The Evacuation Inform Index

Every displacement crisis contains a question that is rarely asked out loud: is leaving actually safer than staying? Roads are mined or blockaded. Checkpoints separate families. The elderly and the sick do not survive transit that a healthy adult would manage. A destination that looked open on Monday has closed by Thursday.

Existing humanitarian indices measure how bad a situation is. They rank crises by severity so that funding and attention can be allocated. What they do not do is model the alternative. A crisis scored as extremely severe tells a planner that people are suffering; it does not tell them whether moving those people would reduce or increase the harm.

That gap matters because the two risks move independently. A siege can be catastrophic to endure and simultaneously impossible to escape. A drought can be survivable in place while the routes out remain open and safe. Treating severity as a proxy for a decision to evacuate conflates conditions that need to be distinguished.

International humanitarian law makes the distinction consequential rather than merely analytical. Under Article 49 of the Fourth Geneva Convention, an Occupying Power may undertake total or partial evacuation of a given area where the security of the population or imperative military reasons so demand, must return those evacuated to their homes once hostilities have ceased, and must ensure proper accommodation, satisfactory conditions of hygiene, health, safety and nutrition, and that members of the same family are not separated. Evacuation is a regulated act with conditions attached, not a self evidently benign one.

The Evacuation Inform Index is a browser based research prototype. It presents a world map of active crises, a panel of live developments for each one, a methodology section, a catalogue of data sources, and a reference list.

It scores 104 active humanitarian crises on two dimensions: the risk of remaining in place, and the risk of attempting to leave. It expresses the relationship between them as a single ratio, while always displaying the two component scores that produced it.

It is not deployed in any operational setting, has no institutional mandate, and produces no output that any organisation is obliged to act upon. The distinction between what is implemented and what is planned is marked throughout the interface rather than left to the reader to infer.

The index expresses its result as a ratio: the risk score for evacuating divided by the risk score for staying. A value above 1.0 indicates that evacuation carries more risk than remaining. A value below 1.0 indicates the reverse. A value near 1.0 indicates that the two courses of action carry comparable risk, which is itself a meaningful finding rather than an absence of one.

Ratios become unstable as the denominator approaches zero, so a floor is applied to the staying score. And the ratio is never shown alone: both component scores accompany it wherever it appears.

The component scores are not invented for the prototype. They derive from the INFORM Severity Index, the monthly crisis severity model produced by ACAPS with the Joint Research Centre of the European Commission. INFORM assembles 31 core indicators into three weighted dimensions: the impact of the crisis at 20 per cent, the conditions of the affected people at 50 per cent, and the complexity of the situation at 30 per cent.

The index maps two of those dimensions onto its own question. The conditions of affected people become the risk of staying, measuring how severe it is to remain in place. The complexity of the crisis, which covers access constraints, safety, and the operating environment, becomes the risk of evacuating, measuring how difficult and dangerous it is to move. The substitution is defensible on its face, since complexity measures precisely the conditions that obstruct movement, and the prototype labels it as a substitution rather than a measurement.

Indicators combine by weighted geometric mean rather than arithmetic mean. The practical consequence is that a catastrophic score on one component cannot be averaged away by comfortable scores elsewhere. If every evacuation route is closed, no amount of favourable weather or food security compensates for it. This is the same non compensatory logic used by the Human Development Index, and it reflects a judgement about the world rather than a mathematical preference: some failures are absolute.

The prototype adopts a structural argument from the CERAI framework, which holds that danger and feasibility must never be collapsed into one score.

The reasoning is legal as much as analytical. The obligation to protect civilians arises from the danger they face. Operational difficulty does not extinguish that obligation. A single blended number would allow low feasibility to drag down the composite, making a desperate situation appear less urgent precisely because it is hard to resolve.

Keeping the two visible produces the opposite effect. High danger combined with low feasibility raises a protection gap flag, signalling that the problem has moved beyond operational reach and requires political engagement rather than a logistical answer.

Endangerment is drawn from the INFORM conditions score and expressed as a percentage, with a 75 per cent obligation threshold marked and a trajectory drawn from recent conflict fatality trends. Feasibility inverts the complexity score and reduces it by live route weather. A trajectory indicator compares fatalities in the most recent three months against the preceding three, answering whether a situation is deteriorating or stabilising, which a static severity score cannot convey.

A crisis score describes a population. It does not describe a person within it. The prototype therefore offers an interactive profile of twelve factors that a user can switch on to represent a particular group, bounded between a multiplier of 0.7 and 1.3.

The multiplier raises endangerment and lowers feasibility, on the reasoning that these are two causally distinct pathways rather than one effect measured twice. A personal characteristic does not add a fixed quantity of risk; it scales the risk already present in the environment. Being elderly in a stable region and being elderly under bombardment are not the same increment.

The twelve factors extend well beyond mobility. Young children, the elderly, the disabled and medically dependent, the wounded and acutely sick, and pregnant women and new mothers all bear on the physical capacity to travel. Unaccompanied and separated minors, people facing gendered violence risk, members of targeted ethnic, religious or political groups, undocumented people with identity gaps, and linguistic minorities with low literacy face a different kind of exposure: they are endangered by targeting, detention, or exclusion rather than by the difficulty of walking. Two factors work in the opposite direction, reducing assessed risk: prior evacuation experience and available financial resources.

The distinction matters. A demographic list alone would treat vulnerability as a question of who moves slowly. The protection based additions treat it as a question of who is hunted, who is turned back, and who cannot read the sign telling them where to go.

The laboratory hosting this work has a lineage in supply chain traceability and forced labour mitigation, and several practices carry over into how the index treats evidence.

A two witness standard separates a fact corroborated by two or more independent reports from a single unverified claim, turning source credibility into a graded ladder rather than a binary.

The score is deterministic rather than generated. Every number comes from a fixed, auditable formula. Where a language model is involved, it supplies source grounded facts and never overrides the arithmetic. A reader can reconstruct any score by hand.

An evidence gap is treated as a risk signal rather than a neutral absence. An unverified corridor segment is penalised, not assumed safe. This inverts the common default in which missing information quietly reads as an acceptable situation. Where the tool falls back to background knowledge rather than a live verified source, that fallback is labelled, capped below verified status, and flagged.

The optional satellite damage layer follows the same discipline. It stays empty until a user connects their own deployment, and the interface reports the real state of that connection rather than assuming it: green only once imagery actually arrives, failure when none does, and partial coverage distinguished from failure. Nothing about the connection is asserted before it is observed.

The repository states nine limitations. They are reproduced here because they are the most important part of the document for any reader considering what weight to give the tool.

Proxy construct. Endangerment and feasibility derive from INFORM sub scores supplemented by conflict and weather data. They should be read as a faithful architectural proxy, not a validated instrument.

Researcher assigned weights. Every weight is a best estimate pending expert validation. This places them at the same evidentiary level as INFORM's own initial weights, which ACAPS likewise describes as a current best estimate to be refined by expert analysis.

No ground truth calibration. Independent ground truth for evacuation decisions does not exist. There is no register of correct historical calls against which the index could be scored.

Population level vulnerability. The demographic profile is a scenario the user sets, not measured household data. It cannot capture individual circumstance.

Static snapshot. The endangerment and feasibility pair has no corridor dynamics, ceasefire windows, or modelling of misinformation. When hosted as a static site, news and conflict data are captured rather than live.

No modelling of political will. The index cannot represent actor behaviour, negotiation status, sudden shifts in belligerent intent, or the granting and withdrawal of consent. In many real evacuations these are the determining factors.

Correlation, not causation. The index prioritises attention. It is decision support and must not be the sole basis for an evacuation decision.

The contribution of this prototype is not a novel algorithm. It is a reframing. By insisting that the risk of leaving be scored separately from the risk of staying, and that danger be scored separately from feasibility, it makes visible a comparison that humanitarian assessment usually leaves implicit.

Its second contribution is its treatment of its own uncertainty. The weights are labelled as unvalidated. The proxies are labelled as proxies. The absent capabilities are listed rather than omitted. The satellite overlay reports honestly that it is not connected. For a tool that touches decisions about who moves and who remains, this reticence is not a weakness in the work; it is a substantial part of what the work is demonstrating.

The index does not know whether anyone should evacuate. It is built to make the question harder to answer carelessly.