Chapter Seventeen The Ground Leads Somewhere

The ground held under every trial brought against it. Here is the proof that it goes somewhere, and the difference is the difference between a floor and a road.

Everything won so far has been, in a sense, defensive. Induction is not groundless; the sceptic defeats himself; the paradoxes were under-specified questions. Good verdicts, all of them, but each one clears the framework of a charge rather than showing what it can build, and a reader would be within rights to ask, after so many acquittals, whether consistent inference does anything more than avoid contradiction. It does. Reasoning correctly does not merely keep you consistent; under conditions named exactly, it carries you towards the truth, and the carrying is not a hope or a habit or an article of faith. It is a theorem.

State it plainly first, then guard it. If you keep updating consistently as evidence arrives, your credence in the true hypothesis converges towards one, provided its conditions hold: the truth represented, the prior keeping it in play, the observations arriving in a stable regime, and the rivals’ predictive differences recurring rather than appearing once. Each condition has a face you have met. Realizability fails in the diagnostician who never lists the rare disease, so no evidence can ever elect it. Prior support fails at the welded door, the live hypothesis assigned zero. That door never opens again. The sampling condition fails in the pollster who surveys one street and reports the nation. Distinguishability fails wherever two theories issue identical forecasts, so the evidence, however plentiful, cannot tell them apart. Of the four, three face the world and carry the price. The first is realizability: the truth is somewhere in the space of hypotheses you are entertaining, not excluded from the start, and here the least-assuming prior earns its keep a second time, since a starting point that welds no door shut is what guarantees the truth, if present, keeps nonzero standing. The second world-facing condition is distinguishability: the truth makes different predictions from its rivals, so that evidence can, given enough of it, pull them apart. And the third is the stable regime itself: the world that grades tomorrow’s updates must be the world that trained them, or near enough, and when it is not, when deployment drifts from the data, you have the grue trial industrialised, and convergence must begin again on the new regime. Grant the four, and convergence is guaranteed, not likely, not usually, but forced by the same consistency that forced the update rule; as evidence accumulates, the truth wins.

The mathematics behind this is associated most closely with Joseph Doob, who in the mid-twentieth century made rigorous the sense in which consistent updating is a process that, in the long run and under the stated conditions, lands on the fact. And now notice what is not among the conditions. There is no global premise that the future resembles the past, no assumption that every feature of nature holds still. There is one named local condition, and this paragraph exists to point at it. Conditional on each hypothesis, the observations must arrive in a stable regime, and the rivals’ differences must keep showing up. The theorem answers Hume conditionally, not from consistency alone. The condition is on the table where he can see it. No appeal to custom or habit. The convergence falls out of the structure of consistent inference itself, which means the very thing Hume declared rationally groundless, the tendency of honest learning to arrive at truth, proves to be a mathematical consequence of reasoning consistently at all, available to a consistent updater under those conditions.

Sit with what that does to the ghost of Part Three’s opening. Hume asked what justifies believing that induction will keep working, and could find nothing, and concluded we run on habit. The theorem does not remove induction’s stability condition; it isolates it, prices it, and puts it on the table. Once realisability, prior support, stable sampling, and recurring distinguishability hold, convergence is mathematics rather than habit. You do not have to assume the method works and then worry that the assumption is circular. You have only to update without contradiction, on a truth your hypotheses can represent and your sampling can reveal, and arriving at truth then comes as a structural consequence of doing so. Hume looked beneath induction for a further foundation and found empty air. The convergence was overhead all along, waiting to be proved rather than assumed.

Honesty requires that these conditions be examined rather than waved through, because a theorem is only as strong as its antecedents, and these four conditions carry the guarantee, and the world-facing conditions carry the price. Suppose realizability fails, and the truth is simply not in your hypothesis space. Then you will not converge to it, and you cannot, because you cannot approach what you have excluded from consideration; and this is a genuine and permanent limit, but notice precisely what kind of limit it is. It is not a flaw in induction. It is a demand on imagination, on the generation of hypotheses, the one place in this whole book where the creative art of the abduction trial does load-bearing work that no mechanical procedure can supply: make sure the truth is among the candidates, because the update rule can only raise what you were willing to consider. Suppose instead distinguishability fails, and the truth makes the same predictions as some rival, forever, under all possible evidence. Then you will not separate them either, but here the failure is not yours and not induction’s; the hypotheses are observationally identical, evidence is silent between them by their own construction, and no method that ever existed, inductive or divine, could do better, because there is nothing left to go on. So the world-facing conditions are not arbitrary fine print bolted to a hopeful theorem. They are the exact statement of when finding the truth is possible at all, and the theorem’s real content is bracing: within a realizable, supported, stably sampled, and distinguishable model, repeated consistent updating finds the truth, and where consistent inference cannot, nothing could.

Which turns even the limitations into a kind of vindication, and lets me answer the friend and the rain one final time before the machines take the example over. You believed the testimony because the friend had been reliable before, and I can hear Hume asking why past reliability should speak to present accuracy, whether that is not the very leap he questioned. Now you can answer him without flinching. You hold hypotheses about the friend, reliable and unreliable and the many grades between, and each past occasion, each time they were right or wrong, updated your credence across those hypotheses, and “this friend is about eighty-five percent reliable” came to fit the record better than its rivals. That hypothesis is not about the past. It spans past and future by its very content; it says they are reliable, not merely were; so when they speak now, the hypothesis that earned its standing on yesterday’s accuracy pays out on today’s, not by an extra inductive leap bolted on at the end but because reliability was always a claim about the person and not merely a summary of the log. The convergence theorem is that same move made general and made rigorous: keep updating honestly on a findable truth, and you close on it. The rain question opened everything as a puzzle about how much to believe a sentence. It ends here as an instance of a theorem about how belief, disciplined, finds the world.

So the ground stands, and it leads somewhere, and the trials close not on a successful defence but on a positive gift. Everything until now has concerned a single mind reasoning: you, at the window, weighing a friend’s word. But minds do not reason alone. They reason in languages they did not invent, with tools forged by the dead, inside institutions built to catch the errors no individual can catch in himself, and lately with engines of inference that are not human at all. How reasoning escaped the single skull is a story we are living through the climax of. Before the machines, the history, and the debt to the dead who built the mathematics.

One more thing deserves saying before the court adjourns, because the argument has earned a headline and should not whisper it. The results of these two acts admit statement as laws, with the same structural role the laws of thermodynamics play for machines made of matter. First a ground, the zeroth law, beneath the others the way a floor is beneath a house: consistent inference is possible. Then three laws standing on it. One: degrees of belief must be probability. Two: the lawful starting point is the one that assumes least. Three: there is exactly one way to change your mind, the way that neither invents information nor destroys it. A ground and three laws; if you carry nothing else out of these chapters, carry them. Laws, because within the stated scope no consistent alternative exists. And the exact scope of the whole architecture deserves stating once, plainly, so the confident verbs elsewhere never need a chaperone. Where the domain is determinate and the logic Boolean, degrees of belief must be probability. Where constraints and a reference measure single out a solution, maximum entropy fixes the starting point, and new constraints enter by the one consistent minimisation, with the classical conditioning rules as its central cases. Where they determine no point, the mandated state is the credal set. And given a representable, distinguishable truth, stably sampled, with the least-assuming prior keeping it in play, repeated updating converges on it. Those are not concessions. They are the boundaries of theorems, and a theory that proves both what follows and where determination ends is stronger than one that claims to answer everything. And the ledger of assumptions closes with one sentence, stated whole: what remains open is the choice of logic, and what remains empirical is the content of the constraints, and no epistemology could close either, and none should try.

The court is adjourned, and the convergence holds, at the price of the conditions that name what truth-finding costs.

And before the story widens to the whole species, pocket what the trials bought, because every verdict doubles as a specification for the reasoners now being built. Hume bought disciplined generalisation, standing earned by calibrated repetition rather than assumed from scale. The guillotine bought the boundary a machine can cross while reasoning flawlessly, between what follows and what is worth wanting. Goodman put representation inside the warrant: what pattern did the learner actually learn, and does it survive outside the frame that taught it. Peirce split generation from evaluation, so that fluent hypothesis is priced as art until it is tested as law. The lottery bought graded belief with stakes-priced action. Gettier bought route integrity: the answer is not the knowledge; the connection is. The sceptic bought local audit without global surrender. Popper bought severe tests that rivals would fail differently. The cube bought lawful refusal. The virtues bought conduct you can score. And convergence bought the conditions under which growing confidence has earned growing trust. Eleven verdicts, one specification sheet. The machines are next.