The aid budget fell by half. The need did not.

Half the money. The same number of people.

In a single year, US official development assistance fell from about $63 billion to under $29 billion. The UN's 2025 appeal took in $12 billion, the lowest in a decade, and twenty-five million fewer people were reached than the year before.

The sector's own response is triage, and it is stated in the open: 239 million people in need, 135 million targeted, 87 million "hyper-prioritized." That framing is not ours. But it does define the problem precisely. When the budget is fixed and the need is not, the only remaining lever is how well each dollar is placed.

A map, a spreadsheet, and a room.

Deciding where a camp's water points, latrines and tubewells go is, today, a workshop. Local knowledge, a printed map, whatever land is available, and whoever argues hardest. The output is a list of sites with no coverage number attached, which means the argument has no arithmetic in it.

This is not negligence — it is the absence of a tool. The people in that room are the best-informed people in the country. They are working without the one thing that would settle the question, and without a number they can defend line by line to a donor — which, this year, is the conversation.

The standards are already the constraints.

Humanitarian response has a published rulebook. The Sphere standards are not aspirations; they are numbers, and every one of them is a constraint a solver can read:

That makes siting a coverage problem with a budget, not a workshop. Pick the set of positions that brings the most households inside those numbers for the money available. It is the same shape of problem as placing charging infrastructure across a province, which is a solver we have already built and run.

Cox's Bazar, one monsoon.

A million people across thirty-three camps on erodible hills — the most densely populated refugee settlement on earth. 152,000 people arrived after late 2023, against a response plan written for 50,000. In one week of July 2026: 286 weather incidents, 95 landslides, 26,119 people affected.

We gave the solver 180 candidate positions, four asset classes and a $2.0 million ceiling, and asked it for a site plan. It returned nine positions — and left $150,000 unspent, on purpose, because the ninth dollar of the last site bought less coverage than leaving it in the envelope. Every line carries what it serves, what it costs, and the Sphere clause it satisfies.

People move toward what you built.

Standard planning treats where people live as a fixed input, solves against it, and is wrong by the second month. Put in a water point and the settlement shifts toward it — the plan changes the thing the plan was solved against.So we don't sample three thousand random draws. We simulate three behaviours, each a household making a rational choice:

Run those three forward nine weeks against the plan just solved and the gap stops being a worry and becomes a coordinate: 1,340 households beyond 500 m by week nine. You now know where the plan breaks, in April, in a model — rather than in July, in a monsoon.

Same settlement. Same season. Same ceiling.

Nothing differs between these two columns except how the positions were chosen. Not the terrain, not the population, not a single extra dollar.

4,822 more people reached in a single week of response,for less money than the plan it replaces. Not because anything was added, but because a single-modality response runs out of envelope at person 21,297 — and cash where markets work, in-kind and road where they don't, air only where nothing else reaches, is an assignment problem with a right answer.

Before the season, during it, and after a loss.

One budget, three moments where it is committed — and they are three different problems, in this order:

Optimized by math. Pressure-tested by simulation.

Built to be checked — and to be wrong.

Every context figure below is public. Every solver figure is model output on synthetic data, and we would rather be corrected on it than quoted on it.