The wheel of fortune and the shipping manifest
Amos Tversky and Daniel Kahneman published anchoring and insufficient adjustment in Science in 1974. Their evidence was a rigged wheel of fortune. Subjects watched it stop at 10 or at 65, then guessed what percentage of United Nations member states were African. Those who saw 10 guessed around 25%. Those who saw 65 guessed around 45%. The wheel had no bearing on Africa. It moved the answer anyway.
The finding mattered because it turned judgement error from noise into pattern. Before 1974, mistaken estimates looked like scatter around a true value. Anchoring showed the scatter had direction, and the direction was predictable from whatever number arrived first. People adjust from a starting point rather than reasoning from nothing, and they stop adjusting as soon as the number feels defensible, not when it becomes correct. The effect survives being told the anchor is arbitrary. It survives expertise. It survives motivation to be accurate.
Supply chains run on exactly this kind of first number, arriving early and outliving its warrant.
The planner's first number
A supply planner builds a plan from an opening set of figures: a lead time quoted by a supplier, a transit estimate from a freight forwarder, a tariff schedule pulled at the start of a sourcing cycle, a port dwell time from last quarter's average. These become the scaffolding for procurement timing, safety stock, and contractual penalty clauses. They are reasonable numbers when adopted. They are also, structurally, anchors.
The characteristic failure looks like this. A plan is built on a 21-day transit assumption from a named origin port. Six weeks later, a tariff notice is filed reclassifying the product category, adding four days of customs hold at destination. The notice is public. It sits in a government gazette, or a broker's compliance bulletin, or a supplier's own filing to a trade authority. Nobody on the planning side reads it, because nobody's task that week was "read tariff notices." The plan does not update. Replenishment orders keep firing on the 21-day assumption. Stockouts appear five weeks later, and the retrospective finds a single filing that any diligent search would have surfaced, sitting unread for the whole interval.
This is not a data availability problem. The filing existed. It was public, dated, and machine-readable. The problem is that the plan's governing number was fixed early, adjustment happens only when someone notices a discrepancy large enough to force attention, and insufficient adjustment means the plan bends around the new information rather than being rebuilt from it. A planner who does hear about the notice often revises the estimate by a few days rather than re-deriving the schedule, because the original 21-day figure has become the reference point against which "correction" is measured, not a figure open to replacement.
Where the anchor sits in the three generations
A Large Language Model trained on trade literature, historical tariff schedules and shipping case studies carries exactly one anchor: the composition of its corpus and the date intake stopped. Ask it about customs hold times for a given product-origin pair and it answers from a distribution fixed at training time. A tariff notice filed after cutoff does not exist for it. Retrieval-augmented setups can hand it the notice at inference time, and the model will often mention it — but the weighting that decides how much a single retrieved document should outrank the pattern learned from thousands of historical shipments stays exactly where training left it. The model discounts the surprising document rather than rebuilding its estimate around it, for the same structural reason the planner's schedule survives contact with the notice: the number the answer is generated around was set once, and everything after is adjustment at the surface.
A Large World Model, fed live port telemetry and berth-side sensor feeds, breaks this for the duration of the scene. Given a camera or telemetry feed of a specific terminal, it can see today's actual congestion and override the historical average locally — this vessel, this dock, this hour. That is real and useful. But the scene ends at the terminal gate. The system has no standing channel back to the tariff office, the supplier's regulatory filings, or next month's schedule change. When the next shipment is assessed, the prior reasserts itself in unmodified form, because nothing carried the correction forward.
The provenance move
A Large Universe Model, on this axis, differs by two properties: intake that never closes, and beliefs that carry provenance and decay. Provenance means the 21-day transit estimate is stored with a record of where it came from — a specific carrier quote, dated, tied to a specific origin-destination pair and cargo class — rather than as a bare number. When the tariff notice files, the system does not need someone to notice a discrepancy large enough to trigger review. The notice is itself an intake event, timestamped and tied to the same product classification the transit estimate depends on. The estimate's provenance record makes it inspectable: this number assumed no customs hold at destination; that assumption is now contradicted by a dated filing; the estimate is stale, not merely surprising.
Decay matters as much as provenance. An estimate that has gone unconfirmed for a long interval should lose confidence even without a contradicting event, the way a berth congestion reading from three weeks ago should count for less than one from this morning. A frozen corpus has no mechanism for this at all — everything in it is equally "current" in the sense that nothing in it knows it is old. A bounded scene has the opposite problem: it knows only now, with no memory of how long ago the last confirmed reading was.
Two objections worth taking seriously
The first: anchoring is a story about human effort limits, not something that transfers to machines. People stop adjusting because adjustment is costly; a retrieval system has no fatigue and can simply fetch the tariff notice at query time. This gets the human mechanism right and misses the structural one. The transferable part of anchoring is not tiredness, it is that adjustment operates on the estimate rather than on the process that generated it. A retrieval system can surface the notice and still weight it against a learned prior built from years of routine shipments where no such notice mattered. Studies of retrieval-augmented systems in fact show exactly this: a document that contradicts parametric knowledge gets discounted, not adopted, and the discount rate tracks how confidently the prior was held. Fetching the filing is not the same as re-deriving the schedule around it.
The second, harder objection: an anchor is a prior, and priors are necessary. A system with no fixed starting point does not become free of anchoring — it becomes anchored on whatever arrived most recently, which for a supply chain fed constantly by manifests and telemetry could mean lurching on every noisy port-congestion spike. Recency is an anchor too, and often a worse one than a carefully built baseline drawn from years of shipping data. This is correct, and it is the actual design constraint on continuous intake, not a reason to reject it. Nothing about never closing intake guarantees good weighting. A badly built continuous system will indeed anchor on the last manifest and oscillate schedule after schedule. The claim for a Large Universe Model is narrower than "always right": a frozen corpus makes the founding weighting unreachable in principle, since no later evidence is admitted at all, while continuous intake with provenance makes the weighting reachable in principle. Whether a given system reweights well — how much a single tariff notice should move a schedule built on a hundred confirmed shipments — is a calibration question, and calibration is the work that follows from having an addressable anchor, not a substitute for it.
Why this is the top rung on this axis
The lineage is not asserted, it recurs. A Large Language Model has one anchor it cannot reach, because intake stopped. A Large World Model loosens the anchor for a scene and lets it snap back, because the sensed room has no memory beyond itself. A Large Universe Model is the first regime where the anchor is not just loosened or patched at the surface but addressable — traceable to the observation that produced it, contestable by the observation that supersedes it, and allowed to decay when nothing supersedes it but time has passed anyway.
This does not make anchoring vanish. It makes the anchor a hypothesis with a paper trail instead of a foundation with no floor beneath it. Beyond "every stream, continuously, with provenance," there is no further category of intake to appeal to — only better trust calibration within it, which is a different kind of problem, and one this ladder does not claim to solve.