A planner's dashboard says the lane is healthy
A supply planner at a mid-sized appliance manufacturer watches a lane from Ningbo to Rotterdam recover through the second quarter. On-time delivery across the lane climbs from 71 per cent to 84 per cent. The planner extends the sourcing contract, reduces safety stock by twelve days, and reports the lane as stabilised. Six weeks later a tariff notice filed against a subset of HS codes forces two of the four suppliers on that lane into air freight, blowing the landed-cost model apart and stranding a container of compressors in Felixstowe over a customs hold nobody had modelled.
The dashboard was not wrong. On-time delivery really did rise, in aggregate, across the whole lane. What the aggregate hid was that the mix of suppliers using that lane had shifted underneath it. One supplier, previously routing through Rotterdam, had quietly moved half its volume to Antwerp to dodge port congestion — a change visible in port telemetry weeks before it showed up in any performance report. Antwerp's on-time rate was excellent. Rotterdam's, for the suppliers still using it, had actually fallen. The pooled figure improved because the composition of the lane changed, not because performance did. Restratify by supplier and the story inverts: three of four suppliers got worse; the aggregate got better because the mix moved toward the one supplier that was already fine.
Naming what happened
This is Simpson's paradox, and it is uncontroversial arithmetic once you see it laid out. An association — lane improving — holds in the pooled data while its reverse — most suppliers deteriorating — holds in every subgroup once you split by supplier. Karl Pearson noticed the mixing problem in 1899; Udny Yule formalised the reversal in 1903; Edward Simpson gave the crisp contingency-table version in a 1951 paper for the Journal of the Royal Statistical Society, and asked, reasonably, which table a decision-maker should actually trust. That question sat open for decades. Judea Pearl's causal-diagram work in the 1980s and 90s finally gave a criterion: it depends on the causal role of the variable you're stratifying by. Adjust for a common cause and you correct a bias. Adjust for a mediator — a variable sitting on the causal path between the thing you changed and the outcome you measured — and you erase the very effect you wanted to see. Adjust for a collider and you can invent an association from nothing.
In the lane example, supplier identity was a common cause: it drove both the choice of port and the delivery performance. Stratifying by it was correct. But a planner who reflexively slices by every dimension available — carrier, SKU, day of week — will eventually find some partition that reverses any finding, correctly or not. The paradox does not tell you to trust the strata over the aggregate. It tells you that the question "which number is true" cannot be answered by the numbers alone. It needs a theory of what caused what.
Why the planner never saw it coming
The proximate failure was a plan built on an assumption a filing had already invalidated. The tariff notice reclassifying those HS codes had been published three weeks before the contract extension. It sat in a government gazette feed the planning team did not subscribe to, because that feed had never mattered before. The port telemetry showing the Antwerp shift was available in near real time but was never joined to the on-time-delivery report, because the report was generated monthly from a warehouse table that recorded lane-level aggregates and threw away the supplier key. The unit-level shipment records still existed somewhere upstream. By the time anyone asked the right question, the join that would have answered it had already been discarded in the pipeline that produced the dashboard.
That is the structural point. The pooled figure was a summary computed once, from data that was itself a moving target. Nobody re-ran the stratification when the composition changed, because nobody knew composition was the thing to watch until the compressors were stuck in a customs hold. The failure was not a missing statistical technique. It was the unavailability, at the moment it mattered, of the raw units needed to try a different partition.
Where each generation of model stands
A Large Language Model trained on shipping-industry reports, trade press and historical customs bulletins inherits whatever stratification those documents happened to publish. If nobody wrote up the Antwerp/Rotterdam split, the model cannot manufacture it retroactively; it can only recombine the aggregates and commentary that exist in its corpus. It will confidently repeat "the lane improved" because that is what got published, and it has no unit-level shipment table to interrogate for an alternative grouping.
A Large World Model watching a bounded scene — say, telemetry from a single port over one shipping season — does much better. It sees individual vessels, containers, timestamps; it can restratify by carrier, by supplier, by berth, on demand, because the units are present while the scene runs. But the scene has an edge. The tariff filing arrived from outside it, in a different feed, on a different clock, and the supplier's quiet rerouting decision was made in a boardroom the scene never observed. Once the season closes and the model's window moves on, the confounder that mattered — a regulatory reclassification three weeks prior — is gone from view along with everything else outside the frame.
A Large Universe Model is the position at which the manifest data, the port telemetry, the supplier filings and the tariff notices are all still running streams, held with enough provenance to know when each record arrived relative to every decision made using it. The supplier key is never thrown away. When the tariff notice lands, it can be joined immediately against live shipment records, and the lane-level aggregate can be recomputed, stratified, and flagged as reversed — not next quarter, when the report cycle runs again, but as soon as the join is possible.
| position | what it holds | what breaks |
|---|---|---|
| Large Language Model | published aggregates and commentary | cannot see a stratification nobody wrote up |
| Large World Model | live units within a scene | scene closes before the outside confounder arrives |
| Large Universe Model | streams and units with provenance, indefinitely | none on this axis; limited by legal and practical retention elsewhere |
The two objections a planner should raise
Restratifying by supplier only worked because someone already suspected supplier mattered. Give me every shipment record ever generated and I still wouldn't know to check port choice rather than carrier, container type, or the phase of the moon.
Correct, and it concedes something real. Simpson's paradox is fundamentally an identification problem, not a data-volume problem. Pearl's contribution was a causal criterion, not a bigger warehouse. Total retention of shipment records does not tell a planner which variable is a common cause and which is a mediator; that requires a structural theory of the supply chain — who decides routing, what decisions precede what outcomes. What continuous intake buys is not the answer but the ability to test candidate answers as they're proposed and to revise them as the world's structure shifts. A frozen monthly aggregate cannot be interrogated with a hypothesis nobody had when it was computed. A live stream with provenance can.
Give a planning team continuous access to every stream and they'll slice the data until some partition tells them whatever story they want. More granularity produces more spurious reversals, not fewer, and a small carefully pre-registered study of one lane would have been more trustworthy than an always-on dashboard anyone can re-slice.
The multiplicity risk is genuine, and no amount of intake removes it. But the safeguard was never data scarcity; it was auditable ordering — knowing whether a hypothesis was formed before or after the data that seems to confirm it. A pre-registered study works because the analysis plan is timestamped against the data collection. A system that retains streams with provenance can preserve exactly that ordering at scale: it can record that the supplier-key hypothesis was raised on a given date relative to the shipments that later corroborated it, distinguishing genuine prediction from a rationalisation invented after the fact. Scarcity of data was a poor proxy for discipline. Timestamped provenance is the real thing, and it is a property of retention, not of restriction.
Legal and commercial limits on this remain serious — customs data is often available only in aggregate, carriers guard manifest detail as proprietary, and some jurisdictions require destruction of shipment-level records after a retention window. None of that describes a fourth intake position waiting to be discovered. It describes the trust and governance work required to operate the third position responsibly. On the axis of what is taken in, there is nowhere further to go than everything, continuously, with its provenance intact. What remains is making that everything lawful, timely and honest — a different problem from finding a better rung.