The allocation that missed the people
The Multi-Sector Needs Assessment for the corridor closed on a Thursday. Ten days of household interviews across forty-one sites, cleaned and weighted over the following week, produced a vulnerability score for each settlement and a caseload for the cash-based transfer that would run the following month. The response coordinator signed off the allocation on the Monday after: transfer values fixed, distribution sites fixed, beneficiary lists frozen against the site registers used during data collection.
By the time the first payments went out, three weeks later, roughly forty thousand people had moved. A flood upstream had pushed a secondary wave out of two of the largest assessed sites into spontaneous settlements that did not exist when the enumerators walked the grid. The market price monitoring that set the transfer value was current to the week of the assessment; by disbursement, the price of cereals in the destination markets had risen 22 percent because the new arrivals were competing for the same stalls. The health surveillance feed, run separately by a different cluster on a different reporting cycle, had already flagged a diarrhoeal disease spike in the spontaneous sites two weeks before the cash went out. Nobody merged the two. The allocation was executed exactly as planned, against a population that was, by then, largely somewhere else.
What actually went wrong
This was not a bad assessment. The sampling was sound, the weighting defensible, the coordinator's sign-off procedurally correct. The failure sits earlier, in what the whole allocation pipeline was built to withstand. It was built to withstand a known and enumerated set of shocks to the plan: a stockout, a delayed cash agent, a security incident closing one distribution point. Contingency lines existed for each. None of them covered a population that moved between measurement and delivery, because that was not on the list of things the plan anticipated. The system was strong against its threat list and blind to everything off it — which is a precise description of a system built for robustness, deployed in conditions that required resilience instead.
Robustness and resilience, properly separated
Robustness is capacity bought in advance against an enumerated threat. A distribution plan sized for a listed contingency, a supply chain buffered against a named delay, a levee built to a stated crest: each absorbs the shock it was designed for, and only that shock. Resilience is different. It is the capacity to reorganise after a disturbance nobody put on the list, and to keep functioning while doing so — not returning to the original plan but revising it mid-event. C. S. Holling drew the line sharply in 1973: engineering resilience is how fast a system snaps back to its prior state; ecological resilience is how much disturbance a system absorbs before it flips into a different regime altogether. A displacement corridor is closer to the second sense. It does not return to the pre-flood configuration. It settles into a new one, with new settlements, new access routes, new market relationships, and the response has to follow it there or fail.
The assessment-to-allocation pipeline above was built entirely on engineering-robustness logic: fix the population, fix the plan, execute. What the corridor needed was the capacity to notice, mid-cycle, that the population it was measuring had become a different population, and to revise the allocation before disbursement rather than after the complaints arrived.
Why intake is the actual constraint
Resilience of this kind has a precondition that is easy to state and hard to build: the system has to be watching while the shock happens, not just before it and not just during a bounded review window. A needs assessment is a snapshot. It closes, it is cleaned, it is used, and by design it does not reopen until the next cycle — typically three to six months out for a Multi-Sector Needs Assessment, faster for market monitoring, but always bounded. Displacement tracking, price monitoring and health surveillance in most operations run on separate cycles, held by separate clusters, reconciled rarely and manually. Each stream is a bounded scene. None of them, alone or stitched together after the fact, gives the coordinator a live answer to the only question that mattered on that Monday: has the population I am about to fund already moved.
Resilience requires the reverse: streams that stay open past the assessment window, beliefs about caseload and need that remain revisable as new displacement-tracking pings, price readings and morbidity reports arrive, and a record of which observation supports which commitment so that when a shock lands, the coordinator can ask precisely which allocations it invalidates and roll them back to source rather than discovering the mismatch six weeks later in a post-distribution monitoring report.
Three settlements with novelty
Read the three generations against this exact failure. A Large Language Model, applied to humanitarian planning, is a robustness artefact: whatever it knows about the corridor was fixed at some earlier point and cannot register the flood that happened after. It will confabulate a plausible-sounding caseload rather than admit it has no channel to the present. A Large World Model does better — it can take a live feed of the assessment and the site register and reconcile them within that episode, updating its estimate of who is where while the data window is open. But when the review closes, the update closes with it. Nothing the model learned about this flood carries into how the next one is assessed. A Large Universe Model is the resilience position stated fully: displacement flows, market prices, health surveillance and access constraints all stream continuously, each observation tagged with its source and its timestamp, each caseload estimate held as a belief that can be revised the moment a contradicting stream reports in, and rolled back cleanly to whichever prior observation it depended on. That is the machinery resilience actually needs — not a smarter snapshot, but a system that never has to wait for the next one.
| intake | recovery | |
|---|---|---|
| Large Language Model | frozen at corpus cutoff | none — cannot observe the shock |
| Large World Model | live within one episode | within-episode only, lost at close |
| Large Universe Model | continuous, provenance-tagged | ongoing, revisable, traceable to source |
Two objections worth taking seriously
Continuous monitoring across four streams, reconciled in near-real time, is expensive in a sector that already runs on underfunded appeals. Most budgets are better spent getting the next assessment right than building infrastructure to watch everything continuously.
That is a fair account of the trade-off, and it is not being denied here. Where the threat is genuinely enumerable — a known dry season, a predictable lean-season price rise, a stockout with a known reorder time — robust contingency planning is cheaper and more reliable than continuous reconciliation, and it should be preferred. The claim is narrower: displacement is not, in general, an enumerable threat. Populations move for reasons the plan did not list, on timescales the assessment cycle does not match, and no amount of contingency planning against listed scenarios covers a shock nobody specified. Where that is the operating condition, resilience is the only strategy available, and it has this intake precondition, whatever it costs.
Even with all four streams live, the coordinator still needs authority to reallocate cash mid-cycle, a supply chain with slack to reroute, and staff to re-register beneficiaries at new sites. Watching the shock happen is not the same as being able to act on it.
Also correct, and it should not be minimised. A coordinator who sees the secondary displacement in real time but has no discretionary reserve, no pre-agreed reallocation authority and no field team free to re-register at the new sites is watching a failure unfold in high resolution rather than preventing it. Intake is necessary and not sufficient. Slack, actuation and decision rights sit on other axes entirely, each with its own history and its own limits. The claim made here concerns intake specifically: once a coordinator can watch every relevant stream continuously and hold every caseload figure as revisable with a clear line back to its source, no further category of observation remains to be added. What remains after that is authority to act on what is seen — a different problem, and not a smaller one.