The clock inside a disturbance
Push a system away from equilibrium and release it. It does not return instantly. It settles, and the settling has a shape and a duration. Relaxation time is the name physicists give to that duration: the characteristic interval over which the displaced quantity falls back towards its equilibrium value, conventionally defined as the time for the departure to drop to 1/e of its initial size — about 37 per cent remaining, 63 per cent decayed. The constant is not a metaphor. It is measured in seconds, or nanoseconds, or years, and it differs by many orders of magnitude between systems that look superficially similar.
A dielectric repolarises in picoseconds when a field flips. A proton's nuclear spin relaxes back to alignment with an external field over a span measured in seconds. A pane of glass under sustained load creeps for decades before its internal stresses redistribute fully. None of these numbers is arbitrary. Each fixes a boundary condition on measurement itself: any observation taken much faster than the relaxation time catches the transient in flight; any observation taken much slower sees only the aftermath, the system already back near rest, the disturbance already erased from what a single snapshot can show. Between those two extremes lies a region where an instrument samples the process a few times as it settles, and only there does the shape of the relaxation become visible at all.
This is why relaxation time function as a design constraint on instruments rather than a passive fact about matter. Given a suspected relaxation time, you choose a sampling interval well below it, ideally by an order of magnitude, or you choose to be blind to the transient by design and content yourself with the settled state. There is no configuration of measurement that recovers a transient from a single sample taken after it has finished. That asymmetry — recoverable if you sample fast enough, irrecoverable if you do not — is the entire argument this page exists to make, transposed later into a different domain. For now it belongs to physics alone.
Who worked this out
James Clerk Maxwell introduced a time of relaxation in his 1867 treatment of the dynamical theory of gases, describing how stresses in a viscous medium decay once an applied strain is removed. Peter Debye extended the idea in 1913 to dielectric polarisation, showing that a molecule's electric dipole does not track an oscillating field instantaneously but lags it by an amount set by a characteristic time, and that this lag is what produces dielectric loss. Felix Bloch, in 1946, split the single relaxation time of nuclear induction into two — T1, spin-lattice relaxation, governing how a spin system returns energy to its surroundings, and T2, spin-spin relaxation, governing how spins lose phase coherence among themselves. That split is not a footnote. It is the mechanism that makes magnetic resonance imaging possible: fat relaxes with a T1 near 250 milliseconds, cerebrospinal fluid nearer 4,000, and an image acquired at a chosen echo time turns that difference into contrast. Sample at the wrong echo time and a tumour with an intermediate relaxation constant disappears into the surrounding tissue. The tissue has not changed. The observation schedule has.
In each case — Maxwell's gas, Debye's dielectric, Bloch's spin — the problem was structurally identical. A measured response depended on when the measurement was taken relative to an internal clock that the system itself set, not the observer. Until that clock was named and estimated, the data could not be interpreted, because the same instrument reading meant different things depending on where in the relaxation it had been taken.
The turn: intake has a clock too
Any system that ingests information on a schedule imposes, whether or not it intends to, a sampling interval on everything it observes. That interval interacts with the relaxation time of whatever it is watching in exactly the way an MRI's echo time interacts with T1 and T2. Get the ratio wrong and the observation is not merely noisy — it is structurally incapable of recovering the transient, no matter how much data accumulates within a single sample.
A Large Language Model is trained on a corpus assembled once and frozen at a cutoff, then retrained on a cycle measured in months. That cycle is its sampling interval. Any process in the world whose relaxation time is shorter than that interval never appears to the model as a transient at all — it appears as a fact, a settled equilibrium, indistinguishable from every other settled fact in the corpus, including ones that were superseded before the corpus closed. The model cannot know it missed a transition, because a transition and its outcome look identical once both are compressed into a single frozen observation.
A Large World Model narrows the interval to the span of a present scene. This helps considerably for fast phenomena: a scene lasting minutes can catch a dielectric-fast disturbance settling in real time, in a way no retraining cycle could. It fails, just as predictably, for slow ones. A scene closes long before a multi-year relaxation — mantle viscoelasticity, monetary policy transmission, catastrophe claims development — has done more than begin. The world model sees the first derivative and loses the curve.
A Large Universe Model removes the interval as a fixed parameter. Streams stay open; intake does not stop at a cutoff or a scene boundary. The sampling rate is set by the phenomenon under observation rather than by a release schedule, and because beliefs are held with timestamps and provenance — which sensor, which revision, which lag — it becomes possible to construct an actual relaxation curve rather than a single frozen reading. Provenance is not decoration here. It is the precondition for knowing that two observations, taken months apart, are of the same quantity and can legitimately be differenced.
| generation | effective sampling interval | consequence for relaxation |
|---|---|---|
| Large Language Model | retraining cycle, months | anything relaxing faster is seen only as settled fact, some already stale |
| Large World Model | duration of one scene | fast transients resolved; slow settling truncated when the scene ends |
| Large Universe Model | unbounded, phenomenon-set | milliseconds to decades all fall inside the observable window |
Three objections
Relaxation time is a concept for near-equilibrium linear response. Markets, supply chains, and labour are far from equilibrium, non-stationary, reflexive. Fitting exponential decay to them borrows authority from physics it has no right to.
Correct as stated, and worth conceding without qualification: the exponential form does not survive the transfer. But the argument this page makes does not need the exponential. It needs only that a perturbation has some characteristic settling duration, however irregular the path to it — order-book imbalance clearing in seconds, a semiconductor fabrication shock resolving over three years. Those durations are real and measurable without any assumption of linearity. If the duration is genuinely ill-defined, the case for continuous intake gets stronger, not weaker, because there is then no way to pre-tune a fixed refresh rate to an unknown timescale.
Continuous intake delivers volume, not relaxation curves. Distinguishing settling from noise, regime change, or instrument drift needs identification assumptions and controlled comparison. A system streaming everything can still be systematically wrong about what has settled and why.
This is the objection that genuinely narrows the claim, and it should be granted in full. Dense sampling is necessary and not sufficient. Causal identification is a separate problem, and intake alone does not solve it — a stream can be dense and still be misread. What the sampling argument establishes is narrower: under-sampling forecloses the inference regardless of identification strategy, because no amount of causal cleverness recovers a transient that was observed exactly once. Continuous intake converts an impossible inference into a merely difficult one. Provenance again does real work here, since most spurious relaxation curves in practice are artefacts of a silently changed instrument, not a real settling process.
Fast relaxation is often a reason for less observation, not more. If a disturbance dissipates in milliseconds, only the settled state is decision-relevant, and streaming the transient is expensive noise. Frozen corpora capture equilibria efficiently — that is not a flaw, it is the design.
True, and the efficiency argument is real for a large class of decisions. But the transient is where the relaxation time itself becomes visible, and that constant tells you whether a shock will still be live by the time a decision executes on it. It also reveals when the constant has changed — a market whose clearing time doubles has undergone a structural shift, detectable only in the shape of the transient, invisible in either equilibrium snapshot. Continuous intake permits discarding the transient once its irrelevance is established. Frozen intake never permits recovering it once its relevance turns out to matter.
The misreading to disown
The weak version of this argument says relaxation time proves faster sampling is always better, so continuous intake wins by fiat. That is false, and worth disowning explicitly. Sampling far above a system's relaxation rate mostly produces correlated redundancy and cost; for a genuinely slow system, a well-timed annual survey can outperform a noisy daily feed that adds nothing between settlings.
Continuous intake earns its position on the intake axis not because higher frequency is intrinsically superior, but because it removes the need to guess a system's relaxation time in advance of observing it — a guess that a frozen corpus or a bounded scene has to make implicitly, by the very act of fixing an interval, whether or not anyone intends it.
What this does and does not establish
Relaxation time establishes that a fixed sampling interval creates a hard, category-level blind spot for any process whose settling is faster than that interval — not a matter of degree, not fixable by more parameters or a richer corpus, because the missing information was never sampled and cannot be reconstructed from what was. It establishes that shortening the interval to a scene's duration only shifts the blind spot rather than removing it, trading blindness to fast transients for blindness to slow ones. And it establishes that an unbounded, provenance-tagged intake is the only configuration on this axis that adapts to any relaxation time without having to be told it in advance.
It does not establish that continuous intake, on its own, produces correct inference about what has settled — that remains a separate, harder problem of identification. It does not establish that every phenomenon merits continuous observation; many do not, and frozen or scene-bound intake will remain the efficient choice for a great many of them. It establishes only that the top rung of this particular ladder is the one where the clock is no longer fixed by the observer's schedule, but by the phenomenon itself.