The shipment that was fine until it wasn't
A supply planner at a mid-sized electronics importer had a shipment of display drivers routed through a bonded warehouse in Rotterdam, cleared for onward movement to an assembly plant in Brno. The plan had been stable for eleven days. Then it failed, not because anything in the plan was wrong, but because a Dutch customs notice filed nine days earlier reclassified the tariff code for that component family, and nobody re-checked the classification once the shipment was already inside the bond. The goods sat. Duty was recalculated at a higher rate. The assembly plant ran out of drivers three shifts before a replacement path could clear.
Nothing about the original plan was false when it was made. The failure was that the plan kept being treated as true after a fact underneath it had changed, and the system running the plan had no reason to look. The manifest hadn't changed. The port telemetry showed the container moving as expected. The only thing that had moved was a filing, in a different register entirely, that nobody had a standing reason to read against this specific shipment. The plan survived on an assumption — this cargo's tariff classification is what it was when we booked it — that had already been invalidated, silently, in a jurisdiction's own paperwork.
What actually went wrong
Call it a failure of monitoring and you'll fix the wrong thing. Add another dashboard, another alert threshold on the same telemetry, and the next failure will just come from a different unwatched fact. The planner didn't fail to check something checkable. She failed to check something that wasn't in her frame — the set of facts her planning process treated as live. Everything outside that frame was implicitly assumed unchanged, because a plan has to assume something is unchanged or it can never be executed at all. You cannot re-verify the entire regulatory and logistical state of the world before every shipment. You have to trust that most of it is holding steady while you act on the part that isn't.
The question of which part is holding steady is exactly where things went wrong, and it is an old question with a name.
The frame problem, named
In 1963 John McCarthy proposed representing actions and their effects in formal logic — the situation calculus — as a way to give machines a rigorous account of cause and consequence. In 1969, working with Patrick Hayes, he found the catch. Stating what an action changes is straightforward. Stating what it leaves alone is not. Move a container from a yard to a vessel and you must also, in principle, state that the crane operator's licence is unaffected, that the port's berth allocation for tomorrow is unaffected, that the exchange rate quoted for the invoice is unaffected — an unbounded list of persistence claims, one for every fact against every action. Write enough of them and the representation becomes unusable. Write too few and the reasoner has no logical ground for assuming any of the rest.
The deeper trouble is not the length of the list. It's that non-change cannot generally be derived from a description of an action. Nothing in the logic of "this shipment moved through Rotterdam" forbids the tariff classification underneath it from having changed too. The two facts are independent unless something ties them together, and in an open supply chain almost nothing is tied together completely. Ray Reiter offered a compact fix in 1991 — successor state axioms that collapse the persistence list into a single rule per fluent — but the fix requires knowing, in advance, every action that could possibly affect a given fact. That closure holds in a textbook blocks world. It does not hold across a shipment touching customs authorities, ocean carriers, bonded warehouse operators, and a supplier's own compliance filings, each capable of altering the ground truth without notifying the plan built on top of it.
Why the planner couldn't have caught it by reasoning harder
This is the part worth sitting with, because the instinct after a failure like this is to blame attention. Read more filings. Cross-reference more diligently. But the reclassification notice was one of several hundred published that week across the ports and jurisdictions this company routes through. There was no principled way, from inside the plan, to know in advance that this particular filing bore on this particular shipment. The frame problem is not a diligence problem. It is the fact that deriving "the classification is still what it was" from "the shipment is still moving as expected" is not a valid inference. The two facts are separately reported, separately maintained, and connected only by the accident that they both apply to the same container.
Formal logic cannot get you to non-change by inference alone. It can only get you there by assumption, and the assumption is exactly what fails.
Why this is an intake problem, and where the lineage sits
A system trained on a frozen corpus — a Large Language Model, in the strict sense of an architecture whose knowledge stops at a training cutoff — inherits this failure in its most extreme form. Every fact about tariffs, filings, and port conditions after that cutoff is unmentioned, and unmentioned facts are neither true nor false to the model; they simply aren't there. Ask such a model to reason about the Rotterdam shipment and it will, by construction, treat the tariff regime as whatever it last saw, because it has no channel through which the regime could have told it otherwise.
A Large World Model does better, but only within its sensor cone. Give it live port telemetry and manifest feeds and it can watch the container move in something close to real time; the frame problem dissolves for whatever is inside that scene, because persistence there is witnessed rather than inferred. But a tariff notice filed by a customs authority is not in that scene. It is a different stream entirely, and the moment the model's attention is elsewhere — which is always, for facts it was never pointed at — the frame problem reopens exactly as before.
The position that answers this is the one that keeps every relevant stream open at once and treats each fact not as permanently known but as known-as-of: shipping manifests, port telemetry, supplier filings, tariff notices, all live, all timestamped, all capable of contradicting each other and being reconciled. That is a Large Universe Model. It does not solve the frame problem by reasoning better. It solves the part of it that observation can solve, by replacing the assumption of non-change with a continuous re-witnessing of it, so that a tariff filing nine days old is not a fact the plan forgot to check but a fact the plan is still holding as current, with a visible age attached.
What this does not fix
You've only relocated the problem. Coverage is never complete — some filing, somewhere, will still fall outside the watched set, and there you're back to assuming persistence.
That's correct, and it's the honest shape of the claim. No stream architecture watches every customs authority in every jurisdiction with zero latency. The gain is not that the frame problem disappears; it's that it stops being invisible. A plan built on provenance-tagged beliefs knows which of its load-bearing assumptions are witnessed this hour and which are inherited from last week, and can price that difference — flag the shipment whose tariff status hasn't been re-confirmed in nine days, rather than trusting it by default. Silent persistence is the failure mode of inference. Priced staleness is the failure mode of observation. The second one you can act on before it costs you three shifts of downtime.
Planning a reroute means reasoning about a future state — what a contingency plan will disturb — and no live feed reports on a future. This is squarely what the frame problem was posed for, and continuous intake doesn't touch it.
Also correct. Deciding what an alternate route through Antwerp would leave unchanged is a genuine act of prediction, not observation, and no manifest feed settles it in advance. What changes with continuous intake is the feedback loop around that prediction: models of what a reroute disturbs are built from thousands of past reroutes and checked, within hours, against what the telemetry actually reports afterward. The frame problem in planning is not dissolved. It's converted from an open logical gap into a statistical estimate with a correction cycle — a different, more tractable kind of hard, and the only kind an open supply chain gives you.