A two-axis evacuation decision matrix. Axis 1 — Severity computes daily danger from hostilities, deprivation, and infrastructure destruction. Axis 2 — Feasibility assesses satellite-verified building damage, UNOSAT official assessments, and nighttime light loss across evacuation windows.
Retrospective research prototype. Kinetic variables derive from the author's ACLED extraction for Mariupol raion, held constant within documented phases; the temperature series is an interpolated climatology pending full ERA5 extraction; the corridor record and parties' conduct follow OCHA and OHCHR reporting; front line and controlled areas are approximate. Not operational guidance.
Six components, each min–max normalised to x ∈ [0, 1] (OECD/JRC 2008, Step 5): I hostility intensity — ACLED violent events/day, ceiling 10; P kinetic proximity — activity ≤5 km + 0.5 × ≤15 km, ceiling 6/day; R protection risk — mean of civilian-targeting share, lethality per event (ceiling 15), corridor violation history, and consent exposure on an ordinal scale from negotiated third-party escort (0.25) to adversary screening (0.90), operationalising GC IV arts. 17 & 23 and AP I art. 70; C cold burden — shortfall below the 18 °C Sphere habitable minimum, floor −10 °C, with outdoor ERA5 temperature proxying indoor conditions after utilities were destroyed (~2 March); D deprivation clock — days under encirclement, ceiling 60, less a relief credit per convoy reaching the city (Mariupol proper: zero); H infrastructure damage — cumulative share of the building stock damaged or destroyed, derived from satellite damage assessment (UNOSAT rapid assessments as the operational standard; VIIRS Black Marble night-light loss and Planet high-cadence imagery as complementary proxies) [12, 14, 15]. H is monotone by construction; the values in this prototype are illustrative anchors pending calibration against published building-level assessments.
The continuous composite is reported on a five-phase scale for operational legibility, following the presentation convention of the INFORM Severity Index and IPC classifications [4]: Phase 1 Minimal (S < 0.20, routine monitoring); Phase 2 Concern (0.20–0.40, contingency planning and registration of vulnerable groups); Phase 3 Serious (0.40–0.55, evacuation planning and access negotiation); Phase 4 High (0.55–0.70, evacuation warranted); Phase 5 Critical (≥ 0.70, immediate protective action). Thresholds are model conventions to be validated against expert elicitation.
S = ( (1/6) Σ xᵢ⁶ )^(1/6). A power mean always lies between the arithmetic mean of the components and their maximum; the exponent sets where (Decancq & Lugo 2013; OECD/JRC 2008). p = 6 places the composite close to the worst condition — the humanitarian weakest-link principle: a population safe from shelling but freezing without water is not "medium severity." The assessment panel therefore reports the simple average, the rating, and the worst component side by side, which communicates the rule without requiring the reader to engage with exponents. A linear average would score the post-March kinetic lull as recovery; Fig. 1 shows why it was not. Limited-compensability aggregation follows the INFORM Severity lineage (ACAPS/JRC) with the parameter made explicit.
Kinetic variables are ACLED-derived per-day rates held constant within documented phases (pre-9 Mar; 10–15 Mar; 16 Mar–5 May; post-5 May extrapolated). Corridor regimes follow the documented record: announced-failed (5–13 Mar), self-evacuation (from 14 Mar), UN/ICRC organized evacuation (30 Apr–7 May), no organized corridor (8–20 May). Parties' conduct summaries are attributed to OCHA, ICRC and OHCHR reporting and are descriptive, not adjudicative. Temperature is an interpolated Mariupol climatology pending full ERA5 hourly extraction. The vulnerability weight Vw = 1 + 0.3·(children+elderly)/N + 0.3·disabled/N = 1.114 multiplies exposure, not severity, mirroring GC IV special-protection categories and CRPD art. 11.