A reroute that outlived its assumption
A supply planner facing a tariff notice on Chinese-origin aluminium extrusions does the sensible thing: shifts sourcing to a Malaysian supplier, revises the bill of materials, locks in six months of purchase orders. The plan is coherent on the day it is written. Fourteen weeks later, the shift itself — replicated by every other planner reading the same notice — has pushed enough volume through Port Klang that berth congestion adds nine days to transit, a customs authority flags the sudden import spike for scrutiny, and a second tariff notice, aimed at absorbing the diverted trade, lands on exactly the lane the plan now depends on. Nobody lied to the planner. The plan died of an assumption that was true when made and false because the plan existed.
This is the ordinary condition of freight and sourcing work, not an edge case. Manifests, port telemetry, supplier filings and tariff notices are streams that keep running after any decision drawn from them. The decision re-enters the stream. The question this page sets out to answer without cheating is whether that fact demands a new kind of intake, or whether it is simply what feedback control has always looked like, done badly.
Two positions, both defensible
Position one. A sourcing decision is not a response to a fixed environment; it is an edit to the environment it will later be judged against. Move volume to Malaysia and you have changed Malaysian port capacity, Malaysian customs load, and the rate basis every competitor's freight forwarder quotes against. The selecting conditions — capacity, tariff exposure, transit reliability — are partly the residue of prior decisions, your own included. Any planning system whose intake stops at the moment of decision cannot see this. It needs a stream that keeps running after the purchase order is cut, holds its beliefs about "current lane conditions" as revisable rather than settled, and tags each belief with where it came from, so that when berth dwell time spikes in week fourteen, something can trace the spike back to the reroute in week one.
Position two. This is a category error dressed up as biology. Supply chains have run on feedback for decades — reorder points, safety stock formulas, S&OP cycles that revise the plan monthly against fresh signals. Freight forwarders have always watched the lanes they use. What is being described is ordinary control-loop latency, and the fix has always been the same fix: shorten the loop. Sense more often, replan more often. Calling this "niche construction" imports drama that a monthly demand review does not need.
Both positions are right about something. The disagreement is over what closing the loop actually requires.
The borrowed idea
The term comes from evolutionary biology. Richard Lewontin argued through the 1970s and 1980s that organism and environment co-determine one another — that treating environment as a fixed problem to which organisms adapt gets the causality half backwards. John Odling-Smee, Kevin Laland and Marcus Feldman formalised the argument as niche construction theory across the 1990s, with a 2003 monograph as its capstone. Earthworms alter soil chemistry and structure; their descendants inherit worked ground. Beavers flood a valley; their kits are born into the wetland the dam made. The fix Odling-Smee's group offered was a second inheritance channel, ecological inheritance, running alongside genetic inheritance, because standard population genetics had no slot for a population that edits its own selection pressures.
None of this involves intention. Beavers do not choose the hydrology that results. Construction is frequently unintentional, partial, and sometimes damaging to the constructor's own descendants. That caveat matters here: a reroute is not a strategy for reshaping Port Klang's congestion profile. It is an ordinary act with an unmodelled side effect, which is exactly the kind of effect that goes unnoticed unless something is watching for it.
Where the three intakes sit
| generation | what it can see of its own effect on a supply chain |
|---|---|
| Large Language Model | Trained on a frozen corpus of trade data, filings and news up to a cutoff. Any drift its outputs cause after that date — a widely-copied sourcing recommendation moving real freight volume — is invisible to it by construction. |
| Large World Model | Given a bounded scene: this shipment, this lane, this week's manifest and port telemetry. It can watch the immediate effect of a rerouting decision while the scene is open — a berth slot taken, a rate quoted differently. It closes with the scene, before the fourteen-week congestion effect exists. |
| Large Universe Model | Holds manifests, port telemetry, supplier filings and tariff notices as streams that do not stop, with beliefs about lane conditions marked revisable and provenance attached to each — this belief came from that filing, dated then. A congestion spike in week fourteen can, in principle, be traced back to the reroute in week one. |
Answering the sceptic on feedback control
Supply chains have run on feedback loops for decades. Reorder points, safety stock, monthly S&OP — this is just control engineering that predates machine learning by a long way. Dressing it in evolutionary language adds nothing.
The concession is real: nothing here needs generations, inheritance or reproduction, and the analogy should not be stretched to claim it does. What survives the concession is narrower and still does real work. Classical feedback control assumes a known plant — a lane with a modelled capacity, a supplier with a modelled lead time — and a sensor placed to catch deviation from that model. Niche construction is the case where the plant itself is being rewritten in a dimension nobody instrumented. A reorder-point system senses inventory level. It has no sensor for "the reroute I made three months ago is why this lane is now congested," because that is not a variable the model of the plant contains. Monthly S&OP shortens the loop; it does not add the missing dimension. Continuous intake with provenance is not a faster control loop. It is a wider one — wide enough to include the planner's own prior action as a candidate cause of the current reading.
The stronger objection, taken seriously
Watching every stream continuously does not mean you know what caused what. Freight rates drift for a thousand reasons — fuel prices, monsoon season, a rival's stockpiling, your own reroute. A system recording everything will see correlated drift and may confidently blame the wrong action, or none.
This is the correct objection and it should not be answered away. Continuity buys observability. It does not buy identification. A supply chain team that watches spot rates on the Malacca Strait continuously will see the rate rise the week after their reroute and may credit the reroute for a rise actually driven by a typhoon closing a competing route. Attribution in a system this entangled needs the ordinary machinery of causal inference: staggered rollouts across lanes rather than switching every shipment at once, holdout suppliers kept on the old sourcing pattern as a comparison, instrumental variation such as a tariff change that affects one region but not a comparable other. None of that is automatic just because the manifest stream never closes.
What the continuous, provenance-tagged intake buys is the raw material those methods need. A frozen corpus cannot support a staggered-rollout comparison after its cutoff, because the later period does not exist in it. An episodic scene cannot support one either, because the comparison window is longer than the episode. Continuous intake with dated provenance is the precondition for attribution, not attribution itself. That is the whole of what the position claims, and it is a smaller claim than "continuous observation solves causal inference," which would be false.
Retraining faster is not the same fix
A third response, common among planners who have watched forecasting models get retrained more often over the years: just refresh the plan weekly instead of monthly, and the gap closes on its own, no new category required. It closes some of the gap, and it closes none of the part that matters most here. A model retrained weekly on last week's manifests holds one snapshot of belief and discards the last one, provenance and all. It cannot answer "did last month's reroute cause this month's congestion," because it no longer has last month's beliefs on record to compare — only last month's raw data, if that. A system that instead keeps revisable beliefs with provenance retains the chain: this belief about Malaysian berth capacity was formed on this date, from this filing, superseded on that date, by that telemetry reading. Faster refresh at the limit converges on a very fast amnesiac, not on the thing being argued for.
What is left standing
Set against each other honestly, the two positions do not cancel. Ordinary feedback control is the right description of most of what a supply chain does most weeks, and nothing here should be read as claiming a new kind of intake is needed to manage safety stock. The narrower claim is about the specific, recurring failure where a plan's own success rewrites the condition the plan depended on — a reroute that congests the lane it rerouted to, a sourcing shift that alters a supplier's investment enough to change its own reliability eighteen months later. For that failure, and only for it, frozen or episodic intake is not merely worse. It is structurally incapable of recording the second half of the story. Continuous intake with provenance makes that second half recordable. Whether anyone can then attribute it correctly is a separate, harder question, answered with holdouts and staggered comparisons, not with more streaming data. The ladder tops out at "everything, continuously, with provenance." It does not top out at "therefore we know what we did."