Real data. RSS = INFORM Conditions of people affected; RSE = INFORM Complexity (access, safety, operating environment), each rescaled 1β5. Bubble size scales with the metric. Click a marker for the full breakdown.
HASTE has no public API β run it yourself (
docker compose -f docker/docker-compose.yml up), open a disaster project, and copy its damage-layer tile URL. See HASTE_SETUP.md.Live Updates & Conflict
Per-crisis real-time developments (Tavily news) and the structured conflict timeline (ACLED) β the same feed as the map drawer, browsable here without the map. Every item links to its primary source.
Select a crisis above to load live dataβ¦
Methodology
The Evacuation Inform Index (EII) is structured as a ratio β the Risk Score for Evacuating (RSE) divided by the Risk Score for Staying (RSS) β synthesising the INFORM Severity 3-dimension structure with the IOM RICD macro/micro split.
EII = RSE / RSSEII > 1.0 β evacuation is riskier than staying | EII < 1.0 β staying is riskier than evacuating
Live data grounding. The map is populated with the real INFORM Severity Index, April 2026 (ACAPS / EU JRC, 104 active crises, via HDX). Each crisis's Conditions of people affected sub-score drives RSS (how severe it is to remain) and its Complexity sub-score β access constraints, society & safety, operating environment β drives RSE (how hard/dangerous it is to move). Both are rescaled from INFORM's 1β10 to the 1β5 EII scale. Layers 2β3 (route availability, personal modifiers) below are the design roadmap; the current build uses INFORM as the live backbone.
Design note. Pure ratios destabilise as the denominator nears zero. A floor of RSS = 0.5 (on a 5-point scale) is applied, and the EII is always shown alongside both component scores β never as the sole output.
CERAI lens β endangerment vs feasibility
The CERAI v3 framework makes one architectural argument: never collapse danger and feasibility into a single score. Under IHL the obligation to protect civilians flows from danger (GC IV Art. 49), not from operational feasibility β low feasibility does not extinguish the obligation, it intensifies the urgency of political engagement (Mariupol 2022 is the proof-of-concept). This index renders that split live, per crisis, on the π΄ Live & Conflict tab:
| CERAI dimension | What it answers | How we compute it here |
|---|---|---|
| Endangerment (Threat Environment) | How dangerous is it to stay? | INFORM Conditions (RSS) β 0β100%, with the 75% IHL Obligation Threshold marked; trajectory from live ACLED fatalities |
| Feasibility (can people move?) | Can civilians realistically move? | INFORM Complexity (RSE), inverted β 0β100%, reduced by live Open-Meteo route weather |
| Protection gap flag | Where is escalation needed? | High endangerment (β₯75%) + low feasibility (β€40%) β flags need for political, not operational, escalation |
| CERAI dimension | What it answers | How we compute it here |
|---|---|---|
| Vulnerability profile (Dimension 3) | Who is exposed? | Interactive demographic toggles set a 0.7Γβ1.3Γ multiplier that amplifies endangerment and reduces feasibility; the 0.7Γβ1.3Γ span is shown as the endangerment range across the population |
| Damage assessment (satellite) | What is physically destroyed? | Microsoft HASTE + Planet AI damage maps overlaid on the Map tab β building loss raises endangerment (infrastructure availability), blocked routes lower feasibility |
Honest scope. This is a faithful proxy of CERAI's architecture built from data the index already carries (INFORM + ACLED + Open-Meteo, plus an optional Microsoft HASTE satellite-damage overlay) plus the interactive D3 profile β not CERAI's full 22-variable IHL engine, source-credibility weighting, Monte-Carlo robustness, or 47-case comparator. CERAI scores map to the INFORM 0β10 scale by Γ·10, so the two frameworks are directly comparable.
Three-layer architecture
βΈ Click a layer to see its variables & proposed weights.
| Layer | Function | Precedent |
|---|---|---|
| Layer 1 β Objective Risk Score (ORS) variables β 50% | Universal factors, identical for everyone in the geography (hostilities, conflict risk, natural-hazard life risk) | INFORM Severity, ACLED, GCRI |
| Layer 2 β Infrastructure & Access (IAS) variables β 35% | Availability of evacuation routes, resources & connectivity | ACAPS Humanitarian Access, IDMC |
| Layer 3 β Personal Vulnerability Modifier (PVM) variables β 15% | Demographic / household factors applied as a multiplicative modifier (0.7Γβ1.3Γ) | CDC SVI, IOM RICD micro-level |
Aggregation β weighted geometric mean
Indicators aggregate by weighted geometric mean rather than arithmetic mean, so an extreme imbalance (e.g. all infrastructure unavailable) cannot be compensated by a low score elsewhere β the same logic used by the Human Development Index and INFORM Risk.
Score = Vβwβ Γ Vβwβ Γ β¦ Γ VβwβVα΅’ = normalised 1β5 score Β· wα΅’ = weight (Ξ£w = 1.0)
Recommended build sequence
| Phase | Method | Purpose |
|---|---|---|
| 1 Β· Variable design | Delphi + Budget Allocation (8β12 experts, 2 rounds) | Set initial layer & sub-variable weights |
| 2 Β· Weight validation | Fuzzy AHP (triangular numbers, Buckley's geometric mean) | Validate contested weights under uncertainty |
| 3 Β· Calibration | Historical case testing (Sudan '23, Ukraine '22, Kabul '21, Lebanon '06, Haiti '10) + PCA | Check the index matches real decisions; prune redundant variables |
Constraint layers (not scored)
Financial feasibility filter β a second-stage check on whether the recommended action is affordable (transport, accommodation, asset-liquidation loss, income disruption). Legal / rights (UDHR Art. 13) β flags exit-visa requirements, travel bans or closure orders that restrict self-evacuation.
Variables & Proposed Weights
Preliminary weights synthesised from INFORM, ACLED and FSI logic β to be validated by AHP expert surveys.
Layer 1 β Objective Risk Score Β· 50%
| Variable | Data source | Weight | Rationale |
|---|---|---|---|
| Active hostilities | ACLED: fatalities, attack types, proximity | 20% | Immediate life threat; fastest-changing β heaviest weight |
| Likelihood of future hostility | GCRI risk score; FSI security; ICEWS | 15% | Forward-looking; less certain than observed events |
| Life risk (non-conflict) | IDMC, FEWS NET β flood, quake, fire | 15% | Natural-hazard exposure alongside conflict |
Layer 2 β Infrastructure & Access Β· 35%
| Variable | Data source | Weight | Rationale |
|---|---|---|---|
| Evacuation route availability | Flight seats, road/border status, satellite imagery | 12% | No route = evacuation impossible |
| Infrastructure availability (stay) | Internet, energy, food, water (IPC, FEWS NET) | 10% | Determines survivability if staying |
| Security / threat alerts | OSAC, embassy alerts, local-language news | 8% | Near-real-time signal, both directions |
| Weather | NOAA, Copernicus | 5% | Modifier on route viability & shelter |
Layer 3 β Personal Vulnerability Modifier Β· 15% (multiplicative 0.7Γβ1.3Γ)
| Variable | Operationalisation | Direction of effect |
|---|---|---|
| Young children (<12) | CDC SVI "age β€17"; self-report | β RSE (harder to move) & β RSS (more vulnerable) |
| Elderly (65+) | CDC SVI "age 65+"; self-report | β RSE (mobility) & β RSS (medical risk) |
| Gender / gendered risk | UNHCR GBV risk indicators | β RSS in conflict zones with GBV risk |
| Prior evacuation experience | Self-assessed preparedness | β RSE (more capable evacuee) |
| Financial resources | Self-reported | β RSE when high; β RSS when low |
Weighting methods considered
| Method | Subjectivity | Data | Defensibility | Use |
|---|---|---|---|---|
| Equal weights | None | None | Low | Baseline & sensitivity |
| Budget allocation (BAP) | High | None | Medium | Rapid prototyping |
| AHP | Medium | Expert survey | High | Published index |
| Fuzzy AHP | LowβMed | Expert survey | Very high | Ambiguous variables |
| PCA / factor analysis | None | Historical | Medium | Validation & pruning |
Variables integrated from ETC evacuation projects
Variables and structures mapped across sibling Ethical Tech CoLab evacuation repos, folded in here (or on the roadmap) to enrich the model beyond a demographic-only vulnerability list.
| Contribution | From repo | Status here |
|---|---|---|
| Protection-based vulnerable groups β wounded/acutely sick, pregnant & new mothers, unaccompanied/separated minors, undocumented / ID-gap persons, targeted ethnicΒ·religiousΒ·political minorities, linguistic minorities / low literacy (each with an IHL basis, e.g. AP I Arts 16, 78; GC IV Art 23; customary IHL Rules 98β99) | Evac-Sim-Melanie | Added to the Dimension-3 profile (Live tab). These are protection/legal vulnerabilities, not just mobility ones β several raise endangerment via targeting/detention rather than slowing movement. |
| Destination-readiness gatekeepers β Security, Authority consent, host Willingness, Capacity, Shelter, Food/water, Medical capacity; a confirmed host refusal hard-caps readiness | India-EvacSimulation | Roadmap β feasibility currently uses a single INFORM-Complexity score; gatekeeper caps would replace it. |
| Seven-dimension model (D1βD7) + CERAI cross-derivation (Endangerment = d1Β·0.45 + d2Β·0.20 + d6Β·0.35; Feasibility inverts d3,d4,d5,d7), NATO STANAG level binning | ercf Β· Exodus | Roadmap β a concrete recipe to decompose Endangerment/Feasibility into scored sub-dimensions. |
| Corridor / checkpoint dynamics β open/closed exit gates, ceasefire windows, congestion queues, siege "trapped" state, information-environment degradation & misinformation | Evac-Sim-Melanie | Roadmap β the static Endangerment/Feasibility pair has no time-varying corridor or information dimension yet. |
| Historical calibration harness β differential-evolution fit to 16 in-scope cases (RΒ²=0.855, LOOCV 0.807), with documented out-of-scope failure modes | ercf | Informs the limitations below; the model-boundary honesty is the borrowed practice. |
| Non-compensatory geometric-mean aggregation + RSS floor of 0.5 | evacmodel | Already used β see Aggregation above. |
Evidence provenance & traceability
The lab's lineage in supply-chain traceability and forced-labor mitigation (director Yorke Rhodes, Microsoft) shapes how this index treats evidence. Patterns adapted from that work and two public forced-labor models:
| Principle | Adapted from | Effect on the methodology |
|---|---|---|
| Two-witness evidentiary standard β separate a verified fact from β₯2 independent corroborating reports from a single unverified one | GFEMS FLARE | Sharpens the source-credibility tiers (UN-verified 1.0Γ β unverified 0.7Γ) into an evidentiary ladder. |
| Deterministic score, not an AI score β the confidence/risk number is a fixed, auditable formula; the language model only supplies source-grounded facts | provenance-search Β· arts-provenance-agent | The INFORM 0β10 mapping and the D3 multiplier stay reconstructable and never overridden by a model. |
| Chain-of-custody: an evidence gap is a risk signal | arts-provenance-agent | An unverified corridor segment is penalised, not assumed safe β mirroring provenance-gap logic. |
| Labeled graceful degradation β a fallback to background/model knowledge is tagged, capped below "verified," and auto-flagged | provenance-search | Absence of a live verified source never reads as verified. |
| Behaviorally-grounded indicator scoring β risk from a weighted set of observable indicators, not one metric | Global Fishing Watch forcedlabor | Precedent for the multi-indicator endangerment/feasibility structure. |
| Decision-support, not a verdict β a human retains the high-stakes call | FLARE ("a decision support tool, not an executioner") | Governance stance stated explicitly (below). |
Honesty guardrail. This index does not use ILO's 11 forced-labor indicators, machine-learning / positive-unlabeled classifiers, or cryptographically signed credentials β those are the methods of the cited forced-labor models and the provenance passport, referenced as design lineage, not claimed as implemented here.
Research limitations
- Proxy construct. Endangerment and feasibility are derived from INFORM Conditions and Complexity sub-scores (plus live ACLED/weather), not CERAI's full 22-variable IHL engine. Treat them as a faithful architectural proxy, not a validated instrument.
- Researcher-assigned weights. All weights are best estimates pending expert validation (Delphi β Fuzzy AHP) β at the same evidentiary level as INFORM's initial weights, but not yet consensus-tested.
- No ground-truth calibration. Independent ground truth for evacuation decisions does not exist. Sibling calibration (ercf: 16 cases, RΒ²=0.855) is face validity, not statistical generalisation, and is explicitly out of scope for genocide, large-enclave precision operations, and sieges beyond ~90 days.
- Population-level, illustrative vulnerability. The Dimension-3 profile is a user-set demographic scenario applied as a 0.7Γβ1.3Γ multiplier β not measured household data β and cannot capture individual circumstances.
- Static snapshot. The Endangerment/Feasibility pair has no corridor/checkpoint dynamics, ceasefire windows, information-environment degradation, or misinformation (all modelled in Evac-Sim-Melanie, not here). Hosted news/ACLED are captured snapshots; weather is live.
- No political-will modelling. The index cannot model actor behaviour, negotiation status, sudden shifts in belligerent intent, or consent dynamics.
- Evidence quality varies. Source-credibility tiering is a design principle; live inputs (ACLED, Tavily) differ in verification and are not yet weighted by it in code.
- Satellite & damage layers are optional. The HASTE damage overlay requires a self-hosted deployment; without it, physical-damage evidence is absent.
- Correlation, not causation. The index prioritises attention; it is decision-support and must not be the sole basis for an evacuation decision.
Data Sources
Feeds that populate the index. All free unless noted.
Reference Indices & Key Papers
Methodological precedents
INFORM Severity Index
3 weighted dimensions (Impact 20% Β· Conditions 50% Β· Complexity 30%), 1β5 scale. Closest analogue.
acaps.org βACLED Conflict Index
Deadliness 35 Β· Danger to civilians 25 Β· Diffusion 20 Β· Fragmentation 20; non-linear root aggregation.
acleddata.com βIOM RICD
Two-tier macro (spatial risk) + micro (community) model β the structural precedent for base score Γ personal modifier.
iom.int βIDMC Risk Model 2.0
Probabilistic displacement from natural hazards β feeds the "risk of staying" dimension.
internal-displacement.org βFragile States Index
12 indicators, CAST text-analysis triangulation β template for structural / minor variables.
fragilestatesindex.org βCDC Social Vulnerability Index
16 variables, 4 equally-weighted themes, percentile ranking β the personal-vulnerability layer.
atsdr.cdc.gov βGlobal Conflict Risk Index
Conflict-onset risk β separates active conflict (ACLED) from forward-looking risk.
jrc.ec.europa.eu βOECD/JRC Composite Handbook
Definitive reference for normalisation, aggregation, weighting & sensitivity analysis.
publications.jrc.ec.europa.eu βMicrosoft HASTE
AI framework turning satellite imagery into building/route damage maps (Azure Maps, Batch GPU, ML). No public API β deploy it, then overlay its damage tiles on the Map tab. Forked under this org for deployment.
aka.ms/HASTE β Β· our fork βKey papers
- Beccari, B. (2016). A Comparative Analysis of Disaster Risk, Vulnerability and Resilience Composite Indicators. PLoS Currents Disasters, PMC4807925 β reviews 106 index methodologies.
- Can severity of a humanitarian crisis be quantified? Assessment of the INFORM severity index. Globalization & Health (2023). DOI:10.1186/s12992-023-00907-y β identifies governance & access as strongest predictors.
- Al Fozaie (2022). A Guide to Integrating Expert Opinion and Fuzzy AHP When Generating Weights for Composite Indices. Advances in Fuzzy Systems. DOI:10.1155/2022/3396862.
- OECD/JRC (2008). Handbook on Constructing Composite Indicators. OECD Publishing.
- Saaty, T.L. (1990). How to Make a Decision: The Analytic Hierarchy Process. EJOR 48(1), 9β26.
- ACAPS Ukraine Severity Model Methodology Note (March 2024) β worked subnational example.