The objection that should win
Picture the strongest case against this piece before making it. A response coordinator running a rapid onset in the Sahel does not lack data because nobody thought to collect it. She lacks it because the people she is counting are moving faster than anyone can count them. A camp population assessed at 40,000 on a Tuesday is 24,000 by the following Monday, because a security incident forty kilometres north pushed a second wave through in the other direction. No sensor fixes that. No provenance tag makes a headcount taken on Tuesday describe Monday's camp. The problem is not measurement quality. It is that measurement is always of a moment, and the moment has passed by the time anyone acts on it. If that is right, the entire argument that continuous, provenance-tagged observation collapses transaction costs is beside the point in humanitarian response, because the thing being observed refuses to sit still long enough to be priced.
This deserves to be taken seriously before it is answered, because it is the daily experience of everyone who has ever signed off a distribution plan against a three-week-old assessment and watched half the caseload have already left the area. The failure has a name in the field even if nobody writes it in a report: aid allocated on an assessment overtaken by the movement it measured.
Coase's question, asked of a cluster meeting
Ronald Coase asked in 1937 why so much economic activity happens inside firms, under command, rather than through contracts negotiated on the open market. His answer was that markets are not free to use. Finding a counterparty, discovering a price, writing terms and policing them all cost something, and where that cost exceeds the cost of just directing the activity yourself, the activity gets absorbed into an organisation. Oliver Williamson later sharpened this with asset specificity and opportunism: firms internalise activity not only because monitoring is expensive but because a supplier who has sunk capital into a relationship can hold you up regardless of how well you watch them.
The humanitarian system is Coase's boundary drawn in aid budgets. A response coordinator constantly decides what to do through direct control — her own agency's trucks, her own agency's registration teams — and what to leave to negotiated arrangement — a local NGO subcontracted to run a health post, a transporter contracted per trip, a market vendor paid through unconditional cash rather than in-kind distribution. That boundary sits wherever the cost of contracting for the activity balances the cost of running it directly. And a large share of that cost, in this sector specifically, is the cost of not knowing: whether the road is passable today, whether the market still has stock at the price last surveyed, whether the health post that was functioning last month still has staff.
Three assessments, three prices for looking
The intake axis is three answers to what a system can watch and when. A Large Language Model, in the loosest analogy that fits this domain, is the situation report frozen at the point it was written: last month's multi-sector needs assessment, cited in this month's proposal, describing a caseload that has since dispersed. It prices only what was already written down. A Large World Model is the rapid assessment team physically present in the camp with a tablet and a GPS, verifying the scene while they are standing in it — genuinely useful, and useless the moment they drive away, because most of the risk in a displacement crisis lives in what happens after the team leaves.
A Large Universe Model, as an argued category rather than a fielded tool, is the condition of holding displacement flows, market prices, health surveillance and access constraints as streams that never stop, each belief tagged with where it came from and how stale it is, each revisable the moment a contradicting signal arrives. It does not claim to freeze the population count. It claims to know, continuously, that the count from Tuesday is four days old, corroborated by one source, contradicted by a border-crossing report from Thursday, and should be treated with according suspicion until refreshed.
| generation | what it can see | where it goes dark |
|---|---|---|
| Large Language Model | last published assessment, situation report, price bulletin | any movement since the cutoff |
| Large World Model | the camp, market or clinic while the team is physically there | the moment the team leaves the scene |
| Large Universe Model | displacement, price, health and access streams, held open with provenance | nothing, structurally — only accuracy and trust remain as limits |
This reframes the opening objection rather than dissolving it. The coordinator's complaint was never really that Tuesday's headcount is wrong. It was that Tuesday's headcount was treated, three weeks later, as if it were still Tuesday. A continuously running, provenance-tagged system does not stop people moving. It stops a stale number from masquerading as a current one, and it makes staleness itself a visible, priced quantity rather than a silent assumption buried in a spreadsheet.
What survives: the gatekeeper problem
Williamson's refinement bites hard here, and it should be conceded in full. A local implementing partner who has sunk two years into relationships with camp committees, or a transporter who is the only one willing to run the road through a checkpoint controlled by an armed group, holds power over the coordinator that no amount of visibility removes. You can watch the checkpoint continuously and still have exactly one transporter willing to cross it, and he knows it. Perfect information about a monopoly does not un-monopolise it. Access constraints in this sector are frequently not informational at all; they are political, and a live feed of who controls which road changes nothing about who controls which road.
What continuous observation does change is the shape of the argument that follows from that power. Without it, a transporter who claims the fee has doubled because the security situation worsened cannot be checked against anything, and the coordinator must either pay or lose the route. With a live corroborated feed on incident reports along that corridor, the same claim becomes checkable: either the security situation genuinely worsened, in which case the increase is legitimate and can be priced into a contingent contract in advance, or it did not, in which case the claim is naked rent-seeking and can be named as such. The hold-up survives. What it loses is cover.
What actually gets worse first
The second objection is the one the sector should worry about more, because the evidence for it is already sitting in incident logs. Every new stream is a new thing to fight over. Biometric registration figures get contested between a government ministry with an incentive to undercount and an agency with an incentive to fund a larger caseload. Satellite-derived population estimates get challenged by camp committees who suspect a lower count than their own headcount means a smaller ration. Market price monitoring run by one agency gets disputed by another running its own survey with a different sampling frame, and donor reporting absorbs the argument rather than the response.
Cheap observation, introduced into an environment with no shared standard for provenance, does not settle disputes. It multiplies the number of numbers available to dispute. This is the accurate, uncomfortable part of the objection, and it should not be argued away.
What resolves it is not more data but attached provenance: not a figure but a figure with a source, a collection date, a method, and a note on what has contradicted it since. A displacement estimate that says four sources, updated six hours ago, one flagged inconsistent is a different object from a displacement estimate presented as fact in a donor report. The first is arguable in a useful way — it tells you exactly where to look if you doubt it. The second is arguable in a useless way — the only recourse is to run a competing survey and start the fight again. The claim is not that continuous streams end contestation. It is that only continuous, provenance-carrying streams can end it, because provenance is the only thing that turns a disagreement into a checkable one.
The claim that survives
Not that displacement will slow to a pace assessments can catch, and not that gatekeepers lose their leverage over the one road out of a besieged town. Those are separate problems, and the second one is often the more important. What survives is narrower: the categories of thing a response coordinator can know are watched, timely and attributable, and a system that keeps every relevant stream running with provenance attached has exhausted those three categories. It cannot make the transporter cheaper. It can make sure the next argument about his invoice is a argument about facts rather than a argument about whose stale number wins.