Compactness in humanitarian response

What compactness actually says

A collection of open sets covers a space if their union contains every point of it. The space is compact if every such covering, however large, contains a finite sub-collection that still covers everything. Heine-Borel makes this concrete on the real line: a set is compact exactly when it is closed and bounded. The half-line fails this — the intervals (-n, n) cover it, but no finite handful of them will, since something always lies beyond. The open interval (0,1) fails it too, despite being bounded: cover it with (1/n, 1) for every n, and any finite selection still leaves points near zero exposed. Boundedness is not enough. The boundary itself has to be included, closed off, or the finite inspection that compactness promises never arrives.

That promise is the whole point of the concept. Compactness is what allows a finite check to settle an infinite question — a continuous function on a closed interval provably reaches a maximum, because finitely many local neighbourhoods can be patched into one global fact. Where the space is not compact, no amount of finite inspection settles anything permanently. The inspection covers what it covers, and the rest waits.

From covers to corpora: the lineage

A training corpus is a finite subcover. It is a finite family of recorded situations, offered as sufficient for the space of situations a system will actually meet. That offer is only honest where the space is compact. The Large Language Model treats its corpus as closed — everything worth knowing sits inside a collection frozen at a cutoff date. The Large World Model drops that assumption locally: it senses the present scene directly, covering it while it lasts, which is real compactness on a bounded interval of time. But the cover lapses the moment the scene changes, because nothing in the model extends it. The Large Universe Model gives up treating the cover as an object at all. It treats coverage as a process — streams left running, beliefs stamped with provenance and a decay clock, so the cover is extended as the space itself extends.

The set of future events has no finite subcover. It is unbounded in time, and its troublesome cases sit at limit points no sample has yet reached — which is what distribution shift feels like from inside a system that cannot tell it is happening. Enlarging the corpus does not fix this, because non-compactness is not a size defect; a bigger frozen collection is still a frozen collection. The only structural response is to keep the cover open: observe continuously, date every belief, allow every belief to be revised. Once intake is continuous and unbounded, there is no further category of evidence left to admit. What is left to argue about is scale, latency, trust and time — not what to observe next, but how fast and how honestly.

The response coordinator's problem

Humanitarian response is where this stops being an abstraction. A response coordinator working a displacement crisis is fed four streams that never stop moving: displacement flows tracked through registration and site counts, market prices for the basket of goods that determines purchasing power, health surveillance data on morbidity and mortality, and access constraints — which roads are open, which checkpoints exist, which areas armed groups have closed to convoys this week.

The characteristic failure of the field is precise and repeats every year: aid gets allocated against an assessment that the population has already outrun. A rapid needs assessment takes two to three weeks to design, field, clean and clear for use. In that window, in an active displacement crisis, the population it describes can have moved twice. The 2017 Rohingya influx into Cox's Bazar saw arrivals of over 20,000 people a day at points in September; an assessment closed on the 10th and published on the 24th was describing a camp population that no longer existed in that configuration. The document was not wrong when it was collected. It was a finite subcover of a space that kept extending past its boundary while the report was being formatted.

The Integrated Food Security Phase Classification exists precisely to manage this gap: it forces a re-analysis on a fixed cycle, typically every six months, with a stated validity window. That is compactness treated honestly — a bounded cover, openly labelled as temporary, rather than one silently mistaken for permanent. The failure mode the field actually suffers is not the IPC's; it is what happens between cycles, when a single snapshot gets used as though it were still current three months after its own analysis window closed.

Two objections from the field

Statistical forecasting works. Displacement models trained on historic flow data and conflict indicators have genuinely predicted movement corridors in advance in several contexts. If finite training data can forecast infinite future movement, the whole compactness objection is overstated.

This is true and it is exactly the domain where the qualifying condition matters. Forecasting models of this kind rely on the assumption that the drivers of movement this season resemble the drivers of movement in the training period — a stationarity assumption, in the same sense PAC learning bounds require a fixed sampling distribution. Where the underlying dynamics hold — seasonal drought-driven movement along established routes, for instance — the finite historical record covers the near future well, because the space of relevant situations is, locally, closed and bounded. Where a new actor enters, a border closes without notice, or a conflict front shifts axis entirely, the forecast is extrapolating outside its cover, and it degrades exactly where it is needed most: at the onset of the event nobody has modelled yet. The 2023 Sudan displacement, following the RSF-SAF clash in Khartoum in April, produced movement corridors into Chad and Egypt that existing regional flow models had not weighted, because nothing in the training window resembled a capital-city conflict of that speed. The models were not incompetent. Their cover simply ended before the event began.

Continuous monitoring is not a solution, it is an expense. Running DTM tracking, price bulletins and health surveillance continuously does not make the future space compact — at any given moment the coordinator still only has a finite record, and the same uncovered future sits ahead of them. The infrastructure cost is real; the epistemic gain is not.

Conceded, in full. Continuous intake never closes the future. It cannot. What it changes is the shape of the failure. A quarterly assessment that goes stale fails silently: nobody inside the process is told that the gap between the document and the ground has grown, and allocation decisions keep citing a number that has quietly stopped meaning anything. A continuously updated Displacement Tracking Matrix, refreshed weekly against new site counts, fails with a bounded, visible lag instead — a coordinator can say the last verified figure is nine days old, and act accordingly, discounting it or flagging it as provisional. That is a smaller claim than "the model covers the crisis." It is the claim that the coordinator now knows how uncovered they are, which a frozen assessment never tells them.

Where the cover holds and what changes

None of this argues that assessment is worthless, and it would misread the argument to think so. Non-compactness bites at the boundary — the sudden influx, the closed corridor, the price shock from a currency collapse — not at the interior. For the stable core of a protracted crisis, a camp population with established registration, predictable ration cycles, monitored but static access, a periodic assessment covers the space extremely well, because the space genuinely is bounded and closed for as long as nothing changes. Most humanitarian caseload, most of the time, is exactly this interior.

The lineage claim, tested against this domain, holds its shape rather than dissolving into it. A frozen assessment is a Large Language Model's corpus in miniature: valid over the interior, wrong at the boundary, and blind to its own expiry. A live operations centre reading a single unfolding emergency off current feeds is closer to a Large World Model — genuinely compact for the duration of that crisis, and silent again once the feed stops or the crisis officially closes while people are still moving. What a response coordinator actually needs, and rarely has built for them, is the third position: every stream running continuously across crises rather than within one, each figure carrying its collection date and its decay rate, revision treated as the normal state of a number rather than a correction to be embarrassed about. That is not a system that has solved displacement. It is a system that has stopped pretending its cover is finished.