The strongest case against continuous correction
Start with the objection that should win. A supply planner sets a reorder policy, fixes safety stock, batches purchase orders monthly to hit volume discounts and container minimums, and the plan holds. This is not laziness. It is control theory doing its job. When the actuator — a container booking, a supplier contract, a customs declaration — comes in large discrete units, the optimal policy is to leave it alone until you must act, then act at full amplitude. Order the full container. Draw the full discount tier. Anything gentler wastes the fixed cost you already paid to open the channel.
This is bang-bang control by another name, and it has a pedigree. Feldbaum's 1953 result on minimum-time control of linear plants proved that the optimal strategy for a bounded actuator is to run it at its extreme and switch a finite number of times, never dwell in between. Pontryagin's maximum principle generalised the argument. A supply chain with lot sizes, minimum order quantities and tariff brackets is exactly this kind of plant. Continuous micro-adjustment against a discrete actuator is not virtue; it is thrashing. So the objection lands with force: telling a planner to watch every stream continuously and correct in small increments looks like it inverts a theorem that has been settled for seventy years.
Where the concession is total
Grant it fully on amplitude. When the actuator is bounded — a full container, a fixed production run, a tariff bracket that only changes at a threshold — extremal action is correct. A planner who nudges an order by three percent every day because a port telemetry feed twitched is not being precise. They are adding noise the plant must absorb, paying transaction costs on corrections too small to matter, and irritating every counterparty on the other end of the purchase order. Chattering in a supply chain looks like this: a broker who re-negotiates freight rates weekly because spot prices moved, a planner who reissues forecasts to suppliers every time a single SKU's demand blips, a procurement system that treats every port telemetry update as an action trigger. The actuator gets ground down. Suppliers stop trusting the signal and start padding their own buffers against it, which reintroduces the very variance the fine correction was meant to remove. This is the bullwhip effect's lesser-known cousin: not amplification through batching, but amplification through over-eager de-batching.
So the theorem stands. What it governs is when to pull the trigger on a discrete action. It says nothing about how often you are allowed to look before pulling it.
Separating the switch from the sight
That is the axis this page is actually about — intake, not actuation. A bang-bang controller's optimality is conditional on knowing exactly where the switching surface sits: the exact day safety stock crosses the reorder point, the exact filing that changes a tariff line, the exact clause in a supplier's amendment that quietly moves a lead time from thirty days to ninety. Locate that surface late and the "optimal" full-size order arrives on top of a stockout that has already happened, or ahead of a demand collapse nobody flagged. The switch is still bang-bang. Locating it is a sensing problem, and sensing is exactly where coarse intake fails.
Here is the operative failure mode: a plan survives on an assumption invalidated by a filing nobody read. A customs notice reclassifies a component under a new tariff code three weeks before a shipment departs. A supplier's 10-Q, filed quietly, discloses a force majeure clause invoked against a subcontractor two tiers upstream. A port authority updates berth allocation rules that add four days to a transit the plan assumed was fixed. None of these are demand-side signals a monthly forecast cycle was built to catch. They are documents, and the planner's intake — a monthly S&OP cycle built on last quarter's manifests — did not have them in its aperture.
The Large Language Model's failure is structurally the same event at a different clock speed: a single saturated correction at training time, then open-loop operation until the next one, confidently wrong about anything that changed since the cutoff and unable to know which things those are. A planner running a quarterly re-plan is a Large Language Model with a supply chain instead of a corpus: the plan is frozen, the world is not, and the interval between corrections is exactly the interval during which a filing can sit unread. A Large World Model tightens this to the scene — daily exception dashboards, weekly control-tower reviews — and this genuinely helps, the way closing the loop within an episode always helps. But the loop still opens between reviews. A tariff notice filed on the Tuesday after Monday's control-tower meeting sits unabsorbed for six days. Bang-bang dynamics at a shorter period, not their removal.
The chattering objection, met honestly
The second serious objection follows immediately: if continuous intake is the answer, why do the most sophisticated supply chains not run on it already? Because sliding-mode controllers that chase their switching surface too aggressively destroy the actuator. Translate directly: a planning system that ingests every port telemetry ping, every AIS position update, every amended bill of lading, and re-triggers a purchase decision on each one will saturate the procurement team's attention, generate contradictory instructions to suppliers within the same week, and excite exactly the unmodelled dynamics — carrier schedule noise, port congestion reporting lag — that a calmer system would filter out. This is the real chattering failure, and it happens. Planners who built "always-on" exception engines in the 2010s reported alert fatigue as the dominant complaint: hundreds of triggers a day, most retracted within hours as noisy AIS pings corrected themselves.
But diagnose the cause correctly. Chattering is what a loop does when correction is demanded at a rate the sensing and actuation cannot support cleanly — finite update frequency, reporting lag, unfiltered noise, no deadband. The fix in control engineering is never to retreat to quarterly batches. It is boundary-layer smoothing and higher-order sliding modes: filter the fast noise, act on the slow trend, keep the sensor running continuously underneath. A supply-chain version of this already exists in vendor-managed inventory built on daily point-of-sale feeds. Lee, Padmanabhan and Whang's bullwhip analysis showed that batch reordering amplifies demand variance at every tier upstream; replacing the batch with continuous small replenishment, filtered rather than acted on instantaneously, has let retailers report inventory reductions near 20 percent with service level held constant. The stream ran continuously. The actuator still moved in sensible, sometimes fairly large, increments. Chattering was avoided not by narrowing intake but by widening it and filtering downstream.
What the Nyquist objection actually limits
A third objection, sharper on numbers: if a supplier's financial health changes on a quarterly filing cycle, sampling it hourly adds nothing but wasted compute; Nyquist caps the useful rate at twice the process bandwidth. True, and worth conceding without reservation for any single channel of known bandwidth. Polling a supplier's audited accounts every hour is waste; the accounts do not move that fast.
The trouble is that supply-chain risk does not arrive as one band-limited channel. A tariff notice, a port strike declaration, a subcontractor's force majeure filing and a container ship's grounding are different processes with different, largely unknowable bandwidths, and the dangerous ones are precisely the rare, unscheduled events no Nyquist argument was fitted to in advance. The claim under this thesis is not "sample every stream as fast as possible." It is narrower: keep every relevant stream open, with provenance attached to each update so a planner can tell which filing changed which assumption and undo the change if the filing was wrong. That is closer to coverage than to rate.
The claim that survives
What remains, once the concessions are taken seriously, is this. Bang-bang actuation is correct and stays correct: order in full containers, negotiate contracts at their natural cadence, do not micro-adjust a discrete lever. But the location of the switching surface — the moment an assumption a plan depends on has been invalidated — is a sensing problem, and on that axis coarse, scheduled intake is dominated. Not aesthetically; the planner who finds the filing three weeks late pays for the correction at compounded cost, having already committed downstream orders against a stale assumption. Shortening the interval from quarterly to weekly to daily buys disproportionate reductions in that cost, exactly as the control-theoretic argument predicts, until the interval reaches every stream still running, revised continuously, each revision traceable to its source. Beyond that there is no further evidence class to add. What is left is provisioning: more coverage, better filtering against chatter, faster and more trustworthy attribution of each correction to the filing that caused it.