Chapter 6: The Physics of Persistence

In the winter of 1860, before a room of restless children at the Royal Institution, Michael Faraday held up a lit candle and told them there was no better door into the whole of natural philosophy than the physical phenomena of a burning candle. He spent six lectures on it. The flame, he showed them, is not a thing but a transaction: wax drawn up the wick by capillary action, vaporized in the heat, torn apart and recombined with the oxygen of the air, the light itself only the visible receipt of the reaction. What the eye reads as a small stable object is a region of space through which matter is pouring — nothing in the flame at the end of this sentence is what was in it at the beginning. Every particle has been consumed and carried off and replaced. The shape alone remains.

A flame, in other words, is a form maintained through the total and continuous turnover of its own matter — durability purchased by reconstitution rather than by permanence — and it keeps its shape for exactly as long as the loop that feeds it runs, and not one instant longer. Pinch the wick and the form is gone, though not a single atom has been destroyed; the atoms have merely stopped being organized. This is the picture of persistence the present chapter needs, and it is worth pausing on because it is the exact inverse of the one this book began with. Plato’s Form persists by standing outside time and requiring nothing. Faraday’s flame persists by standing inside a process and requiring everything, and it is the second picture, not the first, that the last two chapters have forced upon us.

Those chapters made their case from wreckage. Forty years of institutional experience, read as a single natural experiment, returned one verdict: representations behave like the flame and not like the Form. They persist only while some loop keeps re-instantiating them, and the loop is enforcement — divergence made to fail visibly, attributably, and early. That was an empirical finding, a pattern read off the graveyard and compressed into a causal claim, and we conceded at the close of Chapter 5 that a fair skeptic could accept every autopsy and still call the thesis a heuristic — true of software, perhaps, for reasons peculiar to software. This chapter answers the skeptic. It shows that the same criterion falls out of physics, reached by people who had never seen an enterprise ontology and were trying to repair something else entirely. When a claim read from forty years of failure turns out to coincide with a claim derived from the foundations of physics, the coincidence is itself the evidence, and supplying it is what this chapter is for.

I. The Constructor

The physics in question is recent and deliberately strange. In 2012 David Deutsch proposed reorganizing the ambition of fundamental theory around a different kind of statement. The usual form of a physical law is a trajectory: give me the initial conditions and the equations of motion, and I will tell you what happens next. Deutsch’s proposal — constructor theory — expresses laws instead as statements about which transformations are possible, which are impossible, and why. Not what will happen, but what can be made to happen, and what no arrangement of matter could ever be made to do.

The theory takes its name from the constructor, defined as anything that can cause a transformation and then remain able to cause it again: a catalyst that emerges from a reaction unchanged, a robotic arm that stamps one part and turns to the next, a factory, an enzyme. But the constructor is the theory’s namesake, not its subject. Having defined it, Deutsch abstracts it away, and what remains as the basic furniture of the world is the set of tasks — the transformations themselves, sorted into the possible and the impossible. A task is possible, in this vocabulary, precisely when a constructor for it could exist: when there could be something that performs the transformation reliably and, crucially, retains the capacity to perform it once more. Possibility, in constructor theory, means repeatable capability.

It takes a moment to feel how large this reframing is, because it lifts the fundamental question about any object off the object and sets it on what can be done with the object. The old question — what is this made of, what state is it in — is displaced by a new one: which transformations does its existence make possible, and which does it forbid? A thing is characterized not by its substance but by its repertoire. Readers of the first three chapters will hear the echo, because this is the counterfactual restatement of a move Shannon made by hand in 1948, when he defined information by what could have been sent rather than by what was. Deutsch generalizes the deferral into a physics: what a thing is matters less, at the deepest level, than what could repeatedly be done to and with it. And done repeatedly is the phrase to hold onto, because a capability that can be exercised only once is not a constructor at all. Every constructor is, by its definition, a loop.

II. Knowledge That Keeps Itself

Inside the 2012 paper is a definition easy to read past and central to everything that follows. Among all the abstract things that can play the role of a constructor, one kind is singled out and given a name borrowed, pointedly, from Karl Popper and stripped of any knowing subject: knowledge. Knowledge, in Deutsch’s sense, is information which, once physically instantiated in a suitable environment, tends to cause itself to remain so instantiated — information that survives noise, error, criticism, and the wholesale replacement of the matter that carries it, because it acts, through its environment, to get itself re-instantiated. The canonical case is the gene, which causes the organism around it to build fresh copies of the very sequence that specifies the organism. There is no knowing subject anywhere in the definition. There is only a pattern that participates in its own persistence.

This is the flame made rigorous and given a mechanism, and it quietly answers a question the earlier chapters left hanging. Chapter 3 named λ, the effective decay rate of a represented piece of knowledge, called it the quantity on which the whole economics of representation turns, and then confessed that the model was silent on what fixed it. Deutsch’s definition supplies the missing principle, if not the number. What sets λ is not any property the representation has in isolation — not its expressiveness, its formality, or the polish of its authorship — but whether the representation sits inside a loop that causes it to be re-instantiated. Knowledge, in Deutsch’s sense, simply is information whose decay rate has been driven low by such a loop; everything without one decays at the ambient rate of its surroundings. A representation does not endure because it is well made. It endures because something keeps remaking it, and it would decay like anything else the instant that something stopped.

And the loop is never free — a point to carry forward, because it is exactly where the natural case and the organizational case will part. A self-instantiating pattern is not a pattern exempt from cost; it is a pattern that something is continuously paying to maintain. The flame keeps its shape only while the wax keeps feeding it. The gene — Deutsch’s own example of knowledge — endures only while the organism spends, generation upon generation, the energy to copy it and copy it again. Nothing in the account makes persistence cheap. It makes persistence conditional: on the existence of a loop, and on that loop being paid for. What will differ, when we turn from candles and genes to organizations, is not whether the loop costs something. It is who decides whether to keep paying.

III. The Ship That Athens Kept

Deutsch completes the thought with the oldest thought-experiment in the Western catalogue, and gives it a reading this book will use twice. The ship of Theseus, preserved by the Athenians, has its planks replaced one by one as they rot, until no original timber remains; the traditional puzzle is whether it is still the same ship. Deutsch’s move is to stop asking about the timber and ask what stays constant across the whole process. Not the planks, which are all replaced. Not the shipwrights, who die and are succeeded. Not even the written specification, which is recopied as its parchment decays. The one invariant across the centuries is the abstract program that keeps causing Athens to re-instantiate the ship, the specification, and the shipwrights alike. The information is the constructor. Durability, on this reading, is never resistance to change; it is a causal loop that reproduces a pattern through change, and Plato’s eternity has been quietly replaced by a fixed point of maintenance.

But notice what kind of loop keeps the ship, because it is not the kind that keeps the gene. The gene’s loop runs on chemistry and differential survival; it convenes no committee. The ship’s loop runs on Athens — on a community that had decided the ship was worth keeping, that funded its repair, trained its shipwrights, and would notice a rotten plank and replace it. The ship of Theseus is the first maintained pattern in this chapter whose loop is institutional rather than natural, and that single fact makes it the ancestor of the central practical test of Part III. We will call it the model-swap test, and Chapter 8 will make it operational, but its logic is already visible in the harbor at Athens: take any representation an organization believes it holds, and swap out its carriers one at a time — the application, the vendor, the foundation model, the people who understood it — and ask whether the meaning survives the swap. For the ship it survived, because the specification was load-bearing: the shipwrights consulted it, and a plank cut against it failed, visibly, to fit. For most of what an organization “knows,” the meaning does not survive, because nothing consults the specification and nothing fails when it is wrong. The swap simply loses the knowledge, silently, the way a flame goes out when the wax runs dry — no drama and no alarm, just a shape that was there and then is not.

IV. Escaping the Circle

The strongest form of the argument arrived in 2014, when Deutsch, now working with Chiara Marletto, brought information itself into constructor theory and in doing so repaired a flaw at the foundation of Shannon’s edifice that had gone unremarked for sixty years. The flaw is a circularity. Information is standardly defined in terms of distinguishable states; distinguishability is defined in terms of what a measurement lets you learn; and learning is the acquisition of information. Each concept leans on the next, and the last leans back on the first. Shannon’s theory, magnificent at counting distinctions, never says what the act of distinguishing physically consists of — it takes the distinguishable states as given and builds upward. For an engineering theory of transmission that was a harmless economy. For a would-be physical theory of what information is, it is a hole at the center.

Deutsch and Marletto’s repair is to ground the whole structure in tasks, in the one order that avoids the circle. Computation is defined first, as the set of possible permutations over a substrate’s physical attributes — no mention of information yet. Information variables are then defined as those computation variables that can be copied, that is, cloned from one system onto another. Distinguishability arrives last, downstream of both, and non-circular because the concepts it rests on were fixed before it. The result is a definition of information as a physical property constituted by which copying and distinguishing tasks are possible on its carrier. To carry information is to be such that certain transformations — copyings, distinguishings — can reliably be performed. Take away every task that could be performed on a thing, and you have taken away its status as information. What could be done with it is not evidence of the information; it is the information.

Now translate that criterion into the vocabulary of Chapter 5 and watch what happens. A representation carries meaning only insofar as some task can actually be performed on it — some check run, some copy made, some distinction drawn — and it persists only insofar as the task that reproduces it remains possible and is in fact performed. Strip away every process that reads it, checks it, and re-instantiates it, and you have not merely neglected the representation; on this account you have removed the thing that made it a representation at all. That is the load-bearing thesis, stated in the language of physics, and it was reached from the bare requirement that the concept of information be non-circular — reached by two physicists with no stake whatever in canonical data models, change-control boards, or the fate of the Semantic Web.

This is the convergence the skeptic was owed. Chapter 4 read the criterion off the headstones: the artifacts that survived were the ones something checked, and expressiveness, formality, and funding sorted the record not at all. This chapter derives the same criterion from first principles, for reasons internal to physics. The two arguments share no premises — one begins in a graveyard of enterprise software, the other in the axioms of a candidate fundamental theory — and they arrive at the same line: what is real and durable about a representation is what can be done with it, which is to say what checks it. Convergence of independent arguments is not proof; organizations are under no obligation to obey constructor theory, and this book does not pretend they are. But it retires a particular suspicion. The load-bearing thesis is not a parochial lesson of one young industry, a rule of thumb that happens to hold in software. It is the local, institutional face of something general about how any pattern persists in a universe that offers nothing for free.

V. Same Law, Opposite Side

We can now say precisely what was meant, at the end of Chapter 1, by calling this an engineering essay written under a physical law. The law is this: a pattern persists when, and only when, some loop re-instantiates it at least as fast as it decays, and no loop of any speed can preserve a pattern whose decay rate exceeds the loop’s corrective capacity. Genes obey it, flames obey it, the ship of Theseus obeys it, and a database schema obeys it. It is substrate-independent in the strong sense — the same indifference to medium that Chapter 2 found beneath the bit, now applied not to the encoding of a pattern but to its survival. Persistence is not a property of a thing. It is a property of a thing’s relation to a loop, exactly as load-bearing, in Chapter 5, turned out to be a property of a wall’s relation to the structure rather than of the wall.

What differs between the gene and the schema is not the law. It is the provenance of the loop, and the whole remaining argument of this book lives in that difference. Nature’s maintenance loops come installed with the laws. Differential survival maintains the gene whether or not any creature intends it, because the lineage whose copies drift too far is simply not among those that continue; the loop runs whether or not a budget was approved, because there was no meeting at which the budget could have been denied. The loop is found, not chosen. Selection discovers it, installs it, and pays for it out of the deaths of everything that lacked it, and the organism inherits a maintenance apparatus it never had to design and cannot decline.

Organizations inherit nothing of the kind. Their corrective loops — the build that breaks on a bad schema, the validation that bounces a malformed message, the evaluation that refuses to let an agent act on a stale rule — come with no law at all. Each one must be designed by some particular person, funded against everything else the money could buy, staffed by people who could be doing more visibly rewarded work, and defended, every planning cycle, against the entirely reasonable question of why the effort cannot be spent elsewhere. Nature’s maintenance is subsidized by extinction; an organization’s is subsidized by a line item, and a line item is negotiable in a way that extinction is not. This is the asymmetry — the same law, approached from opposite sides. Physics gets to describe loops that already exist and were never in doubt. Organizations face the prior problem: bringing the loop into existence, and keeping it funded while every incentive whispers that the money is needed elsewhere and the loop, having quietly prevented the failures that would have justified it, looks like a cost with nothing to show.

That asymmetry is the whole reason this is an engineering book and not a work of physics. The physicists have told us where self-maintaining information is found; the graveyard has told us it is almost never found in an organization by accident; and the two findings, set side by side, define the task exactly. If an organization’s understanding of itself is to survive its authors, its applications, and its models — if it is to be a pattern kept rather than a document aging — then someone must build the loop that re-instantiates it, and someone must keep paying, in attention and in money, for that loop to run. Nature will not supply it. It never has.

What remains is to design the loop so that building it is cheap and skipping it is hard — to give an organization’s meaning the one property the graveyard’s casualties all lacked and its survivors all, by luck or by instrumentation, possessed. That requires an artifact: something an organization’s knowledge compiles into, structured so that the processes which consume it fail visibly when it drifts, and structured so that leaving it unenforced is not the path of least resistance but a deliberate and conspicuous choice. Part III builds that artifact. We turn first to its anatomy — its parts, its properties, and the one component, absent from everything in the graveyard, that makes it structurally difficult to leave unchecked. It has a name already, borrowed as an homage and understood in hindsight, and we can at last say what it is made of.