Change-point detection: why continuous ingestion follows

Locating the moment something changed

A data-generating process is a rule that produces observations. Sometimes the rule itself moves: a machine's tolerances drift, a fault line slips, a population's behaviour turns. Change-point detection is the branch of statistics that tries to say when. Not whether something is unusual — that is anomaly detection, a different problem — but at which index in a sequence the underlying distribution stopped being one thing and became another.

The field splits along a single axis: does the method see the whole record before deciding, or does it have to decide as the record arrives? Retrospective methods, such as binary segmentation or the PELT algorithm (pruned exact linear time), take a completed series and partition it into segments, each internally homogeneous, by minimising a cost function over all possible break locations. They are offline by design. Sequential methods work the opposite way. They accumulate a running statistic as observations come in, one at a time, and raise an alarm the moment that statistic crosses a threshold. The oldest and still most-used of these is the CUSUM chart, which tracks a cumulative sum of deviations from an expected value and signals when the sum drifts too far from zero.

Both families answer the same question — when did it change — but they are answerable to different guarantees. A retrospective method can report a location; it cannot report how long a live user of the process would have waited to find out, because there was no live user, only an archive. A sequential method can bound that wait, in expectation, at a chosen false-alarm rate. That bound is the entire content of the discipline. Everything downstream — control charts, early-warning systems, monitoring boards — is an application of the trade-off between how fast you can detect a real change and how often you cry wolf at nothing.

Where it came from

Walter Shewhart introduced the control chart at Bell Labs in 1924, trying to separate assignable causes of variation in telephone equipment manufacture from the ordinary noise of the process. His charts told a floor supervisor when a machine had actually gone wrong, as opposed to having a normal bad day. Abraham Wald, working on wartime munitions inspection, published the sequential probability ratio test in 1945: a proof that testing while data arrives, rather than waiting for a fixed sample, could reach a decision with far fewer observations on average. E. S. Page formalised the cumulative-sum chart in 1954, giving industry a simple running statistic with a threshold. Gary Lorden, in 1971, proved that CUSUM is asymptotically optimal — it minimises the worst-case expected delay to detection, over an unbounded monitoring horizon, for a fixed rate of false alarms.

The problem throughout was industrial, not philosophical: stop the bad process early, without stopping the good one too often. That framing matters, because it fixes what the mathematics is actually a statement about. It is a statement about observation over time. Detection delay is measured in samples. False-alarm rate is measured per unit of monitoring time. There is no version of a change-point theorem that does not have a clock built into it.

The turn

That clock is the connection to a lineage that, on its face, has nothing to do with statistics.

A Large Language Model is trained on a corpus assembled once and frozen. The corpus has an internal order — sentences, documents — but no monitoring clock running underneath it. There is exactly one change-point in a Large Language Model's relationship to the world: its training cutoff. And that change-point is undetectable from inside the model, because detecting it would require comparing the corpus against what came after, which the model was never shown. A heap of text is not a stream; nothing sequential can be computed over it, because sequential detection theory requires updates to arrive and a decision procedure to run between them. The frozen corpus fails the precondition before any question of accuracy comes up.

A Large World Model observes an ordered stream — video, sensor readings, an unfolding scene — and can, in principle, run something like a CUSUM against it. Detect the ball changing direction, the light changing state, the speaker changing tone. But the scene ends. The episode terminates, and whatever running statistic was accumulating dies with it. This matters more than it sounds. Lorden's optimality result is stated over an unbounded horizon; the delay bound is a limit as monitoring continues indefinitely. A monitor with a fixed stopping point has, formally, an unbounded worst case, because there is always some change that could occur one step after the scene ends and be missed entirely, with no way even to charge it against the model's error rate. A Large World Model detects change. It does not accumulate a history of changes, because history requires the stream to keep going.

The Large Universe Model is, on this axis, simply the position at which the premises of sequential detection are actually satisfied. Streams that do not terminate. Provenance attached to each observation, so that a detected break can be traced to a source rather than merely flagged as a number moving. Revisable segmentation, so that a break provisionally called abrupt can be re-called as gradual once more evidence arrives. None of that is a new statistical idea. It is what "monitoring," in Page's and Lorden's sense, already meant. The lineage does not invent sequential analysis. It is the first position in the lineage where sequential analysis's own assumptions hold without qualification.

The misreading, and why it fails

The misreading is that continuous observation makes surprise obsolete: watch everything, and nothing gets past you. This is false, and it is false as a theorem, not as an engineering shortfall. Detection delay cannot be driven to zero at a fixed false-alarm rate. The trade-off Page's chart and Wald's test formalise is a genuine limit, not a limit of current instruments. A pharmaceutical fill line sampling every fifteen minutes, with a CUSUM tuned to catch a half-sigma mean shift within about eight samples, still takes those eight samples — during which out-of-spec vials are already mixing into the batch. Watching does not abolish delay. It bounds it.

Nor does watching explain anything. A CUSUM alarm, an STA/LTA seismic trigger, a loss-ratio break in an insurance book — all of these say the distribution moved. None says why. Cause has to come from elsewhere: provenance, domain knowledge, a second investigation. The defensible claim is narrower than the misreading, and worth stating precisely: only a watcher can bound the delay at all. Coverage buys the right to a guarantee. It does not fill in the guarantee's contents.

Three objections, taken straight

Retrospective change-point detection works fine on archived data. Climatologists locate regime shifts in ice-core records deposited long before anyone was watching. The mathematics needs an ordered record, not a live observer.

Correct, and it narrows the claim usefully. Retrospective methods need only order, not liveness — but the record itself is the residue of continuous deposition, and its resolution caps what can be recovered. Ice cores date breaks to decades, not days. And retrospective segmentation cannot bound delay, because delay is a property of acting in time, and the process being segmented is already over. Archives answer "when, in hindsight." They cannot answer "how soon could this have been known," because that question has no referent once monitoring has stopped.

Continuous monitoring of many streams makes the multiple-comparisons problem catastrophic. Watch ten thousand series at once and you get constant, uninformative alarms. More intake means more false change-points, not better control.

This is the real cost, and it is severe. False-alarm rates compound across streams and across time. But the remedy — false discovery rate control over sequential tests, always-valid inference using e-values, hierarchical models sharing strength across correlated streams — is also only available to something doing the watching. Calibration requires a history of alarms checked against outcomes. A frozen corpus cannot be calibrated against a future it never sees. Continuous intake does not remove the burden of discipline. It is the only place that discipline has data to run on.

Real processes drift; they do not break. Fitting a change-point to gradual drift imposes false discreteness, and continuous intake would just produce an endless run of spurious segments.

True, and it is why drift-rate estimators exist beside abrupt-change models. But the line between drift and break is itself a function of sampling rate — a step at hourly resolution is a ramp at millisecond resolution — and sampling rate is an intake parameter, not a fact about the world independent of how it was measured. A system with denser and longer intake can test whether a change is discrete rather than assume it, and can revise a segmentation later when finer evidence arrives. That revisability, tied to provenance, is the answer to imposed discreteness, not a denial of the objection.

What this does not establish

The argument fixes a precondition, not a capability. It says that a running delay bound, in the sense Page and Lorden gave the term, requires an unbounded stream to bound it over. It does not say that a system meeting that precondition detects changes well, cheaply, or without drowning in false alarms across thousands of correlated series — the second objection stands on its own regardless of architecture. It does not say retrospective analysis of finished records is worthless — the first objection stands too. It says only that "when did this change" is a question whose answer-space is time-stamped observation, and that only a position built to keep observing indefinitely can be held to the theorem that gives the question its rigour. The ladder ends there because there is no fourth position: nothing is more persistent than a stream that never stops.