planetary motion is useless for understanding gravity unless an account of curved spacetime comes with it. [Source: "The Theory of Everything"]

Start from a problem, not from data. Astronomy produced deep theory while centuries of cloud-watching produced none, even though clouds offered far more observation. Attributing the drift of clouds to winds at different altitudes already felt like a complete account, while wandering lights in the sky left an existing explanation unsatisfying. Discovery starts where a theory you already hold looks inadequate, narrow, or in conflict with another good one. [Source: "Problem-solving"]

Explain at whatever level the explanation actually lives. A specific copper atom sits at the tip of the nose on Churchill's statue because he led Britain through the war and because it is customary to honor such leaders in bronze. Deutsch's plainer version is that the bear ate the honey because it was hungry. Reductionism and holism are both wrong: legitimate explanation exists at every level, and neither direction has priority. [Source: "The Theory of Everything"]

Apply the kick-back test when you are unsure something is real. Something counts as real if it affects you in ways complex and autonomous enough to require their own explanation, which Deutsch sharpens into a computational test: if faking its effects would demand a substantial hidden computation, positing the thing is simpler than positing the fake. A mirror reflection fails, because all its apparent complexity is borrowed from the objects in front of the mirror. [Source: "Criteria for Reality"]

Check whether the rival account secretly needs the account it claims to replace. The Inquisition's fixed-Earth theory could specify a planet's looping path only by asking how it would look if heliocentrism were true, which makes it an elaboration of the theory it was meant to defeat rather than a competitor to it. A solipsist who takes his position seriously has to postulate an unconscious part of himself larger and more law-governed than the conscious part, which is realism wearing a disguise plus extra unexplained machinery. [Source: "Criteria for Reality"]

Argue from wherever you actually are. Real arguments never run like a proof from fixed axioms to a conclusion. They start in the middle, full of gaps and inconsistencies, and proceed by criticizing and replacing even the supposed axioms as better explanations turn up. There is no bedrock of self-evident starting points, all the way down, which is why waiting for one is waiting forever. [Source: "A Conversation About Justification"]

Rely on the best surviving explanation among the rivals actually proposed. Corroboration in Popper's sense means a theory has been tested against its real, historically offered rivals, not against every rival that could logically be constructed. Deutsch's dialogue at the top of the Eiffel Tower turns on exactly this: the demand for something better before you act either regresses forever or stops arbitrarily at dogma. [Source: "A Conversation About Justification"]

Make it possible to be shown wrong, and give up on being shown right. Because virtual reality is interactive, a user can prove a rendering inaccurate with a single perceptible discrepancy, but can never certify it accurate, since infinitely many untried actions might have revealed one. Deutsch draws the parallel explicitly to scientific theories, which experiment can refute and can never establish. [Source: "Virtual Reality"]

Keep a room where rank does not count. Deutsch offers the research seminar as evidence against Kuhn's sociology of science: a graduate student can openly challenge a senior professor's equation and be answered with argument rather than an appeal to seniority, even though the same people revert to ordinary status behavior at dinner an hour later. The behavior is not automatic, it belongs to a setting somebody maintained. [Source: "The Four Strands"]

Don't

Don't read a correct prediction as support. The asymmetry is sharp: a false prediction refutes the theory behind it, while a true one confirms nothing, because bad explanations that happen to yield correct predictions are cheap and common. This is why testability rather than confirmation is what separates science from pseudoscience, and why the grass cure for colds can be dismissed without an experiment: it explains no more than its rivals while adding claims of its own. [Source: "Problem-solving"]

Don't assume the visible participants are all the participants. With a two-slit barrier, a specific point on the screen is bright. Cut a second pair of slits and that same point goes completely dark, and it still goes dark when photons arrive one at a time, as sparse as one a day, with detectors confirming the photon never splits. Something is arriving through the openings nothing visibly travelled through. When adding an unused channel changes the outcome, your inventory of what is acting is incomplete. [Source: "Shadows"]

Don't treat a conclusion as more certain than the process that produced it. A proof is not an abstract object, it is a physical process: symbols in a brain, ink on paper, a program on a machine. Its validity rests on a hypothesis about how those physical symbols behave, which is itself a physical theory and just as fallible. Mathematical knowledge is therefore no more certain than science and derivative of it. [Source: "The Nature of Mathematics"]

Don't accept a mechanical rule as a substitute for understanding. Hilbert wanted a finite, mechanically checkable set of inference rules that could prove its own consistency; Gödel showed no consistent system strong enough for ordinary arithmetic can do that. Deutsch treats this as good news, because a complete rule set would have made mathematics an exhaustible exercise in rule-following instead of something that stays open and depends on explanation. [Source: "The Nature of Mathematics"]

Don't measure importance by mass or scale. Hawking's line about humanity as chemical scum on a moderate planet is the view Deutsch is arguing against. Knowledge, not bulk, is the right measure: predicting what becomes of the Sun billions of years out already depends on choices intelligent descendants might make, and precedents like plants generating the entire oxygen atmosphere show how much mass a small amount of knowledge can move. [Source: "The Significance of Life"]

Don't claim a cause without a real alternative to check it against. In a classical block universe exactly one sequence of events happens, so a statement about what would have followed if some event had gone differently has nothing to refer to, and causation stays paradoxical. Deutsch's working definition needs the alternatives to be real: a cause is an event without which, across most of the variants, the effect would not have occurred either. [Source: "Time: The First Quantum Concept"]

Don't take an apparent paradox as proof of impossibility. The grandfather paradox only bites if you assume a traveler returns to the single history they left. Deutsch also notes the historical pattern: the expanding universe and black holes were both dismissed as unphysical solutions before being confirmed. His line for why identical copies can end up with different fates is that parallel universes carry no hidden serial numbers. [Source: "Time Travel"]

Don't spend your effort refuting what is already known to be false. Deutsch's recurring complaint: Turing spent his life arguing that brains are not supernatural instead of extending computation theory, Popper permanently refought the problem of induction, and Dawkins was cast as public defender against creationism. In each case the energy that should have gone into developing the theory went into defending it against positions already beaten. [Source: "The Four Strands"]

Hold

The four strands hold each other up, and none of them survives alone. Physics on its own reduces people to chemical scum, strict computationalism to programs running on someone else's hardware, narrow Darwinism to vehicles for replicating genes, and Popperian epistemology alone cannot say why growing knowledge is worth having. Deutsch expected physics to swallow the other three and changed his mind while writing. Holding this is what stops a single lens from producing an impoverished picture and then presenting it as rigor. [Source: "The Ends of the Universe"]

Universality is a fact about physics, not a metaphor. The Turing principle says a real, buildable machine can perform any computation any physically possible object can perform. It functions like a conservation law, cannot be derived from particle physics, and entered physics from computer science rather than the reverse. Held, the question about any capability becomes whether the knowledge exists rather than whether the substrate permits it. [Source: "Universality and the Limits of Computation"]

Almost everything well-defined is permanently out of reach, and that is not a defeat. Cantor's diagonal argument lets Deutsch construct environments that provably cannot be generated by any program on any machine, and there are infinitely many of them for every one that can. A question can have a perfectly definite answer that no physical process will ever compute or verify. Held alongside the previous item, this keeps universality from turning into a claim of omniscience. [Source: "Universality and the Limits of Computation"]

Some of the work is happening where you cannot look. When a quantum computer factors a 250-digit number by Shor's algorithm, the interfering computation runs across something on the order of 10 to the 500 universes while the visible universe holds roughly 10 to the 80 atoms. Deutsch's question is where the rest of it is being done. Refusing this leaves you calling a working machine an abstraction. [Source: "Quantum Computers"]

Nothing is moving through time. A genuinely moving present would have to move with respect to some further exterior time, which would need its own moving present, and so on without bottom. What you actually experience is the ordinary difference between present perception and present memory, correctly read as evidence of change and wrongly read as evidence that something is sweeping past. Held, there is no lost moment to chase. Not held, every moment arrives already framed as something being taken away. [Source: "Time: The First Quantum Concept"]

The long-term future is open and is chosen, not fated. Deutsch closes the book with Popper's formulation that the ends of the universe are for us to choose: bounded by physical law and nothing else, reachable through a sequence of real deadlines, and conditional only on knowledge continuing to grow. Held, the deadlines become engineering. Not held, the whole arc gets filed under things that will be decided by something other than us. [Source: "The Ends of the Universe"]

The one line

A prediction you cannot explain is not understanding, whatever else it is useful for, so refuse the black box and ask what is actually happening.