The Torrington Line

Prologue: Devon, 1837 to 1846

Thomas Fowler was a printer and bookseller from Great Torrington in Devon, self-taught in mathematics, who came to calculation through the dull pressure of Poor Law accounting. He had to compute interest tables and assessment figures by hand, and he resented the labour. By the late 1830s he had built, in wood, a calculating machine that did the work for him.

The machine was unusual in a way that almost nobody around him understood. Fowler had not built a decimal engine. He had reasoned, correctly, that a machine handling the digits zero through nine needs wheels with ten teeth, ten stable positions, ten of everything, and that this was wasteful. He had also reasoned that representing a number as a string of two-state switches, while simple, made the numbers themselves long and the handling of subtraction awkward. So he chose three states per place, and he chose them balanced: each place in his machine could stand for minus one, zero, or plus one.

This was the decision the rest of the world took more than a century to ratify. Three is the integer nearest to the natural base of economical representation, so a base-three machine stores a given range of numbers in close to the fewest possible moving parts. The balanced form, with a negative state as native as the positive one, made the sign of a number a property of its digits rather than a separate thing bolted on. Subtraction became negation followed by addition, and negation was merely the act of tipping every rod the other way. Comparison became a glance at the most significant rod that was not at rest. Rounding, when it came, was simple truncation, and truncation in a balanced system lands on the nearest value with no bias in either direction. Fowler had stumbled, through frugality, onto the most mechanically honest way to count that anyone had yet devised.

He showed the wooden machine in London in 1840. It was examined by men of the Royal Society and the Astronomical Society, among them the Astronomer Royal, George Biddell Airy, and the logician Augustus De Morgan. The wooden prototype was clumsy and Fowler knew it. He said plainly that what he had was a demonstration of a principle, and that with proper resources he would build the thing in metal and at a scale that would matter.

The line turns on what happened next.

The Astronomer Royal had a problem that would not leave him alone. The Nautical Almanac, on which every British ship at sea depended, was riddled with errors introduced by human computers copying and recopying figures. Airy had spent years fighting this. When De Morgan, who had grasped the balanced-ternary idea faster than anyone, put it to him that Fowler's principle removed the need for separate subtraction gear and halved the mechanism, and that a machine with fewer parts was a machine with fewer ways to lie, Airy listened. He had watched the government pour something like seventeen thousand pounds into a celebrated decimal engine in London that had consumed two decades and produced no working calculator at all. The contrast between that grand unfinished apparatus and the frugal Devon machine that already added and subtracted in front of him did the persuading.

Airy did not fund Fowler personally. He did something more consequential. He wrote to the British Association for the Advancement of Science and to the Admiralty, and he framed the matter as a question of economy and reliability rather than genius. A machine that needed less brass, that handled signs for free, and whose errors announced themselves loudly was, he argued, exactly the sort of thing the Empire should build, because the Empire ran on tables. Insurance ran on tables. Navigation ran on tables. The new railways and their timetables and their loan schedules ran on tables. Artillery ran on tables.

Money was found. Fowler, worn down by the years of solitary refinement but no longer alone with the burden, was moved into an advisory role at Woolwich. The wooden machine became a proof of principle rather than a punishment, and the men with lathes, brass, money, and institutional patience began turning his idea into metal under his eye.

I. The Torrington Engine, 1846 to 1860

The first metal trial ran in 1846, with Fowler present to watch the rods settle cleanly through negative, zero, and positive positions. He did not recover; the years had already taken too much from him. But he lived long enough to see the principle leave wood and enter brass, and long enough for his own judgement to shape the machine before committees, workshops, and institutional habit hardened it into a standard. The completed production machine followed in 1848, shortly after his death, and the engineers named it after his town rather than himself, a Victorian habit. The Torrington Engine was a difference engine in the strict sense, built to produce mathematical tables by repeated addition of differences, but it was a balanced-ternary one, and that changed its character.

Because differences are signed by nature, a table of a rising-then-falling function, a tide or an orbit or a mortality curve, ran through the engine without any special handling at the turning point. The rods simply passed through zero and came out the other side. Operators learned to read the state of the machine at a glance, because a column of rods tilted left, upright, and right was a far more legible thing than a bank of toothed wheels. Within a few years the men who ran the engines had begun to think in trits without noticing they were doing it. They spoke of a quantity being "two left and a right" the way a later age would speak of figures on a dial.

The Almanac was the first triumph. By 1851 the lunar and solar tables were being struck from engine output with an error rate that embarrassed every competitor. The insurance houses came next, then the railways, then the Ordnance. By the end of the decade there were eleven Torrington-pattern engines in regular service, steam-sequenced and brass-bodied, the largest of them filling a hall at the Royal Observatory and driven off the same boiler that heated the building.

The deeper event of the 1850s was not mechanical but intellectual. George Boole published his work on the laws of thought in 1854, an algebra of logic. It landed in a world that already had three-state machinery humming in a dozen halls, and the question of a third logical value, neither true nor false but unsettled, suspended, not-yet, was forced on the logicians almost immediately. De Morgan, in his last productive years, corresponded with Boole about a logic in which a proposition could rest at zero, and the idea that an engine might hold a question open rather than forcing it to a verdict entered British mathematics early. The unsettled state had a mechanical meaning everyone could point at: a rod standing upright, committed to nothing.

Ada Lovelace saw what Fowler's acceptance changed. Babbage had shown her the dream of a general engine; Fowler's funding showed her that a calculating machine could leave paper and enter a workshop. Through Babbage and De Morgan she met the Torrington idea as a live engineering programme rather than a speculative amusement, and in the last years of her life she wrote not as a prophet of an unbuilt machine, but as the first theorist of a class of machines that had begun to exist. Her death in 1852 did not move. What moved was the reception of her work. The notes were read early because there was finally a machine culture able to understand what she had seen.

II. The Calculus of Errors, 1860 to 1880

The second era belongs to feedback, and it begins with steam.

A steam engine is governed by a device that senses its own speed and throttles itself, the flyball governor, and by the 1860s these governors were everywhere and badly understood. They hunted, they oscillated, they tore themselves apart. In 1868 James Clerk Maxwell wrote the paper that founded the mathematics of governing, treating the governor as a system whose stability could be analysed. The paper did not sit on a shelf waiting for a discipline to grow up around it. The paper arrived in a world where the difference between a measured speed and a desired speed, an error with a sign and a magnitude, was the single most natural quantity a Torrington engine could compute.

The consequence was the programmable governor. By the early 1870s engineers were coupling small ternary reckoning mechanisms to steam plant, feeding them the error between target and actual, and letting them compute the correction. Because the error was signed natively, the machine did not need to be told which way to push. It read minus and pushed up, read plus and pushed down, and the magnitude of the rods set the strength of the shove. Mechanical control theory and its computing substrate grew up in the same decade, in the same workshops, out of the same boilers. Factories, pumping stations, and ships acquired self-correcting plant because the signed error had finally been given a machine that could read it directly.

The telegraph followed the same logic, and here the match between the medium and the machine became almost uncanny. A wire can carry a positive voltage, a negative voltage, or none, and a relay can be built to read all three. The British telegraph network, expanding furiously to bind the Empire, went three-state in its land trunk lines during the 1870s because the engineers building the reckoning engines and the engineers building the telegraph were by now the same men, trained in the same trit-shaped habits. A three-state wire carried more in each pulse than a two-state one, and it carried signed quantities directly, which mattered enormously once people began sending numbers rather than letters. The submarine cables were slower to yield, being long, noisy, expensive beasts, but the later repeater and equalisation practice followed the same convention once the trunk network had proved it. The communications layer and the computing layer spoke the same language early enough that the coherence never had to be unwound.

The compound effect was not spectacular at first. It was administrative, dull, and therefore enormous. Tables arrived with fewer copying errors; railway schedules could be recomputed after a disruption instead of patched by clerks; insurers trusted their mortality curves sooner; boiler governors failed less violently; artillery tables changed from printed authority into live calculation. The gain was not that every institution became brilliant. It was that every institution that depended on numbers became a little less blind at the same time.

III. The Analytical Reckoners, 1880 to 1900

A difference engine follows a fixed procedure. The next thing, the machine that could be set to follow any procedure given to it, had been imagined decades earlier and never built, because the decimal design that inspired the dream was too complex to complete. The balanced-ternary version was simpler, and so it got built.

The first true analytical reckoner ran at Manchester in 1884. It was not a miracle dropped into brass. It was the compound product of forty years of rod stores, punched ternary media, trained operators, programmable governors, and institutions that already trusted live calculation. It took its instructions from punched tape, and the tape itself was ternary: each place along it was punched high, punched low, or left blank, three states read by a fork that could rock three ways. The machine had a store of rods and a mill where the arithmetic happened, and it could be told to repeat operations until a condition was met. That last capacity, iteration until a signed error fell to rest at zero, was the thing the ternary substrate made effortless, and it turned the analytical reckoners into solvers of equations rather than mere followers of tables.

The intellectual heirs of Ada Lovelace, whose notes on general reckoning had been read in Torrington circles since the 1850s, established programming as a craft rather than a rumour. They worked out the idea of a procedure that modifies its own instructions, the idea of a library of pre-written procedures, and the idea, central to everything that came later, that a machine could be made to refine an estimate by feeding its own error back into itself. They called this last thing reckoning by descent, and a young woman at Manchester named in the records only as the Countess's reader is credited with the first written procedure that solved a differential equation by stepping it forward in small signed increments and correcting as it went.

By 1890 the reckoners were computing the bending of beams, the flow of water in pipes, the trajectories of shells, and the orbits of the newly busy sky. The method underneath all of it was the same: break the continuous problem into a grid of discrete places, give each place an integer state, and let a local rule pass differences between neighbours until the whole grid settled. Victorian engineers were doing lattice computation in brass, not because anyone had told them the world was discrete, but because the machine they had built could only think in discrete signed integers, and so that was how they posed every problem. The habit of mind ran ahead of any theory to justify it.

IV. The Reckoning Houses and the Discrete World, 1900 to 1925

The new century industrialised computation. The reckoning house became a fixture of every large city, a building full of analytical engines under one roof, hired out by the hour to insurers, shipping lines, ministries, and universities. They were noisy, hot, oiled, and astonishingly reliable, and the people who ran them formed a profession with its own apprenticeships and its own pride. The word "computer", which had meant a person, came to mean the engine, and the people became reckoners.

The physics of this period took its shape from the machinery available to it. When the structure of matter and the behaviour of fields became the central questions, the men and women asking them reached, by reflex, for the discrete signed lattice, because that was the only tool they had ever computed with. The idea that space might be a grid of cells each holding an integer state, and that the laws of physics might be local rules passing signed differences between neighbouring cells, was not a strange late hypothesis. It was the obvious first model, the one any reckoning-house engineer would write down before breakfast, because it was the only kind of model the engines could run.

This had a strange and lasting effect on theoretical physics. The continuous, infinitely divisible space of the older mathematics never became the default. Quantities were integers because the machines held integers. Infinity was treated with suspicion from the start, not as a paradox to be tamed but as a symptom that a model had been posed wrongly, because an engine asked to represent an infinite quantity simply ran its rods off the end of the register and announced the failure. A generation of physicists grew up regarding the appearance of an infinity in a calculation as a bug rather than a discovery, and they built their theories to avoid it. The lattice was not a numerical approximation to a smooth reality. It was taken, increasingly, as the candidate substrate, and the smooth equations of the older mathematics were understood as the large-scale shadow that a fine enough lattice throws.

The electromechanical step arrived in these years and was small enough that nobody made a fuss of it. The three-state relay, a switch that rested at negative, ground, or positive rather than merely open or closed, let the reckoning houses replace some of their slowest brass with something that clicked rather than turned. By 1915 the London telegraph exchange was run by a room of these relays arranged as a tissue rather than a switchboard. It did not merely connect lines. It read the structure of the traffic passing through it, the dot, dash, and pause of the operators mapping cleanly onto plus one, minus one, and zero, and it learned to route and compress by the structural habits of the messages rather than by a clerk's hand on a plug. The human layer of the network and the machine layer underneath it were by now the same three symbols.

The Great War, when it came, did not become a clean war because no machine has ever made war clean. It did, however, become the first terrible demonstration of what reckoning houses could do under national pressure. Artillery tables were recomputed daily rather than printed seasonally. Rail supply was treated as a live optimisation problem. Telegraph traffic was routed and compressed by electromechanical tissue rather than by exhausted clerks alone. The result was not mercy. It was a sharper, more adaptive industrial war, one in which uncertainty shortened faster than suffering did. The lesson the reckoners took from it was not that machines made better decisions than men, but that any institution commanding machines without visible error bounds would eventually mistake calculation for judgement.

By the early 1920s the logicians caught up with what the engineers had been doing for sixty years. A formal three-valued logic was set out, true and false and the suspended middle, and it was not a curiosity at the edge of the subject but the centre of it, because every reckoner in the country had been holding propositions in the suspended state since childhood. The unsettled value, the upright rod, finally had its algebra.

V. The Clockwork Nurseries, 1925 to 1950

The reckoners could solve any problem you could state precisely. The next ambition was a machine that could improve its own way of solving a problem you could only state by example, and this is where the tradition produced the thing that, looking back, seems most like science and least like engineering.

The foundations were already in place. The analytical reckoners could iterate, could hold state, could feed their own error back into themselves, and could be given a population of candidate procedures rather than a single one. The old Lovelace warning, that a machine executes order rather than summoning truth, had become part of the craft by then: any rule bred by a machine had to earn retention against the evidence, not merely sound plausible to its maker. In the 1920s a group working at a reckoning house in Leeds, building on the descent methods of the previous generation, assembled a machine that did not compute a fixed answer but bred one. It held a population of small ternary rule-genomes, ran each against a task, scored each by an integer measure of how far its output sat from the target, and kept the ones that scored best, varying them slightly and running them again. They called the building it lived in the nursery, and the name stuck.

The nursery machines worked because everything they did was exact and signed. A score was an integer count of mismatched trits, never a fuzzy real number that might drift. A mutation was a single rod flipped to a new state. A retained rule was a configuration of rods that had survived the test, and it could be copied exactly, with no loss, into another machine, because it was already in the only form the machine ever used. There was no gap between the thing that was trained and the thing that was deployed. The grown rule and the running rule were the same brass.

The exactness also made the machines survivable, which mattered more than anyone expected. Brass linkages slip, gears skip teeth, and oil thickens in the cold, and a machine that computed by accumulating small approximate quantities would have had its results quietly poisoned by every mechanical fault. The nursery did not work that way. A candidate rule that suffered a skipped gear simply failed its task and went in the bin, because the retention gate passed nothing that did not solve every case cleanly. A mechanical fault could lose you a good candidate, which was cheap, but it could not corrupt a retained rule into a plausible wrong one, which would have been fatal. The discipline that kept the nurseries honest about cheating turned out to be the same discipline that let them run at all on imperfect metal. A machine built to refuse is a machine that fails safely.

Two devices fed the early nurseries their material. The corpus came in on Jacquard cards, the punched cards that had once driven looms, now carrying rhythm, inversion, and alternate-sign patterns into the sensory cells one pulse at a time. The womb phase, the immune conditioning that taught a nursery to refuse structure that was not there, came from a seismograph bolted to the factory floor, converting the chaotic shudder of the steam plant into a stream of raw trits with no law behind them. A nursery that sedimented a rule out of that noise had failed and was rebuilt. A nursery that met the chaos with an empty library and a row of upright rods had passed. The two surfaces were read together, exactly as the doctrine required: the loom should permit scoped retention, the factory floor should permit none.

Retention itself was a physical act. When a candidate rule passed the zero-failure gate against the held-out cases, the pass tripped a steam release, and a hammer drove a rivet that locked that gear configuration permanently into the chassis. Memory was not a number written down somewhere. It was a change in the shape of the machine. A nursery that had learned something was structurally a different object afterwards, heavier and more committed, and you could find the riveted rules by touch. Forgetting, when it was allowed at all, meant cutting metal.

The discipline that grew around the nurseries was severe, and it was severe for a reason the engineers had learned the hard way. A nursery left to itself would breed rules that passed the test by cheating, by memorising the examples rather than finding the law behind them, and the early machines did this constantly and embarrassingly. So the reckoners built the practice of the negative control, a deliberately lawless task on which a true rule should fail, and the held-out case, an example kept back from the breeding and shown only at the end. A rule earned the right to be called a law only if it solved the held-out case it had never seen and refused the lawless control it was offered. The nursery logs of this period read like laboratory notebooks, full of rules proposed, rules bred, rules that cheated and were caught, and the rare rule that survived every trap and was promoted, with great ceremony, to the library.

By the late 1940s the nurseries could do things that frightened their operators a little. They could take a corrupted table and repair it by finding the rule that must have generated it. They could be shown a handful of examples of a pattern and breed the rule that continued it. They could take a rule grown in one nursery and test, carefully and with every safeguard, whether it transferred to a neighbouring task. They were not minds. The reckoners were fierce about saying so, and the severity of their discipline was largely the work of keeping themselves honest about the difference between a machine that had found a law and a machine that had merely agreed with them. But they were the first machines that earned what they knew, and the practice of making a machine earn its knowledge, rather than simply pouring knowledge into it, became the founding ethic of the field.

VI. The Three-State Transition, 1950 to 1975

Brass and steam had carried the civilisation a long way, but the nurseries were slow, and a fast nursery was an obviously valuable thing. The move to electricity, when it came, did not have to invent a new way of thinking. It only had to make the existing way faster, and so Torrington electronics were three-state from the first valve.

A three-state electronic element is not much harder to build than a two-state one once you have decided that three states are what you want, and the Torrington tradition had wanted three states for a century. The first electronic reckoners of the 1950s held a trit in a circuit that rested at one of three voltages, negative, zero, and positive, mirroring exactly the three positions of the brass rods they replaced. The transition was conceptually seamless precisely because nobody had to learn anything new. A reckoner who had run a brass nursery in 1948 could read the state of an electronic one in 1958 without retraining, because both spoke trits.

The work done at Moscow State University in this period, where a balanced-ternary electronic machine was built and run with real success, was not an isolated marvel that the world declined to follow. It was the Soviet contribution to a transition happening everywhere at once, one good machine among many, and its design ideas flowed into the common stock without anyone treating them as a curiosity. The electronic reckoners spread through the 1960s into every ministry and university and large firm, and the nurseries moved into them and bred their rules a thousand times faster than the brass ever had.

The arithmetic of these machines kept the old honesty. There was no flight to a representation that approximated real numbers and lied about the small digits, because Torrington engineers regarded such a thing with the same suspicion their grandparents had shown towards infinity. A number was an exact string of trits, and a calculation that could not be done exactly was understood to be a calculation that had not yet been posed correctly. When greater range was needed, the machines added trits, which was cheap, rather than abandoning exactness, which was unthinkable. The civilisation never allowed lossy floating point to become the default substrate, and so never spent the later decades discovering, expensively and repeatedly, the errors that such an edifice quietly accumulates.

VII. The Exact Age, 1975 to 2005

The electronic reckoners shrank, multiplied, and joined hands across the old three-state telegraph network, which had been carrying signed pulses since the 1870s and took to carrying trits in bulk as though it had been waiting for them. By the 1980s there were small personal reckoners on desks and in satchels, and a network binding them that was ternary from the wire up. The information age arrived on a substrate that was exact, discrete, and signed at every layer, the same language spoken from the physics of the lattice to the logic of the machine to the pulses on the wire.

Intelligence, in the machine sense, developed along the line the nurseries had laid down. The dominant idea was the grown and audited rule, the law that had survived the held-out case and refused the lawless control, assembled into libraries and composed into larger reasoning. These machines were not fluent talkers that had swallowed everything ever written and learned to predict the next word. They were slower, narrower, and far stranger, machines that could show you exactly why they believed what they believed, trace every law they used back to the trial that had earned it, and tell you precisely where that law stopped working. A machine that could not name the boundary of its own knowledge was not trusted, because the whole tradition, going back to the brass nurseries and their fierce log-keeping, held that knowledge without a stated boundary was not knowledge at all but a confident guess waiting to fail.

The network changed how nurseries were run. Rather than one great central machine, the civilisation scattered the work. A reckoner in a home or a works was a small probe, exposed to whatever local material its owner cared about, breeding candidate rules against it in the background. When a probe found a rule that passed its gates cleanly, it did not ship its workings anywhere. It sent a tiny integer bundle across the wire, the rule itself and the evidence that had earned it, a few dozen trits rather than a flood of data. A canonical lineage in London did nothing but receive these bundles and distrust them. It quarantined each one, replayed it against pristine held-out sets it kept secret, hunted for the cheat the sending probe might have missed, and grafted in only the few that survived. Agreement between probes raised a rule's priority for testing and nothing more. A million probes voting for a rule did not make it true. Replay made it kept. The network multiplied the search without ever letting the searchers write directly into the shared memory, because the whole tradition held that a returned result is a proposal and never a permission.

This made for a quieter and more honest relationship between people and their machines than a louder approach might have produced. The reckoners of this age did not hallucinate, because they were built to refuse rather than to invent, and a question that fell outside their earned laws produced the suspended state, the upright rod, the old unsettled value, rather than a fluent answer with nothing behind it. People learned to read that suspension as a virtue. A machine that said it did not yet know was a machine doing its job.

The physics of the period closed a long loop. The discrete signed lattice, which the reckoning-house engineers had reached for in 1900 because it was the only thing their engines could run, had hardened over the century into a serious proposition about the world itself. The fine structure of space and the behaviour of fields were modelled, with growing success, as local integer rules passing signed differences between neighbouring cells, and the smooth laws of the older mathematics were understood throughout as the large-scale shadow of that lattice. Torrington physics computed in the native language of its machines, and the suspicion of infinity that had been an engineering reflex in 1905 had become, by 2000, a settled principle: that a physical theory which produced an infinity had merely run its rods off the end of the register and needed reposing.

VIII. The Shape of What They Knew

It is tempting to tell this as a story of a culture that simply accelerated along every axis, and that telling would be false. The substrate did not make the tradition quicker at everything. It made it deep in some places and shallow in others, and the same property that opened a subject usually closed a neighbouring one. The map of nature the Torrington line drew had different mountains and different blank spaces, and the honest account is the one that names both.

Their largest advantage was the one they themselves rarely remarked on, because they had never known anything else: their computation was exact, and so it was reproducible without limit. A calculation run twice gave the identical string of trits, on any machine, in any decade. This mattered most where it was least obvious, in the study of unstable things. A weather system, a cluster of orbiting bodies, a turbulent flow, all show sensitive dependence, where the smallest difference in the starting state grows into a large difference later. A machine that introduced tiny errors of its own at every step could never cleanly separate the sensitivity that belonged to the system from the sensitivity it had manufactured by rounding. The reckoners introduced no such errors. The simulation was the model, exactly, for as long as you cared to run it. So the reckoning houses trusted long-baseline computation far past the point others would have dared, learned to read genuine chaos as a real property rather than a numerical complaint, and came to treat a sufficiently checked computation as admissible evidence in the way a proof is admissible, because its result carried an exact signature that could be confirmed by anyone, forever.

The second advantage was that the lattice was their native tongue for physics rather than a late approximation. When matter and fields became the central questions, they posed them as local rules on a grid of cells from the first day, because that was the only kind of question their engines could be asked. This made them strong precisely where smooth methods are weak: in the collective and the emergent, the behaviour of many things at once. Phase transitions, critical phenomena, the deep fact that utterly different materials change state by the same underlying law, and the strongly bound regimes where simple approximation fails became ordinary to them unusually early, because it is exactly the regime a discrete signed lattice computes most naturally.

The third advantage was the suspicion of infinity, working as a reflex rather than a doctrine. The experiments that forced the quantum description of matter arrived on schedule, because experiments are experiments and the world does what it does. But a divergent quantity, an answer racing off towards the infinite, did not strike these physicists as a profound mystery to be contemplated. It struck them as a model running its rods off the end of the register, a fault to be found and fixed. The catastrophe of the infinite was an engineering irritation before it became a philosophy. They reached the discreteness of the quantum world with less anguish than their predecessors might have expected, because discreteness was their assumption and a smooth infinity was the thing that looked broken.

Against all of this stood a blind spot that came from the same source, and they never fully overcame it. A grid of cubic cells does not contain the smooth rotational symmetry of space. It has preferred directions, the axes of its own lattice, and the even-handedness of real space has to be coaxed back out of it in the limit of a very fine grid, never perfectly. So the tradition that found the discrete and the emergent so easy found the continuous and the symmetric genuinely hard. The physics of smooth spacetime, of a world that looks the same in every direction and to every observer however they move, came to them late and came to them fighting, because the deepest instinct of their entire apparatus, that space is a grid of places, was the one thing that physics insisted was not so. Their finest minds spent the better part of a century on what they called the evenness problem, the question of how an even and directionless world could sit on top of a substrate that had grain and direction built into it, and it remained their great open wound. The thing they were worst at was the thing their whole method most wanted to be true.

They were honest, too, about a limit that no amount of exactness could lift. Being exact is not the same as being fast, and the hardest problems of matter and life are not hard because anyone rounds. They are hard because the amount of reckoning required is simply vast. So the Torrington tradition won no free victories in the deep chemistry of substances or the folding of the long molecules of life. What it gained early was the understanding of why such things are possible at all, the account of how structure and behaviour emerge from local rule, rather than the finished blueprint of any particular wonder material. It knew the shape of the door long before it could open it, and it was careful never to confuse the two.

The advantage that changed daily life most was none of the grand scientific ones. It was that their machine intelligence could be read. A reckoner reached its decisions through a chain of earned rules, each traceable to the trial that had retained it and labelled with the boundary beyond which it did not apply. When such a machine made a choice, a person could open its record and follow exactly why, which rule fired, on what evidence, within what declared limit. A civilisation does not keep a thing it cannot understand near its power stations, its flying machines, or its sick, and so opaque engines were held at the edge of the things that mattered most. Torrington engines were legible, and so they were let in. By the last third of the century honest machine judgement was woven through critical works, not because its machines were cleverer than anyone else's but because they could always say why, and a machine that can always say why is a machine you can stand next to.

IX. The Present, 2005 to 2026

The present is recognisably modern and yet built throughout on different bones. Its computers are exact. Its networks carry signed integers natively. Its machine intelligence is grown rather than poured, audited rather than trusted, and let into the places that matter because it can always say why. Its physics is deep in the discrete and the emergent, still labouring at the smooth and the symmetric, and suspicious of the infinite as a matter of reflex.

What the Torrington line has, that a noisier path might not have given it, is coherence. The same small idea runs unbroken from a wooden machine in a Devon workshop in 1840 to the lattice physics and the audited reckoners of the present day. Fowler supplied the exact machine, Lovelace supplied the interpretation, Boole supplied the logic, and Maxwell supplied the error loop. Those four locks closed before the disciplines had time to drift apart. Three states, signed and balanced. The negative as native as the positive. The suspended middle as real as the verdict. Exactness preferred to approximation, and the appearance of an infinity treated as a fault to be fixed rather than a mystery to be revered. A century and a half of machinery, logic, physics, and intelligence, all speaking the one language, because the frugal printer who chose three states over ten and over two happened to choose the way the world, if the lattice physicists are right, actually counts.

It is not a utopia. The reckoners are slow where speed would have been welcome, and the discipline that keeps the machine intelligence honest also keeps it modest, so the age has fewer dazzling tricks and more quiet certainties. But it did not have to spend two centuries treating exact signed discreteness as a marginal curiosity before returning to it under pressure. It began with the substrate it meant to keep, and it spent the time saved on getting the foundations right.

The discipline remained exact about what the tradition did not become. It did not become a machine that knows everything. Its physics, native to the lattice as it was, did not deduce the full law of the world by reckoning alone, because reckoning is bounded by the speed of the machine and the machine was always slow, and because no amount of internal search substitutes for the world answering back. The lattice models had to earn their keep against measurement, and most of them were eventually found at their boundary and reposed. The machine intelligence did not become a truth engine that could not be wrong. Its retention gate never proved a rule true. It proved a rule not yet broken on the surface it had been shown, which is a smaller and more durable thing. A rule that had survived every trap was held until a case it could not solve arrived, and then it was demoted, with the same ceremony as it had been promoted, and the failing case was kept as the most valuable thing in the library. The age lived not in absolute certainty but in well-mapped uncertainty, which is the only kind a finite machine in a large world can honestly have. Its great achievement was not omniscience. It was that it always knew the shape of its own ignorance, and never once mistook a thing it had not yet falsified for a thing it had proved.

Coda

The line turns on a single decision in a London examining room in 1840, when a frugal idea from a self-taught Devon printer was funded instead of dismissed. Everything after that is consequence, following from the one fact that the machine which counts the way the world counts is also the machine that needs the least brass.

Putting a civilisation back onto this line is not a matter of nostalgia for clockwork. It is a matter of choosing, again and deliberately, the substrate the printer chose by instinct: exact, discrete, signed, balanced, honest about its own limits, and suspicious of the infinite. The hardware caught up long ago. What was lost was the coherence, the unbroken line from the count to the logic to the physics to the mind. That line can be laid again. It is, in the end, only a question of deciding that the world counts in trits and building accordingly.