An assortment plan that is wrong the day it ships
Take the objection at its strongest. A category manager builds next season's assortment from a demand curve estimated across thousands of SKUs, hundreds of stores, and a supplier calendar that stretches months ahead. Each line in that plan is individually defensible: the forecast for a given SKU used the best available sell-through history, the safety stock was sized against a sensible lead-time distribution, the price point was set against last year's elasticity. And yet the category manager knows, with something close to certainty, that some of these lines are wrong. Not "might be wrong" — wrong. A promotion will cannibalise a neighbouring line nobody modelled. A supplier will slip a container. Weather will do something to outerwear that the algorithm did not anticipate. The plan, taken claim by claim, is rational. The plan, taken as a whole, is known in advance to contain failures.
This is the preface paradox, and retail operations lives inside it permanently. David Makinson's 1965 point was narrower and more devastating than "big plans have errors": he showed that a set of claims can each be individually justified and yet jointly inconsistent, because conjoining many high-confidence claims mathematically drags the confidence in the whole conjunction down, often close to zero. A category manager who signs off five thousand SKU-level forecasts at 95% confidence each is not signing off a plan that is 95% likely to be right. She is signing off a plan that is, as a whole, almost certain to be wrong somewhere. She knows this. She writes it, in effect, into the margin of the planning meeting: some of these numbers will be off, and I own that. That is the preface.
The strongest objection to everything that follows is this: fine, that is a fact about probability, and no amount of data changes it.
Makinson's paradox is about the logic of belief aggregation, not about inventory data. However many POS feeds or supplier notices you pipe in, the multiplication of independent probabilities still drags a large enough conjunction toward zero. Streaming more numbers into the planning system does not repeal a theorem. Calling this an argument for continuous intake is a category error — you are treating an epistemological result as if it were a resourcing problem.
This lands, and it should be granted in full before anything else is said. The formal result is permanent. No supply of inventory telemetry, however dense, makes a five-thousand-line assortment plan jointly consistent. If every SKU forecast carries independent 95% confidence and there are five thousand of them, the conjunction's confidence is not recoverable by better sensors. That mathematics does not care about POS feeds.
But the paradox has two halves, and retail operations only ever needed the second one. The formal half says the conjunction is improbable — permanently, unfixably. The practical half asks a different question: given that you know some lines are wrong, can you find out which, and when? A printed plan cannot answer that question about itself. A plan under live observation can. The demand curve a category manager planned against was a snapshot; it started decaying the moment it was frozen. What changes with continuous intake is not the truth of the theorem. It is whether "some assortment is wrong" stays a confession or becomes a queue with named items in it: this SKU, in these forty stores, is now off-curve, as of this morning's sell-through.
Three feeds, one demand curve, no fixed point
The characteristic failure in this domain is specific and worth naming precisely: an assortment is planned against a demand curve that has already moved by the time the plan reaches the floor. This is not an edge case. It is close to the default condition, because the inputs that would reveal the movement arrive on different clocks than the planning cycle does.
Point-of-sale streams report what already sold, store by store, often within minutes — but they report the past, and a category manager reading last week's sell-through is reading a demand curve that existed last week. Inventory telemetry — stock-on-hand, backroom counts, in-transit units — tells you what you have, not what you will need, and it lags physical reality by the count-to-system latency of the store's own processes. Supplier notices carry the other half of the picture: a delayed container, a substituted material, a minimum-order-quantity change, arriving on the supplier's schedule, not the planner's. None of these three streams alone falsifies the plan. Together, read continuously, they can.
The plan was built as a single frozen artefact: a forecast, a buy quantity, an allocation, fixed at a moment and then executed against for a season. Everything that happens after that moment is discovered by the store, not by the plan. A stockout in week three is not new information to the retail floor. It is new information to the category manager only if something routes it back to her — and in most planning systems, nothing does, until the quarterly review, by which point the demand curve has moved twice more.
What a bounded scene actually buys you
It is tempting to say the fix is simply "more real-time visibility" — a live dashboard, a control tower, a scene the category manager can watch. This is real progress and should not be dismissed. A live view of shelf-level stock, updated store by store, lets her catch an emerging stockout in a specific location before the weekly report would have surfaced it. That is a Large World Model's contribution to this domain: bounded, sensed, and correctable within the window it covers.
But the boundedness is the point. A dashboard shows the stores it is wired to, for as long as someone is watching it, and it says nothing about a supplier notice that has not yet arrived or a competitor promotion that has not yet shown up in the sell-through. The scene ends — the shift ends, the dashboard is closed, attention moves to the next category — and whatever was true outside that scene stays unaudited. A category manager who caught this week's stockout in flagship stores because she happened to be watching the live feed has not solved the plan's underlying problem. She has patched one item, once, because she was present.
Provenance is the difference between a confession and a queue
The category manager's honest preface — some of these SKU forecasts are wrong, and I signed all of them — becomes actionable only if each forecast carries enough of a trail to be re-tested against whatever arrives next. This is where the second serious objection belongs.
Most demand forecasts are blended from thousands of weak signals — historical sell-through, weather indices, promotional calendars, competitor pricing — with no single traceable cause. Demanding that every SKU-level number carry a clean provenance chain is demanding an architecture that does not exist for statistically emergent numbers, and then blaming the planning process for not having it.
This is fair, and full provenance for a blended forecast is not achievable today, possibly not achievable at all in the sense the objection implies. But provenance in this domain does not need to be complete to be useful; it needs to be graded. A forecast tagged with which inputs most recently updated it — this SKU's number was last touched by August's sell-through and a promotional-lift model trained on last year's event — is enough to route a new supplier notice or a fresh sales anomaly to the right forecast without having to re-derive the whole model's reasoning. Retailers already do a version of this for markdown decisions and out-of-stock alerts; the gap is that it stops at the alert and rarely reaches back into the plan that generated the exposure. Partial provenance, applied consistently, converts "the plan is wrong somewhere" into "the plan is wrong in this category, on this signal, as of this date" — which is the only form of the preface a category manager can act on before the season ends.
The discipline continuous revision demands
The remaining risk is real and specific to a domain that already suffers from over-correction: chasing every POS blip. A category manager who reallocates buy quantities on a single bad week, or panics on one supplier delay, will destabilise an assortment worse than a stale forecast would have. Continuous intake without discipline produces thrash, not accuracy — and retail has plenty of institutional memory of promotions and reorders whipsawed by noisy short-term data.
The answer is not less intake but weighted revision: a single-store anomaly should not move a national forecast, but a corroborated signal across a cluster of stores, sustained over more than one reporting cycle, should. This is ordinary Bayesian discipline, not a new invention — treat the existing forecast as a prior with real weight, and require the new stream to earn its correction. Retail operations already has the primitive it needs — safety stock buffers, reorder thresholds, forecast confidence intervals — and the work is wiring live streams to those thresholds rather than to a person's judgement call made under time pressure.
The narrower claim
Retail's version of the preface paradox cannot be dissolved. A five-thousand-line assortment plan will always be almost certainly wrong somewhere; that is arithmetic, not a process failure. What continuous intake — POS streams, inventory telemetry, supplier notices, read against forecasts that carry even partial provenance — offers is not a plan that is finally correct. It is the difference between a category manager who apologises for error in the abstract at the quarterly review, and one who can name, this week, which SKUs are off-curve and why. That is the entire claim, and it is enough.