Scientific model cards
Contract review: 14 September 2026. These cards describe the reviewed candidate models and isolated evaluation results. They do not certify that prospective validation is already running live. Predictive accuracy and probability calibration are not established by software checks. This reference is currently available in English.
Physical priority index — version 5
Ranks recorded physical events for attention. I is the hazard-specific intensity index; P is a normalized distance-weighted urban exposure index; C is a rule-based cascade scenario weight; H is a recorded archive reference or current-batch anomaly fallback. None is a probability of harm. Historical version 4 remains identifiable and is not silently presented as version 5.
CS = round(0.40 I + 0.30 P + 0.20 C + 0.10 H), bounded to 0–100.
Inputs and coverage. Source observations retain their units and timestamps. Fire active-pixel counts are not hectares. The historical earthquake reference transforms original archived USGS magnitudes onto the version 5 intensity scale (reference version usgs-raw-magnitude-linear-v5-90d-2), using selected episodes in a closed 90-day window, requires at least 30 episodes and 14 reporting days, and has a 36-hour freshness limit. Its version, window and availability must accompany the score; otherwise the fallback is identified. These requirements do not establish worldwide completeness.
Reproducible example. Synthetic arithmetic check: I=40, P=25, C=20, H=10 gives round(16+7.5+4+1)=29. This verifies arithmetic, not predictive skill.
Limitations. Thresholds and weights are heuristic. Equal scores across different hazards do not imply equal losses, impact or likelihood. Source support, input completeness and freshness are separate evidence dimensions. A legacy confidence scalar measures a mixture of data-support signals, not accuracy or calibrated confidence.
News attention index — version 1001
Describes the volume and diversity of recorded media attention. It is a separate index family and must not be ranked as physical severity.
News = min(min(3 × articles, 50) + min(5 × source labels, 30) + min(5 × languages, 20), 80).
Inputs and coverage. Recorded article counts, source labels and languages. Syndication, language coverage and common upstream reporting can make these inputs dependent.
Reproducible example. Synthetic example: four articles, two source labels and one language produce 12+10+5=27.
Limitations. Source labels are not proof of editorial independence or event verification. More coverage does not establish a larger disaster; absent coverage does not establish absence of an event.
Calamity Forecast Index — cfi-2
A country/hazard activity and scenario index, not a probability that an event will occur within seven days. Observed activity, target cascade scenario and target localization are distinct channels. The current candidate selection covers at most 100 countries with recorded activity; an absent pair is unassessed.
CFI = round(0.30 C + 0.25 F + 0.20 S + 0.15 T + 0.10 N). Components are indices from 0 to 100.
Inputs and coverage. C uses target scenario weights. F = min(100, round(100 × log(1 + episodeRecords90d / max(1, reportingDays)) / log(101))) and measures recorded episodes per reporting day, saturating at 100/day. Episode identifiers fall back to record identifiers when missing. S is a global calendar-month prior, not a local seasonal calibration. T compares recent eligible activity with its reporting-day baseline. N measures recent media attention level, not acceleration. The five weights remain fixed.
Reproducible example. Synthetic C=40, F=20, S=30, T=10, N=0 yields 12+5+6+1.5=24.5, rounded to 25. A value of 25 does not mean a 25% chance.
Limitations. Reporting density, missing days, coupled inputs and global seasonal priors affect the index. Display the recorded computed_at, valid_until, model_version and coverage. Current outputs have a two-hour validity contract; expired or missing validity is never displayed as current low risk. Refresh cadence is not a guarantee that a new assessment succeeded.
Country activity rating — activity-rating-2
Grades A–E summarize recorded current activity, with A=0–20, B=21–40, C=41–60, D=61–80 and E=81–100. They are not structural vulnerability, resilience, creditworthiness or expected loss ratings. Operational status follows event-severity rules, not a validated prediction of escalation.
A=min(100, ΣeventScore/500 × 100); B=min(100, 20 × cascadeCount + 50 × maxWeight); P=min(100, ΣurbanPEI/1,000,000 × 100); F=min(100, 1.5 × meanCFI). Index = round(0.40 A + 0.25 B + 0.20 P + 0.15 F).
Inputs and coverage. The displayed component values are the four persisted terms actually used. Hazard breakdowns are separate. CFI shares event and cascade inputs with this rating and is not independent confirmation. Missing CFI contributes zero arithmetically while remaining explicitly unassessed in coverage. The selected country scope and two-hour validity accompany each output.
Reproducible example. Synthetic example: two event scores of 100 produce A=40. With the other terms zero, the index is 16. If the CFI input is absent, its coverage remains unassessed; this is not evidence of no future hazard.
Limitations. A grade change can result from reporting, event expiry, corrected data or model changes. Recorded deltas show before/after components and their direction. A dominant event is not proof of causation. The aggregate urban term can count overlapping contributions repeatedly.
Urban PEI and gridded modelled population
Urban PEI is a proximity-weighted index, not affected people or unique people. Zero means no covered urban contribution, and does not establish that nobody is present. An unavailable estimate remains unassessed. Dataset, reference year, geometry and model version identify each estimate.
Urban PEI rounds Σ cityPopulation × distanceWeight × sourceWeight, with a heuristic minimum of 100 when a populated city is within R. Search extends to max(1 km, 1.2R): Gaussian exp(−3(d/R)²) can contribute beyond R; linear and step modes have their own support.
Inputs and coverage. The separate gridded output covers Italy only: WorldPop Italy 2025 R2025A v1 alpha estimates, aggregated by summing 3-arcsecond source cells into 30-arcsecond cells, with no extrapolation outside Italy. Cell centres sample the nearest available USGS ShakeMap grid node at MMI≥5. The method is shakemap-mmi5-worldpop-cell-centres-v1. High-resolution CoverageJSON is preferred, with medium resolution as fallback; product source/version, update/process times and grid spacing accompany the dataset/year, missing cells and coverage. This does not replace urban PEI in the priority index.
Reproducible example. Synthetic grid example: two overlapping footprints covering the same cells with populations 100 and 80 have a union of 180, not 360. A footprint moving to different cells may expose new modelled population even if its total is unchanged; the delta uses newly covered cells.
Limitations. Grid populations are modelled residents, not observed individuals, occupancy, casualties or displaced people. Cell-centre inclusion approximates boundary overlap. Missing footprints, missing cells and places outside Italy are unassessed, not zero. The gridded union avoids repeated cells within the supported grid; it does not identify people moving between cells.
Cascade scenarios and evidence
A scenario describes a plausible secondary hazard and its recorded localization assumptions. A weight of 0.8 is not an 80% occurrence probability. Observed secondary events and rule-generated scenarios remain distinguishable.
Rules assign relative scenario weights; chained weights and scores remain heuristic indices.
Inputs and coverage. Trigger observations, geographic rules, hazard-specific assumptions and recorded source support. Source availability, completeness and freshness must be visible independently of a priority value.
Reproducible example. A high earthquake-to-landslide scenario weight indicates rule-based attention. It is neither an observed landslide nor a calibrated chance of one.
Limitations. Weights have not been shown here to be calibrated against independent target outcomes. Explanatory text must not turn a relative scenario weight into a probability claim.
Validation and provenance
Arithmetic, unit, coverage and replay checks test implementation contracts. Predictive validation requires a defined future target, preserved input vintages, chronological holdouts and comparison with declared baselines. No validated predictive skill or calibrated occurrence probability is claimed by these cards. Computation time, assessment expiry, source observation time and page delivery time are distinct. Historical records lacking a version or validity interval are explicitly incomplete.
Executed retrospective evaluation
A restored public archive contained 614,562 observations across 16 hazard types. All 16 representative shared contracts passed. Ten retrospective current-model cutoffs were evaluated, with 2,500 inputs per cutoff at most; seven reached that cap. Held-out evaluation covered only 37 country/hazard weeks across seven types. The seven-day target was subsequent archived priority, including mixed historical score versions, conditional on reporting coverage. It was not independently observed loss or complete disaster occurrence. Silent windows were not negative outcomes.
Persistence, reporting-frequency and available prior-month baselines were compared on covered windows, with ordinal and threshold metrics and country-block bootstrap intervals. Few geographic blocks limit those intervals. Cascade, seasonal and forecast-proxy ablations and duplicate sensitivity checks ran; triplication changed some rating scores by up to 38 points. The news ablation was uninformative because these retained cuts contained no media inputs. Original model-input vintages were unavailable, so this replay does not establish what a deployed model knew at each historical cutoff. There were zero mature real prospective outcomes in this evaluation.
Limited external comparisons
Seven compatible USGS/PAGER earthquake cases gave a descriptive Spearman correlation of 0.411 between archived priority and maximum MMI, with an approximate bootstrap interval of −0.535 to 1.000. These are different quantities and may share physical inputs; this is not calibrated impact accuracy. A bounded EONET comparison mapped 186 cases (182 fires and four cyclones), of which 175 had approximate archive matches. Catalogue limits, selected geometry points and possible upstream overlap prevent interpreting this as complete recall or independent confirmation.
Amatrice: resolution and product vintage
With the same WorldPop 2025 population grid, the 2016 Amatrice event yielded 107,339 modelled residents in MMI≥5 cells using low-resolution Atlas 2020 shading, and 161,577 using its high-resolution grid. The physical XML grid reproduced the high-resolution result. A different USGS 2016 footprint matching the PAGER product family yielded 2,952,932. PAGER’s MMI class bins imply a broad MMI≥5 interval of 1,338,122–7,091,026 for its own population product. Product vintage, population year and bin boundaries matter: agreement with that interval does not validate exposure accuracy.
Independent raster calculations verified axes, CRS, units and aggregation, with relative population-mass difference about 2.34 × 10⁻¹⁰. Using original 3-arcsecond population cells with the Atlas high-resolution shaking grid yielded 171,283, versus 161,577 after aggregation: about 5.7% lower with coarse cell-centre inclusion in this case. Twenty-eight geometry and union/delta regression assertions passed. These are implementation and retrospective spatial checks, not counts of affected people or a reconstruction of what was known in 2016.
The USGS ShakeMap documentation describes ground-motion products. USGS PAGER combines shaking, population and regional loss models; its validation does not validate Calamity indices. WorldPop methods describe modelled population mapping; each published estimate must retain its actual dataset and year.
Source references: Bondarenko et al. (2025), WorldPop Italy 2025, R2025A v1, DOI 10.5258/SOTON/WP00839; USGS retrospective ShakeMap Atlas; NASA EONET v3 catalogue documentation. Shared source inputs and modelled quantities remain explicit limitations.