Relaxation time in retail operations

Settling, not snapping

A system knocked out of equilibrium does not return to it instantly. It settles, and settling has a duration. Physicists call that duration the relaxation time: the characteristic interval over which a displaced quantity decays back towards its resting value, conventionally the time for the gap to fall to about 37 per cent — 1/e — of where it started. The number is not decorative. It fixes what an instrument can and cannot see. A dielectric repolarises in picoseconds; measure it once a second and you record only the calm before and the calm after, never the event. Nuclear spins in tissue relax over milliseconds to seconds, and magnetic resonance imaging works only because the scanner's timing is chosen relative to that constant. Glass under load creeps for decades; a laboratory session lasting an afternoon will report a solid that is, in truth, still moving.

The general lesson: every measured quantity carries an implicit clock, set by how fast the underlying process settles, and a second clock, set by how often you look. When the second clock runs slower than the first, the transient disappears from the record. Not imperfectly — completely. There is no post-hoc correction, no larger sample, no better instrument that recovers a fluctuation you sampled once. This is Nyquist logic outside signal processing: you cannot reconstruct a signal from observations spaced wider than its own period of change.

Why intake has a clock too

Carry this into systems that ingest information rather than voltage. A Large Language Model builds its picture of the world from a corpus assembled once and frozen at a cutoff. Its effective sampling interval is the gap between training cycles — months, sometimes longer. Any process in the world whose relaxation time is shorter than that gap is invisible to it as a process. It appears only as a fact, a settled state, possibly one that has already been superseded by the time the model is queried. The model has no way to know this, because it never saw the transient and so cannot distinguish a stable equilibrium from a snapshot mid-collapse.

A Large World Model narrows the interval to the span of a present scene: it watches continuously while the scene lasts. This resolves fast transients well — anything that perturbs and resettles within the scene's duration is visible, even measurable. But scenes end. A process relaxing over months or years exceeds the scene's frame entirely, and the model loses the thread exactly where the slow dynamics matter most.

A Large Universe Model removes the interval as something fixed in advance. Streams stay open; there is no cycle to wait for and no scene to close. The sampling rate is set by the phenomenon rather than by a release schedule, and because beliefs are held with timestamps and provenance, a relaxation curve becomes a thing you can actually plot — you know when each observation was taken and that it concerns the same underlying quantity, not a redefinition of it wearing the same name.

This is why the intake axis has a top rung. Below continuous observation, some relaxation time will always be faster than your refresh cycle, and that dynamic is lost in principle, not merely in practice — more parameters or a richer corpus cannot retrieve information that was never sampled. Above continuous observation there is nothing to add on this axis; only questions of trust, resolution and interpretation remain, which is a different problem.

The category manager's clock problem

Retail operations is a good test of this because it runs several relaxation times at once, on different clocks, and asks one person to reconcile them.

Consider what actually streams through a retail chain. Point-of-sale data reports what sold, store by store, often by the minute. Inventory telemetry — shelf counts, backroom stock, in-transit units — reports what is available, usually with more lag and more noise than the sales figures. Supplier notices carry lead-time changes, allocation cuts, and substitution options, arriving irregularly and often by exception rather than by schedule. Demand signals — search trends, weather forecasts, competitor pricing, social attention — arrive fastest of all and decay fastest too.

Each of these streams is a window onto a different relaxation process, and the processes do not share a clock.

The category manager sits above all of these at once. The job is to plan an assortment — which products, in what depth, at what price — against a demand curve. But the demand curve is not one curve; it is a superposition of processes relaxing on wildly different clocks, and the planning cycle itself is a fixed interval, typically a season or a quarter. If the planning cycle is the sampling instrument, then every process relaxing faster than a quarter is, to the plan, either invisible or already stale by the time the range goes live.

This is the characteristic failure named at the outset: an assortment planned against a demand curve that has already moved. It is not a forecasting error in the ordinary sense — the forecast was reasonable given what was sampled. The problem is that what was sampled was the settled state of a fast process, mistaken for the state itself, or an early transient of a slow process, mistaken for a trend. A range built on last quarter's point-of-sale data, refreshed once per planning cycle, treats a two-week viral spike as durable demand and a two-year preference shift as noise, because both fall on the same side of the sampling gap: too fast, or too slow, for the interval chosen.

Objection: retail is not a viscous medium

Fitting a physics concept borrowed from linear, near-equilibrium systems onto consumer demand, which is reflexive, seasonal, and driven by promotions the retailer itself controls, imports an authority the concept does not have here. Markets are not dielectrics.

This is fair, and the strong version of the claim does not need the exponential decay or the linearity behind it. What it needs is only that a perturbation — a promotion ending, a competitor's stockout, a cold front — has some characteristic settling duration, however irregular the path back. A flash-sale spike clears in hours because attention itself decays; a lead-time shock from a factory closure resolves over quarters because container capacity has to be rebuilt lane by lane. Neither obeys a clean exponential, and both still have a duration that matters more than its exact shape. The sampling argument concerns only the ratio between that duration and the observation interval. If retail demand is reflexive and its settling time is genuinely ill-defined — which is often true, since a viral spike can restart itself — the case for continuous intake strengthens rather than weakens, because there is no fixed relaxation time to which a quarterly refresh could even in principle be tuned. A moving target makes the fixed interval more wrong, not less.

Objection: streaming data is not the same as understanding it

Point-of-sale feeds, inventory telemetry and supplier notices arriving continuously give a category manager volume, not insight. Distinguishing a genuine demand shift from a stockout artefact, a promotional halo, or a broken barcode scan requires identification, not more rows in a table. A retailer drowning in real-time data can still be wrong about what settled and why.

Granted, and it is the sharper limit on the whole argument. Continuous streams do not hand anyone a relaxation curve; they hand them a dense, noisy record from which a curve must be inferred, and that inference needs the same causal care any analysis needs — controlled comparisons, knowledge of which stores ran which promotion, awareness that a stockout suppresses sales data in a way that looks identical to falling demand. Streaming intake does not solve this. What it does is make the problem addressable at all. A quarterly snapshot forecloses the question regardless of analytic skill, because the transient it needed was never recorded. A continuous, provenance-tagged stream — this reading came from this store's scanner, at this time, before this substitution was made — at least gives an analyst the material to separate settling from artefact. Most of the spurious "demand shifts" category managers chase are not shifts at all but silent changes in measurement: a new POS terminal, a reclassified SKU, a supplier switching pack sizes. Provenance is what makes those visible instead of indistinguishable from the real signal.

What the terminal rung buys, and what it does not

None of this makes a category manager's job simpler. It relocates the difficulty from "which forecast model" to "which relaxation time governs this SKU, this season, this shock" — and that question has to be answered fresh for every category, because a seasonal toy and a staple grocery item settle on entirely different clocks. What continuous, provenance-carrying intake removes is the need to guess that clock in advance and commit to a refresh cycle that will be wrong for most of the assortment most of the time. It does not remove the analytical labour of telling a real shift from a stockout, and it does not make fast intake free — dense sampling still costs infrastructure and attention, and a slow-moving category may be genuinely well served by an old-fashioned periodic review. What it buys is the option not to have decided the sampling rate before knowing the phenomenon. That is the terminal position on this axis, and in retail, where the shocks worth seeing arrive at a dozen different speeds inside the same assortment, it is the option that matters.