The recursion Kálmán published
Rudolf Kálmán's 1960 paper, "A New Approach to Linear Filtering and Prediction Problems," did something the electrical engineering establishment initially found unremarkable and mechanical engineering, oddly, did not. It carried a hidden state through time using only two objects: a best estimate of that state, and a covariance describing how wrong the estimate might be. Nothing else. No archive of past measurements, no re-fitting.
The recursion runs in two steps, forever. Predict: push the estimate forward through a dynamics model — the physics, or the assumed physics, of how the state evolves — and let the uncertainty grow, because time passing without new information always costs you certainty. Update: a measurement arrives, noisy, and you weigh it against the prediction in proportion to their relative reliabilities. A trustworthy sensor gets a large gain; a noisy one gets a small gain. The estimate shifts, the covariance shrinks, and the measurement itself is then discarded. It survives only through its effect on the estimate and the covariance. The past is never re-read.
Kálmán proved something stronger than convenience: under his assumptions — linear dynamics, Gaussian noise, a correctly specified model — this recursive estimate is optimal. No amount of batch reprocessing over the full measurement history can beat it. That is the result that matters for everything downstream of it, including systems that have never heard of state-space models.
What the recursion fixes about intake
If a filter's estimate carrying forward all the information a full reprocessing would have given you, then continuous intake does not require unbounded storage. A sufficient statistic — estimate plus covariance — stands in for all history. That is a structural claim about how much a system needs to keep, not just an engineering trick.
It also fixes the shape of the terminal case on any axis that measures how a system takes in the world. The Large Language Model is a single batch estimate over a corpus assembled once and frozen at a cutoff. It is the least-squares fit Gauss would have recognised, run at planetary scale, never revised. The Large World Model runs Kálmán's actual predict-update cycle, but only for the duration of a bounded scene: sensors stream, a state is tracked, and when the episode closes the covariance is thrown away with it. The Large Universe Model is the same loop with the scene boundary removed and the stopping condition deleted. Persistent state, per-source noise models doing the work that provenance does elsewhere, corrections applied forever across every stream running at once.
Once a system holds a calibrated belief, accepts a measurement with a stated noise model, and corrects in proportion to relative reliability, there is no further move to add on the intake axis. You can widen the state vector, sharpen the dynamics model, admit more sensors, tighten the noise estimates. All of that is more of the same three moves. There is no fourth. That is why the Large Universe Model is not a new category of intake sitting beyond the Large World Model — it is the World Model's own sensing loop, made unbounded in time and source. That is where the ladder ends.
The desk that has to run this loop by hand
Energy trading is where this stops being a diagram and starts being a Tuesday morning. A desk quant running a position on regional power or gas is tracking a hidden state — true available transmission capacity, true plant availability, true weather-driven demand — through measurements that never directly reveal it. Grid telemetry arrives at five-minute or sub-minute resolution and is itself noisy, subject to sensor drift and reporting lag. Outage notices arrive as discrete, asynchronous text, sometimes hours ahead of physical effect, sometimes retracted. Weather reanalysis updates on a slower cycle but revises the recent past as satellites and models improve. Regulatory filings — a transmission constraint lifted, a curtailment order rescinded, a capacity market rule amended — arrive on their own clock entirely, set by hearings and gazettes, indifferent to the market's need to know.
None of these streams is the state. Each is a measurement of it, with its own noise characteristics, and the desk's job is exactly Kálmán's job: predict the state forward using a model of grid and market dynamics, then correct that prediction against whatever arrives, weighted by how much that source deserves to move the belief.
The characteristic failure on this desk is precise and recurring: a position is held against a constraint that was lifted overnight. The quant's model predicted forward correctly from yesterday's confirmed constraint. The regulatory filing releasing that constraint arrived at 11pm, low-volume, unglamorous, buried in a docket update rather than a headline feed. Nothing in the desk's overnight processing treated it as a high-gain measurement, because nothing recognised it as a measurement of the state at all. The estimate did not update. The position, correct against yesterday's world, is wrong against this morning's, and it is wrong specifically because an old belief was held past its warrant rather than because new information failed to exist. The information existed. The gain on it was effectively zero.
This is a covariance failure in the precise technical sense, not a metaphorical one. The desk's implicit noise model treated regulatory filings as slow and low-priority, appropriate for the majority of filings, which are indeed procedural. But the noise model was not conditioned on content — it could not distinguish a routine docket entry from one that changes the feasible region of every position on the book. A Large World Model, scoped to the trading session, would have thrown away yesterday's covariance at the close and rebuilt from scratch this morning, which is no better: the constraint-lifted fact needs to be carried across the boundary, not rediscovered. What the failure calls for is exactly what the Large Universe Model formalises — a per-source noise model tuned to content and provenance, not just to channel, so that a regulatory filing capable of moving the feasible region gets treated, structurally, as a high-gain measurement the instant it lands, at 11pm or any other hour, with no session boundary to hide behind.
Two objections the desk cannot avoid
The first objection is the sharpest one in this domain, because it describes a failure the desk has actually lived through. Kálmán's optimality holds only when the dynamics model is approximately right and the noise statistics are approximately honest. Off that footing, filters diverge with confidence: the covariance shrinks while the true error grows, because the filter is trusting a model it should be doubting. A desk running continuous intake against a misspecified grid model — one that under-weights the probability of sudden transmission topology changes, say — will produce a stream of increasingly confident, increasingly wrong positions. That is arguably worse than a static model reviewed weekly by a human who at least knows to distrust it.
The answer is not faith in the loop but instrumentation of it. Filters predict their own residuals, which means divergence is detectable from inside the system: normalised innovation squared, tested against chi-square bounds, flags exactly the situation where measurements are landing further from prediction than the stated covariance should allow. A desk running this discipline treats a string of surprising outage notices not as bad luck but as a signal that its process noise is mis-set, and adapts it. A frozen weekly model has no equivalent alarm. It cannot know it is stale. Continuous intake, done with this machinery, makes the failure legible in a way batch estimation structurally cannot.
The second objection is about depth rather than direction. Filtering is lossy by construction — the forward pass discards each measurement after using it, and a backward smoothing pass over the full record strictly dominates it at every interior point. Retrospective reanalysis, done by risk teams reconstructing exactly why a position failed last Tuesday, genuinely does beat the live filtered estimate, and no desk should pretend otherwise.
Streaming is just filtering with worse accuracy than the archive would give you if anyone bothered to look back.
Granted. But smoothing requires retained observations, which is a claim about storage, not a rival claim about intake. A desk that has been running continuous assimilation across telemetry, outages, weather and filings can smooth backwards over its own retained streams whenever it needs to, which is exactly what post-trade forensic review does. A desk that treats its model as fixed between quarterly recalibrations has nothing to smooth over except what it happened to log. Retrospection is a capability that unbounded intake enables. It is not an alternative to it.
Where the analogy stops being free
The formal optimality result does not survive contact with regulatory filings and discrete outage events, and it would be dishonest to claim it does. What survives is the architectural shape: a persistent estimate, corrected by new measurements in proportion to stated reliability, with no session boundary and no terminating condition. That shape is what a desk quant is actually running, whether implemented as a literal extended Kalman filter over a state-space grid model or as a looser system of weighted belief updates over disparate feeds. The optimality is a bonus available only in the linear corner. The discipline — predict, measure, weigh, correct, never stop — is the part that has to hold regardless, and it is the part that failed the night the constraint was lifted and nobody's noise model was listening.