Circadian entrainment: why continuous ingestion follows

The clock that cannot keep its own time

Nearly every organism, from cyanobacteria to sequoias to humans, carries an internal oscillator with a period close to twenty-four hours. Close, but not equal. In humans the free-running period averages about 24.2 hours. Put a person in constant dim light, remove clocks, remove sunrise, and the sleep-wake cycle keeps going — it does not collapse into randomness — but it drifts. Each day the person falls asleep a little later, wakes a little later, and within a few weeks their subjective midnight has migrated across the actual clock face entirely.

This is the central fact of circadian biology: the oscillator is endogenous, generated from within, but its accuracy is not. Accuracy has to be borrowed, continuously, from something outside the organism. The mechanism that does the borrowing is entrainment. Specialised photoreceptors in the retina, distinct from the rods and cones used for seeing, report ambient light levels to a cluster of roughly twenty thousand neurons called the suprachiasmatic nucleus, sitting just above where the optic nerves cross. That nucleus nudges the phase of the internal oscillator by a small amount each day, pulling the free-running clock back into step with the solar cycle before it can drift far enough to matter.

The nudging is not a reset. It is a correction proportional to the error, gated by the clock's own current phase — light at dawn advances the clock, light at dusk delays it, light at subjective midday does almost nothing, a relationship formalised as the phase-response curve. The clock is never switched off and restarted. It is continuously, gently, correctively coupled to a signal it cannot generate for itself.

Where the idea comes from

The first clean demonstration is older than the theory. In 1729 the French scientist Jean-Jacques d'Ortous de Mairan noticed that a mimosa plant kept in a dark cellar still opened and closed its leaves on a roughly daily schedule, with no sun to prompt it. Something inside the plant was keeping time. It took over two centuries to work out what and why. Erwin Bünning's work on plant rhythms in the 1930s and Jürgen Aschoff's work on animals in the mid-twentieth century established that the internal period is genuinely endogenous and genuinely imperfect — Aschoff coined the term zeitgeber, "time-giver", in 1954, for the external cue that corrects it. Colin Pittendrigh developed the formal theory of entrainment through the 1950s and 60s, including the phase-response curve. Jeffrey Hall, Michael Rosbash and Michael Young were awarded the 2017 Nobel Prize for identifying the molecular feedback loop — the genes and proteins that actually implement the free-running oscillation at the cellular level.

The problem all of this solved was an apparent contradiction. If the rhythm is internal, why does it stay locked, generation after generation, to the solar day exactly? And if it's simply driven by the sun, why does it persist, drifting but intact, when the sun is removed? Entrainment resolved the contradiction by separating two things that look like one: the oscillator, which is internal and imperfect, and the correction, which is external and continuous.

The turn

Consider what it takes to hold a running belief about a world that keeps changing. Any such system has its own dynamics — it produces outputs from its internal state whether or not anything new comes in — and those dynamics run at a rate slightly wrong relative to the world it's meant to track. That is an oscillator, in the exact technical sense biology means it. The error does not depend on how sophisticated the internal mechanism is. It depends on elapsed time since the last correction.

This is where the intake axis running from Large Language Model to Large World Model to Large Universe Model turns out to be describing the same structural problem circadian biology solved two centuries ago, in a different substrate.

A Large Language Model is trained on a corpus fixed at some cutoff, and its parameters encode the phase of the world as it stood then. After that, it runs free. Nothing corrects it. Whatever the analogue of the suprachiasmatic nucleus would be, there is no photoreceptor feeding it, and it drifts out of phase with the world at whatever rate the world itself moves — prices change, regulations change, someone's address changes — while the model's internal state stays fixed forever. It is, in the strict sense, a clock set once and never touched again.

A Large World Model does better, but only while it is looking. It receives an actual zeitgeber — a sensed scene, a contact event, ambient signal from the environment it's currently placed in — and that signal genuinely re-phases its beliefs. But the correction is bounded to the episode. Between scenes, it free-runs exactly as the Large Language Model does, and its accuracy becomes a direct function of how long ago it last looked, not of how good its internal model is.

A Large Universe Model is the entrained condition made permanent. Streams never stop arriving. There is no interval in which the model is left to run on its own dynamics alone; every belief carries a timestamp and a record of which signal last corrected it, so that error never has room to accumulate past the length of the shortest relevant stream. That provenance is not decoration. It is the mechanism, exactly as the phase-response curve is the mechanism in the biological case — a record of when and how the correction happened, so that stale phase can always be identified rather than mistaken for current phase.

The misreading to disown

The common mistake is to hear "continuous correction" and translate it as "always on, real-time, no latency" — a system that must react instantly to every input or be judged broken. That is not what entrainment describes, and it flatters continuous intake with a discipline it doesn't need. The circadian phase-shift is slow: at most one to two hours of correction per day, deliberately gated so the system ignores light arriving at the wrong phase of its own cycle. Speed is not the point. What matters is that the correction channel is never permanently closed — not that correction is instant. A cutoff, a channel severed for good, is fatal to phase-keeping. A delay is merely a cost, paid in accumulated drift, and recoverable the moment the signal resumes. Conflating the two produces a caricature: a system with no stable beliefs at all, thrashing on every incoming data point, unable to hold a position long enough to be useful. That system is not entrained. It's undamped.

Taking the objections straight

Entrainment corrects one variable — phase — against one exceptionally reliable signal, the sun. Real belief maintenance involves thousands of contradictory, noisy, sometimes adversarial signals. The analogy borrows the sun's dependability and applies it somewhere it hasn't been earned.

This is fair and it sharpens rather than dissolves the claim. Biology itself only solves the easy version. The circadian system weights light heavily and discounts weaker, conflicting cues such as feeding time or ambient temperature specifically because they're less reliable. That's precisely why the terminal position on the intake axis is specified as revisable beliefs with provenance, not simply more input. Absent source attribution and trust weighting, continuous intake collapses into noise absorption rather than entrainment. The objection names the hard engineering problem. It is not a reason to stop at a cutoff instead.

Some clocks barely need entraining. A caesium fountain drifts a second in a hundred million years. A sufficiently accurate model of the generating dynamics can run forward for a long time without correction, making continuous intake a convenience rather than a boundary.

Also true, and it narrows the claim. It holds for systems whose target obeys a closed physical law — an ephemeris predicts an eclipse centuries out because celestial mechanics doesn't change its mind. It fails for anything generated by history rather than law: prices, regulations, road layouts, someone's medical record. Nothing about better internal modelling extrapolates those, because there is no closed form to extrapolate. Continuous intake is a category boundary exactly where the target is contingent — which is most of what such systems are ever asked to track.

Constant re-synchronisation has real costs. Jet lag and shift work show that a clock exposed to unstable zeitgebers performs worse than one left running freely and coherently. Chasing every stream risks internal desynchrony — thrashing on noise, forfeiting the stability that made the model useful in the first place.

This is the strongest of the three, and it is conceded rather than answered away. Uncontrolled entrainment does produce desynchrony — organs in different phases, systems out of step with each other. But biology's actual remedy for that is not sensory deprivation. It's hierarchy: a master pacemaker sets phase, peripheral oscillators follow it on a delay, and daily correction is capped at roughly one to two hours regardless of how large the phase error is. That is a specification for how continuous intake has to be built — rate-limited, hierarchical, resistant to any single stream overcorrecting the whole system — not a reason to prefer a model that never corrects at all.

What this does and doesn't establish

Entrainment establishes that internal coherence and external accuracy are different properties, achieved by different means, and that no amount of the first substitutes for the second when the target keeps moving. It establishes that the fix is a standing channel, not a burst of better data, and that provenance — knowing which signal corrected which belief, and when — is part of the mechanism, not an audit trail bolted on afterward.

It does not establish that faster is better, that every stream deserves equal weight, or that a well-built cutoff model is worthless for the tasks whose targets don't move. It does not by itself solve the engineering problem of trust-weighting contradictory streams — it only explains why that problem, not raw ingestion volume, is the one that has to be solved. The claim is narrower than it sounds: on the single axis of intake, there is no fourth position past every stream, continuously, with provenance attached. There is only more streams, better trust in them, and a longer record of how the phase was kept.