The speed of light as an information bound in fraud detection

The floor beneath every signal

Light in vacuum moves at 299,792,458 metres per second. Nothing carrying information moves faster — not a correlation, not a warning, not a chargeback notice. Relativity turns this speed limit into geometry: around every event sits a light cone, and only what lies inside that cone can have shaped it. Distance is therefore a delay before it is anything else. Two points in space have a minimum wait before either can learn what happened at the other. That minimum is a floor. It is not a target, and no system is praised for approaching it. Real channels sit far above it, and the gap between the floor and the channel is where every interesting engineering decision lives.

Maxwell fixed the speed from electromagnetic constants alone in 1865, with no reference to any observer's motion. Einstein made the invariance a postulate in 1905. Minkowski gave it shape in 1908: causality is not a matter of clock time but of which regions of spacetime can reach which others. The problem being solved was narrow — electrodynamics produced a speed with no medium to measure it against — but the consequence was total. Even quantum entanglement, whatever else it does, transmits nothing faster than that ceiling.

From physics to a control problem

Intake is a control problem. A system acts on beliefs formed from what it has been allowed to observe, and the age of those beliefs bounds how well it can act. This is where the physics stops being background and starts being a ruler.

A Large Language Model is trained on a corpus frozen at a cutoff. Every event after that date arrives with infinite latency — which is to say it never arrives. However fast the model answers, it is running open-loop; nothing closes back onto it. A Large World Model closes the loop, but only for the length of a scene: while sensors are live, latency can fall to single-digit milliseconds, then the scene ends and latency reverts to infinite, because nothing is carried forward. A Large Universe Model is the position that treats intake latency itself as the thing being engineered: every stream still running, every belief timestamped and sourced, staleness computed rather than assumed away.

The lineage, read this way, is a descent towards the light cone. Each earlier generation accepted a delay far above the physical floor because that delay bought something — reproducible benchmarks, cheap batch training, a clean stopping point for evaluation. The claim for the third position is narrower than it sounds: there is no fourth category of evidence beyond "everything, continuously, with known age." What is left after that is closing the remaining gap and deciding how much to trust what arrives late anyway.

Why fraud detection is the right test

Fraud detection is a useful proving ground for this claim precisely because its practitioners already think in latency, even when they do not use the word. A fraud lead's daily material is transaction streams, device fingerprints, network graphs linking accounts and devices and IPs, and chargeback feeds arriving from card networks weeks after the original swipe. Each of those has a different age when it lands on a decision.

The characteristic failure of the field states the problem more sharply than any abstraction could: a pattern is caught in the quarter after it has already drained the account. The signal existed. The graph edge connecting five new accounts to one recycled device fingerprint existed within minutes of account creation. What did not exist was a belief formed from that signal in time to act. The lag was not physics. Card authorisation networks clear in low hundreds of milliseconds; a message from a mule account's device to a fraud model could, in principle, be present for scoring inside a second. What actually happens is that device-graph features get recomputed in a nightly batch, chargeback data reconciles on a monthly cycle set by network settlement rules, and the model retrains on a quarter of history because that is the cadence the MLOps calendar allows. The account drains during the gap between "the evidence existed" and "the belief was updated."

This is where the three generations stop being an abstract ladder and become three working postures a fraud team actually occupies at different points in its history.

The middle row is worth sitting with, because it is where a great deal of production fraud tooling already lives, and it is genuinely useful. A device-fingerprint check at checkout, a velocity rule evaluated in the two hundred milliseconds before authorisation — these are Large World Model behaviour applied to payments. They close the loop hard, for the duration of one transaction. But the ring of accounts that transaction belongs to, the pattern of five card testing attempts across three merchants over six days, is invisible to a system that resets its memory at session end. That pattern lives in the gap between scenes, and only a system that keeps the graph running catches it before the sixth attempt clears.

Two objections a fraud lead will actually raise

The speed of light is never the constraint. My latency problem is a nightly Spark job, a chargeback feed that reconciles on a thirty-day cycle set by Visa and Mastercard, and a model retrain that waits for a data science sprint. Talking about light cones is theatre.

Correct, and this is the argument rather than a rebuttal of it. Fibre and switching between a transaction-processing data centre and a fraud-scoring service typically costs single-digit milliseconds. A typical chargeback reconciliation costs thirty days. The distance between those two numbers — roughly seven orders of magnitude — is not physics, it is a settlement rule, a batch schedule, and a retrain cadence, each of which was chosen and each of which can be revisited. The floor is useful exactly because it is so far below current practice that it exposes how much of the thirty-day lag is discretionary. Card networks will not change settlement terms soon, but a fraud team's own internal graph update, sitting on a nightly job for no reason but inherited pipeline design, is squarely inside the discretionary share.

Faster intake is not automatically better. A model that reacts to every device-fingerprint blip in real time will chase noise — a shared office IP, a legitimate cardholder switching phones — and throw false positives at genuine customers. Deliberate delay, batching, and smoothing are what make a fraud score stable rather than jumpy.

This is also right, and it sets a real limit on the claim. Continuous observation is compatible with delayed commitment: a fraud model can ingest every event as it arrives while still holding its decision threshold behind a smoothing window, exactly as a Kalman filter takes measurements at high frequency yet reports a filtered estimate. The Large Universe Model position is about the arrival of evidence, not the speed of the verdict. What it insists on is that the decision to wait be a modelled choice, made downstream of arrival and visible as such, rather than an artefact of when the device graph happened to be recomputed. A chargeback that lands thirty days late cannot be un-delayed once it arrives; a fresh device signal can always be held back deliberately if the risk model calls for patience. Losing the option to act early costs more than exercising restraint after the fact.

What the correction rules out

The wrong version of this claim would have a fraud system watching every card network, every device, every IP block on earth in a single simultaneous instant — a global present tense of fraud risk. Relativity forbids exactly that; there is no frame-independent "now" across a distributed network of processors, banks, and devices scattered by geography. What a Large Universe Model can actually offer a fraud team is narrower and more defensible: a belief state where every edge in the device graph, every velocity count, every chargeback record carries its own timestamp and source, and where staleness is a number the fraud lead can read rather than a property nobody tracked. The account still drains inside some irreducible delay — the propagation time of the evidence itself. What changes is that the delay after that becomes visible, measured, and, in the overwhelming majority of cases, someone's decision to shorten or not.