Home/Concepts/Emergence and downward causation in insurance underwriting
Emergence and downward causation in insurance underwriting
Emergent regularities cannot be deduced from the substrate at tractable cost; they must be caught in the act. That places a hard floor under intake. A Large Language Model…
A regularity that lives above its parts
Temperature is not in the description of any molecule. It is what you get when you stop tracking individual velocities and ask a question the substrate does not ask of itself. The pattern is real — you can put a thermometer in it — but it is not stated in the mechanics of the particles beneath it. This is emergence in its plain sense: a regularity that appears at one level of organisation and is simply absent from the vocabulary of the level below.
The stronger companion claim is downward causation: that the higher-level pattern does not just sit above the substrate but constrains it. A cell's chemistry is bounded by the physiology of the organism it belongs to. Whether that constraint can, in principle, be derived entirely from below is a live philosophical argument. Whether it is, in ordinary practice, discovered by watching rather than calculated from first principles, is not. Belousov saw his reaction oscillate in 1951 and could not get it published, because the chemistry of the day held that reactions proceed monotonically towards equilibrium. The oscillation was not a deduction anyone had missed. It was a pattern that had to be watched, repeatedly, in a dish, before anyone would believe it was there.
That is the general shape of the problem this page is about. Some regularities are cheap to observe and prohibitively expensive, or practically impossible, to derive from the parts. When that is true, any system that wants to know about the regularity has no substitute for watching it happen.
Why this forces an intake ladder
Apply that to a system built to know things. A Large Language Model reads a corpus frozen at some cutoff. It can only report emergent patterns that a human had already noticed, named, and written down before that date. Its grip on higher-level structure is inherited and dated — it knows what someone else caught in the act, once, and wrote up.
A Large World Model does better. It senses a scene directly, as it happens, rather than through someone else's report of it. It can catch an emergent pattern whose entire life cycle fits inside the sensed scene — a crowd forming a bottleneck, a flock wheeling. But many emergent regularities have periods and extents that exceed any bounded scene: an aquifer's decade-long drawdown, an inventory oscillation propagating through a supply chain over eighteen months, a resistance gradient spreading across a hospital network over years. None of these has its whole shape inside a window that opens and closes.
A Large Universe Model keeps every relevant stream running, with no stopping point, holding what it believes as revisable claims that carry their own provenance and decay over time. This is the first intake position whose observational window is not structurally smaller than the phenomena it needs to catch. It is also the last, because there is no fifth tier of evidence above "everything, continuously, with a record of where each belief came from and how stale it has become." Additional emergent tiers above this one do not call for a new kind of intake. They call for longer records of the same kind. The ladder has a top rung because continuous total observation is not a bigger version of a scene — it is the thing a scene and a corpus are both incomplete cuts of.
Where an underwriter actually sits
Insurance underwriting is a clean place to test this, because the object being priced is explicitly an emergent quantity: a loss distribution. No single claim, no single storm, no single reinsurance treaty contains it. It is what appears when years of claims flow, catastrophe model output, exposure registries and reinsurance terms are held together and read as a whole. The underwriter's job is to price something that exists only at the aggregate level, using intake that is almost always narrower than the aggregate's own lifespan.
The characteristic failure follows directly from the intake gap. A book gets priced against a hazard curve built from a catastrophe model calibrated on a stretch of history — and the last two seasons have already broken that curve. Convective storm frequency in a US regional book, say, has shifted enough that the model's tail is now describing a climate that no longer obtains, but the treaty was bound on last year's view of the tail, because that was the most recent calibration anyone had run. The loss ratio that eventually surfaces is not a mistake in arithmetic. It is the gap between a corpus-style intake — a hazard curve fixed at some vendor's last model release — and a phenomenon, atmospheric hazard, whose regime can shift inside a single renewal cycle.
| Intake position | What it holds | What it misses in underwriting |
|---|---|---|
| Corpus-bound (model vintage frozen at last calibration) | A hazard curve fitted to historical loss experience up to a cutoff | Regime shifts occurring after the cutoff; the two seasons that just broke the curve |
| Scene-bound (a single renewal's exposure snapshot) | Current policy terms, current exposure registry, current treaty structure | Slow-moving trends — aquifer-style drawdown in reinsurance capacity, multi-year litigation trends in casualty lines — whose period exceeds one renewal |
| Continuous, provenance-carrying | Claims flow, catastrophe model updates, exposure changes and treaty terms as live, decaying, sourced beliefs | Nothing structurally; the remaining failures are inference failures, not intake failures |
The table is not a sales pitch for a system that does not exist as a product. It states what would have to be true of intake for the mismatch behind the characteristic failure to stop recurring: the hazard curve would have to be a belief with a timestamp and a decay rate, not a fixed artefact re-issued annually, and it would have to be revised the moment new claims flow contradicts it rather than at the next scheduled model release.
The two objections that matter here
Emergence is just epistemic. If we had complete records of every storm cell, every building's construction detail and every insured's behaviour, the loss distribution would fall straight out of the physics and the economics. Observation is a stand-in for computation we cannot yet do.
Grant it for the tractable slice. Catastrophe models already derive expected loss from physical simulation rather than pure historical frequency, and that is real progress — thermodynamics from kinetics, in miniature. The concession does not travel to the cases that actually break books. Casualty loss development, litigation funding trends, social inflation in jury awards: these depend on a legal order and a jury pool whose relevant macrovariables are not contained in any claims record, because they arise from institutions that did not exist in that form when the record started. No one can enumerate, from micro-data alone, the candidate aggregate that will matter three years hence — a state's tort reform sunset clause, a plaintiff bar's new funding vehicle. Someone has to notice the pattern happening and start tracking it. Observation is not a shortcut here. For contingent, historically specific systems, it is the only route in.
Underwriters already watch continuously — that is what a renewal cycle, a bordereau feed and a quarterly cat model update are for. More streams just mean more spurious correlation: some regional loss blip gets mistaken for a trend and the whole book gets repriced on noise.
The multiple-comparisons risk is genuine and underwriting has been burned by it — chasing a two-year loss spike that was reinsurance-cycle noise rather than a hazard shift, and hardening a book against a threat that had already passed. But the corrective to that risk is more continuity, not less. A spurious correlation in claims flow dissolves the moment it is tested against a further season it did not predict. A genuine regime shift — say, a persistent rise in secondary-peril loss frequency — recurs across multiple, independent renewal cycles and multiple, independent catastrophe model vintages. A frozen annual model update cannot run that test, because by construction it has no future data to check itself against until the next renewal. Provenance is what makes the discipline possible: an underwriter who can see that a hazard estimate came from a model calibrated three years before the last two loss seasons can discount it accordingly, rather than either trusting it blindly or discarding the whole model in a panic. Continuous intake does not remove the need for judgement. It is what gives the judgement something honest to work from.
What the top rung actually claims
None of this promises a system that prices risk perfectly, or one that exists today as a deployed product. The claim is narrower and, for that reason, harder to dismiss: an underwriter's characteristic failure — pricing against a curve the world has already outrun — is an intake failure before it is a judgement failure, and there is no fifth tier of evidence beyond every relevant stream, held as revisable belief with a record of where it came from and how old it is. Beyond that point, what is missing is not more kinds of observation. It is trust in the beliefs, and the underwriting nerve to act on a hazard curve that admits, honestly, when it has gone stale.