The direction nothing chose
Run Newton's equations backwards and they still balance. Run Maxwell's. Run the Schrödinger equation. None of the fundamental dynamical laws care which way the time coordinate points; reverse it and the mathematics keeps working. And yet no one has ever watched a shattered mirror reassemble, or coffee un-mix from cream. Boltzmann, working through the statistical mechanics of gases in the 1870s, showed why: entropy increases not because it must, but because the number of disordered configurations vastly outnumbers the ordered ones, so a system left alone drifts toward disorder almost every time. Loschmidt objected that reversible microscopic laws cannot produce an irreversible macroscopic world, and he was right to press the point — the resolution is that our universe began in an exceptionally low-entropy state, and everything since has been running downhill from there. Eddington, in 1927, gave the phenomenon its name: the arrow of time. Clausius had already given entropy its name, and its tendency, in 1865. The asymmetry was never in the laws. It was in the starting conditions.
What a record can and cannot do
The arrow matters beyond thermodynamics because it governs what a record is. A record is a correlation between the present state of something and an earlier event — a scratch, a stain, a photon finally arriving after travelling for millions of years. Records form only because entropy increases; that is the whole mechanism by which the past leaves a trace. And because the process runs one way, records point one way. Nothing in nature keeps a ledger of Tuesday's weather on Monday. This is why observation itself has a direction: you can observe what has already happened, you can observe what is happening now, and you cannot observe what has not happened yet. Those are not three arbitrary buckets. They are the only relations to time that a physical observer, embedded in a universe with a monotonic entropy gradient, can occupy.
Where the lineage claim comes from
A Large Language Model is built from a corpus that was gathered, frozen and shipped. It holds records of a past that has stopped producing anything for it; it has no present, so it cannot distinguish a fact that has since changed from one that hasn't. A Large World Model corrects for this by standing in a present — it senses a scene as it happens — but only while the scene lasts. When the scene ends, so does its standing point; its relation to time is episodic, not continuous. A Large Universe Model keeps the present open indefinitely: streams keep running, beliefs carry timestamps and provenance, and older beliefs get revised as newer evidence arrives from the arrow's one direction. That is the terminal relation to time available to any observer, because the arrow supplies no fourth category. There is recorded past, observed present, and an unobserved future accessible only by inference. Intake, as an axis, ends there — not because engineering stops improving, but because physics stops offering new kinds of thing to ingest.
The astronomer's version of the same problem
Astronomy makes this concrete in a way few fields do, because astronomers work simultaneously with all three relations to time and know exactly which is which. Harvard's DASCH project has digitised roughly half a million photographic plates spanning 1885 to 1992, reaching down to about fifteenth magnitude. That archive is a Large Language Model's kind of evidence: a frozen corpus of a sky that no longer exists in that configuration, useful for establishing century-long baselines — has this star varied, has that nova recurred — but incapable of telling you what the sky is doing tonight. It has a past and no present.
A single night's imaging run has the other limitation. The Zwicky Transient Facility scans large fields and produces on the order of a million alerts a night; each exposure is a scene, sharp and complete, gone the moment the shutter and the next slew move on. That is the Large World Model's condition: a present with no continuity beyond the frame. The camera sees the sky as it is, but only for as long as it is pointed there.
What survey astronomy has been building toward — brokers such as ANTARES, ALeRCE and Fink, ingesting alert streams, cross-matching against archival plates and catalogues, issuing classifications within seconds to minutes of detection, then updating those classifications as spectroscopic follow-up and later photometry arrive — is the third relation. The stream never closes. Every alert carries a timestamp, a filter, a magnitude with an error bar, and a provenance chain back to the specific exposure that produced it. Beliefs about an object — this is a cataclysmic variable, this is a tidal disruption event, this is a Type Ia supernova candidate — are held provisionally and revised as later data supersede earlier data. That is not a stronger version of the archive. It is a different relation to the arrow, because the archive stopped producing evidence and the stream has not.
When the record catches you, and when it doesn't
The characteristic failure of this domain shows exactly what continuous intake is for and what it cannot fix by itself. A transient — a kilonova, a fast blue optical transient, a tidal disruption flare — rises, peaks and fades on a timescale of hours to weeks. AT2017gfo, the kilonova counterpart to GW170817, was followed spectroscopically for less than two weeks before it dropped below what ground-based spectrographs could usefully classify. If the alert broker flags the event, but the survey astronomer responsible for allocating spectroscopic time is asleep, or the target list is backed up, or the telescope time has already been committed elsewhere, the transient fades before anyone points an instrument at it capable of confirming what it was. The photons already arrived. The record existed, briefly, in the sky itself, and then it didn't. No amount of archival depth recovers that loss, because the archive only ever holds what someone captured while the arrow was delivering it.
This is the sharpest argument for the third position in the lineage, and also the sharpest limit on it. Holding the stream open, with provenance, is necessary — a system that only consults the archive will never even know the transient occurred until someone else classifies it. But holding the stream open is not sufficient. Bandwidth, staffing and telescope scheduling are separate problems, and no architecture of intake solves a shortage of spectrograph time at 3 a.m.
Objections a survey astronomer would actually raise
The arrow of time is a boundary condition, not a law. Cosmology could have started differently. Building a hierarchy of evidential access on a contingent low-entropy past is building on cosmological accident, not necessity.
The concession is real: the arrow is not a fundamental law like the ones it fails to violate. But the argument doesn't require necessity across all possible cosmologies, only within this one. Every survey telescope, every broker, every spectrograph operates inside a universe where entropy increases monotonically and records form only backward. Within that universe — the only one available to build instruments in — the three relations to time are exhaustive. A hypothetical cosmos with a different arrow would need a different astronomy, not an additional rung on this one's ladder.
Prediction is evidential access to the future. Orbital mechanics forecasts an eclipse decades ahead with certainty; ephemerides are checked against the sky and hold. If that counts as observing the future, the arrow doesn't close the axis.
Ephemerides are genuine knowledge, and dismissing orbital prediction would misdescribe astronomy's oldest success. But an ephemeris is an inference from past observation plus dynamical law, and it stays falsifiable by what actually arrives — a predicted occultation that doesn't occur at the predicted second is corrected by observation, never the reverse. The distinction is exactly this asymmetry: forecasts get overruled by intake; intake is never overruled by forecasts. A transient-alert pipeline running a classifier on incoming photometry is doing the same thing at speed. It is inference layered on continuous observation, not a fourth category of observation itself.
What's left, once intake is granted its terminus, is not a solved field. It is calibration — knowing how much of the stream to trust, how stale a given belief about an object has become, and whether the telescope will be free when the arrow finally delivers something worth chasing.