The chronometer and the longitude problem: why continuous ingestion follows

The problem latitude does not have

Latitude is generous. Point a sextant at the sun when it stands highest in the sky, note the angle, consult a table of the sun's declination for the date, and the arithmetic gives a position north or south of the equator. One observation, taken now, answers the question completely. Nothing about the ship's history matters. A castaway with a working sextant and no memory of the voyage could still fix latitude correctly.

Longitude will not cooperate. It is not an angle read off the sky but a difference: the gap between local time, which the sky freely gives up, and the time at some reference meridian, which the sky does not give up at all. Local noon is simply whenever the sun is highest where you happen to be standing. To know how far east or west that is from Greenwich, you must know what time it is at Greenwich at that same instant. And Greenwich is not visible. The only way to know its time is to have carried it with you, running, unbroken, since the last moment you were certain of it.

That is the whole difficulty, and it is worth sitting with before going anywhere near computation. Latitude is answered by an instant. Longitude is answered by a chain. Break the chain — let the clock stop, even briefly — and you do not get a slightly worse longitude. You get dead reckoning: course, speed, elapsed time, and an error that compounds with every hour uncorrected. The quantity itself is defined by accumulation since a reference point. No sharper sextant, no cleverer table, changes that. The information a snapshot can carry and the information longitude requires are different kinds of thing.

Harrison's answer

The stakes became unignorable in October 1707, when a Royal Navy squadron under Sir Cloudesley Shovell, misjudging its longitude in fog, ran onto the rocks of the Isles of Scilly. Estimates of the dead run past a thousand men. Parliament responded with the Longitude Act of 1714: up to £20,000 for a method accurate to within half a degree on a voyage to the West Indies. The Board of Longitude, staffed by astronomers, assumed the winning method would come from the sky — better tables of the moon's motion, perhaps, or of Jupiter's satellites.

John Harrison, a joiner from Lincolnshire with no formal astronomical training, spent four decades building something else: a clock that could survive a ship. Damp, salt air, temperature swings between the tropics and the North Atlantic, and the constant pitch and roll of a wooden hull all defeat an ordinary pendulum clock within days. Harrison's fourth sea watch, H4, was tested on a passage to Jamaica in 1761–62 and came home having lost about five seconds. Mechanical continuity had beaten celestial computation — not on elegance, and not at first on cost, but on what it actually delivered at sea: a reference that kept running through exactly the conditions that broke everything else.

The turn

Set the maritime problem down for a moment and look at three kinds of machine intelligence that have nothing to do with ships.

A Large Language Model is trained on a corpus fixed at some past date and then frozen. Ask it anything within that corpus's competence and it answers with the fluency of something that has read everything — because, in a narrow sense, it has. But it cannot tell you how long ago its knowledge stopped being current, because it has no mechanism for noticing that time has passed at all. It is an almanac: computed once, printed, correct at its epoch, and silently wrong forever after that, in ways it cannot flag.

A Large World Model takes in a present scene directly — sensor data, an image, a live feed — and reasons over what is actually there right now. This is a real advance over the almanac: it is grounded in the instant rather than in someone else's past observation. But it is exact about the present and mute about everything outside it. It cannot say what changed since yesterday, because it was not there yesterday, and it has no chain connecting today's scene to any earlier one. It is the sextant sight: precise, current, and structurally incapable of supplying a reference that only history can give.

A Large Universe Model is the third position: not a frozen corpus, not a bounded scene, but every relevant stream still running, held as beliefs that can be revised, each with a record of where it came from and how stale it is allowed to get before it is flagged as suspect. It is the chronometer. Not a better almanac. Not a better sextant. A third kind of thing, which is what lets the other two produce an actual fix instead of a guess.

The parallel did not need forcing. Longitude splits evidence into what an instant can give and what only an unbroken chain can give, with unusual cleanliness. That is precisely the split between a corpus, a scene, and a stream.

What the misreading gets wrong

The tempting shortcut is to conclude that more continuous data always wins — that a system ingesting everything, always, beats one ingesting less. Ships that carried a single unrated chronometer often did worse than navigators using lunar observations, because an unmonitored clock does not announce its own failure. It just keeps ticking, confidently, at the wrong rate.

What mattered was never continuity by itself. It was the rating certificate: a documented drift rate, a log of comparisons against known fixed points, an error bar that could be reasoned about and that grew, predictably, the longer it had been since the last check. A clock nobody rated is not a reference. It is a liability with a reassuring face. Continuous ingestion without provenance is not the third position on this ladder. It is a faster way to arrive at a confident wrong answer, and worth naming as its own failure mode rather than folding into an argument for continuity generally.

Objections that hold weight

Maskelyne's lunar-distance method, from the 1767 Nautical Almanac, found Greenwich time from the moon's position against fixed stars — an instantaneous sight plus printed tables. No running clock required. That's a frozen corpus and a present scene doing exactly what the chronometer did.

This is correct, and it is a real limit on the claim, not a nuisance to wave off. But look at the cost of the workaround: up to four hours of computation per fix, a sight accurate to a tenth of a minute of arc, and a clear view of the moon, unavailable on perhaps half the nights of any voyage. The sky itself was the continuous reference; the almanac was only a prediction of it, and the prediction failed exactly when the sky was hidden. Chronometers won not because instantaneous methods were impossible, but because continuity you carry beats continuity you must wait for and hope to see.

Chronometers drift. H4 lost about 5.1 seconds over 81 days to Jamaica; ordinary chronometers ran off by a second or more daily. Navigators corrected them in port against known longitudes — which is a return to fixed points, not an unbroken stream.

This states the mechanism rather than undoing it. A chronometer's value was never perfection. It was a known rate of drift, logged, applied between checks, with an uncertainty that grows in a calculable way the longer since the last comparison. That is what provenance and decay mean in practice. The port fix is not a reset to a corpus; it is one update in a record that carries forward through it. Take the running clock away and the port fix, however exact, is worthless the moment the ship clears the harbour.

The whole analogy is flattered by choosing a domain where the target is a single clean scalar — time — that is monotone, universal, and cheap to compare. Most things worth tracking (intent, ownership, whether a structure is still safe) have no chronometer and no rating certificate.

This one genuinely narrows the claim, and it should be allowed to. Nothing here says continuous ingestion turns a messy quantity crisp. The claim is specific: for quantities defined by accumulation since a reference point, no snapshot suffices, and continuity is an entry condition rather than a refinement. Ownership recorded in a title chain, or fatigue accumulated in a bridge, are such quantities, and they remain hard for their own separate reasons even once ingestion is continuous. The chronometer shows what the intake axis can achieve. It does not claim to be sufficient on its own.

What the argument establishes

It establishes that a class of quantities — those defined by accumulation since a known reference — is structurally unreachable from any snapshot, however detailed, and reachable only from an unbroken, provenanced chain. It establishes that this closes cleanly: after Harrison, nobody found a fourth kind of observation. Quartz oscillators, caesium standards, GPS time broadcast — all of these improved accuracy, cost, and redundancy. None added a category.

It does not establish that continuous intake solves problems outside that class, or that provenance alone makes a system trustworthy, or that the Large Universe Model, as an argued category, currently exists as anything one can buy. It says only that on this one axis — what a system takes in, and for how long, and with what record of where it came from — the ladder has a top rung, and the chronometer is the oldest clean picture we have of what standing on it looks like.