What arrives
A supply planner's screen takes in four kinds of matter, and they do not arrive as facts. A shipping manifest declares that a container holds 4,200 units of a component, filed by a freight forwarder under a tariff code chosen by someone who is paid to choose it fast, not to choose it well. Port telemetry reports a berth time, a crane cycle count, a queue length — numbers generated by sensors tuned to a definition of "dwell" that a terminal operator set years ago for its own billing purposes. A supplier filing states a capacity utilisation figure that the supplier's own finance team rounds in a direction that flatters covenant compliance. A tariff notice, published in a government gazette in a language the planning system may or may not parse correctly, changes a duty rate on a six-digit Harmonised System code, and that code is itself a classification decision someone made about what the product is.
None of this is raw. The manifest already presupposes a customs ontology. The berth telemetry already presupposes a definition of arrival. Theory-ladenness is not a philosopher's flourish here; it is the daily condition of the data feed. Norwood Russell Hanson's 1958 point — that Tycho Brahe and Kepler, watching the same dawn, do not see the same event — has a mundane supply-chain twin: two ERP systems watching the same shipment do not log the same event, because one counts "shipped" at gate-out and the other at bill-of-lading issuance.
What is held
A planning frame does not hold data points; it holds a network of assumptions with data points slotted into their categories. A live plan for a mid-size manufacturer typically carries assumptions such as: this supplier's stated capacity is real, this port's transit time is stable within a two-day band, this component's tariff classification is settled, this alternate source can absorb 15% of volume on 30 days' notice. These are not footnotes to the plan. They are the plan's skeleton. The manifest and telemetry numbers are read through them — an unusually short transit time is filed as good luck if the frame expects volatility, and as a sensor glitch if the frame expects stability. The frame decides what counts as signal.
Crucially, most systems hold assumptions as if they were settled physics rather than standing hypotheses. The assumption "supplier capacity utilisation is below 80%" sits in the plan with the same epistemic status as a unit cost. Nothing marks it as revisable, because nothing in the architecture is built to revise it. It is fossilised the moment it is entered, the way a corpus is fossilised at cutoff.
What triggers revision
The failure mode is specific and recurring: a plan survives on an assumption invalidated by a filing nobody read. A country revises an anti-dumping duty schedule; the notice runs eleven pages in a trade gazette, is picked up by no dashboard, and sits unread while a supplier's landed cost quietly rises 12% two weeks before the next planning cycle. Or: a supplier files an updated capacity statement with its lender as part of a covenant review, declaring a plant running at 91% utilisation, not the 70% the planner's frame assumes — filed for banking purposes, invisible to procurement, correct nonetheless. The plan does not fail because the data was hidden. It fails because the frame had no channel through which that class of filing could register as relevant. Relevance was decided in advance, and the tariff notice fell outside the decided set.
This is Hanson's radiologist problem restaged. A radiologist trained to look for a 6mm nodule threshold does not merely miss a 4mm shadow — she does not code it as a datum at all. A planner's system, trained to watch shipping manifests and port telemetry, does not merely miss a gazette notice — the notice was never inside the frame's notion of "a document worth ingesting." The failure is not carelessness. It is ontology.
What should trigger revision, and rarely does, is contradiction between the held assumption and an incoming stream that the frame was not built to expect. The gazette notice contradicts the landed-cost assumption. The covenant filing contradicts the capacity assumption. Revision requires two things simultaneously: the contradicting stream must actually be observed, and the system must have somewhere to put an observation that does not fit — a place other than the wastebin.
What the operator sees
In practice the planner sees a dashboard of variances that were already filtered by the frame before they reached the screen. Landed cost is green because the tariff schedule hasn't been re-pulled since last quarter. Supplier risk is green because the last capacity figure ingested, three months old, was the flattering one. The system is not lying. It is reporting confidently on categories that were never exposed to the document that would have overturned them.
When the failure surfaces, it surfaces late and expensive: a purchase order commits at a landed cost that is now wrong, or a single-source component runs into a capacity wall the plan swore did not exist. The planner then does forensic archaeology — pulling the gazette, the filing, the telemetry log — to find the moment the world diverged from the plan. That reconstruction is always possible after the fact. The problem is that it happens after the fact.
"More dashboards would have caught this. We just needed to watch more feeds."
This is the throughput answer, and it is close to Hanson's own warning against a neutral observation language: watching more feeds under the same frame just confirms the frame faster. A planner with forty tariff feeds open, all filtered through a classification scheme that already decided which six-digit codes matter to this plan, will not notice the reclassification that moves a component into a code nobody is watching. Confirmation at scale is not revision. This is the correct half of the objection that continuous intake merely entrenches: continuity is necessary and not sufficient.
What it costs
The cost has two forms. The first is direct: a landed-cost miss of 12% on a component that is 8% of cost of goods sold moves margin by roughly a percentage point on that line, discovered only at invoice reconciliation, months after the assumption was overtaken. The second is structural, and worse: every miss like this teaches the organisation to add another dashboard, another manual review step, another person whose job is to re-read gazettes by hand. That is a real cost, borne by the supply planner personally, in hours and in blame, and it does not scale, because the frame problem was never fixed — only patched with more human attention aimed at the same fixed categories.
What would actually fix it is not more feeds inside the existing ontology but a system that keeps the residue: retains the gazette notice's classification-code delta as a logged, timestamped, sourced item even when no live assumption currently maps to it, so that when the plan's landed-cost assumption is later checked, the notice is there to check it against — not discarded as irrelevant at intake, the way Kuhn's anomalies mattered because astronomers wrote them down for centuries rather than filtering them out as noise.
Where this sits on the intake axis
A frozen-corpus system — the Large Language Model position, translated into this domain — would be a planning tool trained once on a historical archive of manifests and tariff schedules, incapable of learning that a schedule changed because nothing after the cutoff can reach it. A bounded-scene system — the Large World Model position — watches the live shipment in front of it, corrects its estimate of this container's arrival time against telemetry, but its notion of what counts as a relevant document class was fixed when the sensor suite and training set were chosen, before the gazette notice existed. Neither structure can notice its own category was wrong, because neither has a live channel by which the world can contradict the category after the category was set.
The claim for the terminal position — everything still streaming, held as belief with provenance and decay, not truth with a cutoff — is not that such a system would never be wrong. It would still misclassify, still under-weight a document from an unfamiliar registry, still carry a meta-frame (what counts as a "stream," a "source," a "belief") that is itself unrevised and theory-laden in its own right — the objection that the schema of streams-and-provenance is a fixed frame one level up is correct, and there is no answer to it beyond noting that this is a limit on inquiry generally, not a defect specific to supply planning. What continuous, provenance-bearing intake buys is narrower: the gazette notice gets logged with source and timestamp whether or not today's frame thinks it matters, the contradiction with the landed-cost assumption becomes queryable rather than buried, and the moment of override — plan changed, forced by filing X, dated Y — is recorded rather than reconstructed weeks later by a planner working backward from a bad invoice.
That is a difference in what disconfirmation is allowed to do, not a promise of neutral sight. No dashboard, however wide, sees the tariff notice as evidence unless something already granted "tariff notice" the status of a thing worth seeing. The honest claim is that the frame can be made to answer for itself, continuously, with a record of what forced it to change — and that this is the last structural move available on the intake axis, not the last word on judgement, which remains, as it always was, a planner's.