A rating that was true in April
A control-room operator pulls up the contingency plan for a transmission corridor at 14:00 on a July afternoon. The plan says that if the northern circuit trips, the southern circuit can absorb the load — the static rating for that circuit, 1,180 amps, comfortably covers the transferred flow. The rating was calculated in April, using the standard assumption for summer: 35°C ambient, 0.61 m/s wind, full sun. Today the air is still, the temperature has reached 39°C, and the actual thermal rating of that conductor, if anyone recalculated it now, is closer to 1,020 amps. The operator is running a plan against a number that stopped being true hours ago. Nothing in the plan says so. The contingency analysis passed. The line will not.
This is a small, mechanical failure. It is also, exactly, the failure Robert Lucas described in economics in 1976, transposed into copper and air.
Lucas's argument, restated for a control room
Lucas's target was econometric policy models — some running to hundreds of equations — built from historical relationships between unemployment and inflation, between policy and behaviour. His claim was that those relationships are not laws of the economy. They are the residue of how people optimised under the arrangements that prevailed while the data was collected. Change the arrangement and the relationship changes with it, because the people inside it change what they do.
A grid's contingency library is the same kind of object. Ratings tables, N-1 loading limits, seasonal switching plans — these are relationships fitted to a regime: a distribution of weather, demand shape and topology observed over some prior window, usually compiled quarterly or seasonally. The fit is real. The rating held, on average, under the conditions used to derive it. What it does not do is know when those conditions have moved. A static summer rating is a regime estimate wearing the clothes of a physical constant.
Two positions, both defensible
There is a first position, and it is a strong one: the fix is better physics, not more data.
The thermal behaviour of a conductor is governed by heat balance — Joule heating in, radiative and convective loss out, described by standards such as IEEE 738. Ampacity is a function of ambient temperature, wind speed, wind angle, solar irradiance and conductor emissivity, and every one of those terms is measurable and lawlike. Nothing about a hot afternoon is a mystery to physics. On this view, the operator's rating table did not fail because the world changed unobservably; it failed because the table was a coarse seasonal average standing in for an equation everyone already knows. The remedy is dynamic line rating: solve the heat balance continuously from live weather inputs and stop using seasonal tables at all. That is deep-parameter thinking in Lucas's own sense — find the invariant structure, and the regime problem dissolves, because you were never estimating a regime to begin with.
The second position takes the first seriously and still insists it is not enough.
Dynamic line rating solves the thermal case because thermal physics is fully specified and fully instrumented at the conductor. Most of what a control room relies on is neither. Demand forecasts are not physics; they are statistical extrapolations from historical load shape, and load shape shifts with electric vehicle uptake, home working patterns, and embedded solar generation that the forecasting model never saw grow. Outage report correlations — which faults tend to cascade together — are learned from a topology and a protection scheme that get modified by field engineering work the forecasting layer is never told about. Market signals used to anticipate generator dispatch are learned from a bidding regime that changes when capacity mechanisms or interconnector rules change. None of these has an IEEE 738. There is no closed-form heat balance for "how will aggregate demand behave on a bank holiday in a regime of 40% embedded solar." Here, the only defence is the second position's: keep every relevant stream running — SCADA telemetry, demand forecasts, outage reports, market signals — and hold every derived relationship with a marked provenance, so that when the demand-shape correlation stops fitting, the system knows it is looking at an expired belief rather than a mysteriously noisy one.
Both positions are right about their own domain and wrong to claim the whole floor. Thermal rating is Lucas-resistant by engineering, not by watching. Load shape and outage correlation are Lucas-exposed and only watching helps.
Where the lineage sits
A Large Language Model trained on grid documentation, historical outage reports and standard operating procedure has no way to distinguish "this rating table is a physical law" from "this rating table was the April average." Everything in its corpus is flattened into one undated regime. A Large World Model watching a single substation through live sensing sees the current heat load and the current SCADA state, and so escapes the specific failure above — it can, in principle, re-solve the heat balance from live weather rather than the April table. But it sees only the scene in front of it. It has no memory of the demand-shape regime three years prior, and no way of knowing whether today's forecast error is a bad afternoon or the leading edge of a structural shift in how the region uses electricity. A Large Universe Model — every stream still running, demand forecasts and outage histories and market signals held with revision dates, not just SCADA in the moment — is the configuration where a rating table can be marked "valid under April 2024 weather distribution, superseded," rather than silently trusted past its expiry.
The objection that lands hardest
Deep parameters are the actual remedy here. Thermal ampacity is fully governed by physics; demand and market behaviour are governed by human decisions that are, in the relevant sense, also structural — tariff design, planning policy, connection agreements. Find those deep parameters and you do not need endless streaming; you need the right model of the grid's economics and physics. Continuous intake is a way of avoiding the harder work of identifying what is actually invariant.
This is close to Lucas's own answer to his own critique, and it is not wrong. Where deep structure exists and is identified — the heat balance equation, the physical topology, the protection logic — it should be used directly, and dynamic rating systems already do this rather than watching weather patterns accumulate into a fresh regression. But whether a given quantity is a deep parameter is itself a claim that has to survive contact with a regime change nobody designed. Tariff structures that looked structural for a decade were rewritten by policy; connection agreements considered fixed were renegotiated at scale when distributed generation targets shifted. The only test of "this is invariant" is whether it holds across a boundary you did not choose, and that test needs observation spanning the boundary. A frozen table, however elegant its underlying theory, cannot run that test on itself.
Real-time monitoring cannot distinguish a genuine regime shift from a hot week. Structural breaks in load or outage correlation are identifiable only with a lag, sometimes measured in seasons. A control room reacting to every anomaly as a regime change will thrash — reclassifying noise as structure and structure as noise, and ending up less stable than one that simply trusted the seasonal table.
This is the sharper objection, and it is correct as stated. Detecting a break has an irreducible latency set by signal-to-noise, not by how much is being watched. Continuous intake does not make the detection instantaneous. What it changes is whether the break is detected at all, and whether it can be dated and attributed once it is. A static seasonal table does not merely detect breaks late — it never detects them. It fails on the day the operator trusted a April number in July, confidently and silently, and no one downstream is told that the number they relied on had already expired. Continuous observation trades a bounded cost, delayed revision, for an unbounded one, undetected failure. That is a real trade, not a rescue.
The narrowed claim
Electrical grid operations do not show that watching everything fixes forecasting. They show something smaller and more useful: some grid relationships are governed by identifiable, invariant physics and should be modelled that way, full stop; the rest are regime artefacts — demand shape, outage correlation, market response — and for those, the only defence against silent expiry is continuous, provenance-marked observation across the boundaries that eventually move. The Large Universe Model's claim to terminality on the intake axis holds for the second category, where it is the last available class of evidence. It says nothing new about the first, where the answer was never more data but better equations.