How to read
the signals.
CaribEye brings public sources into one Caribbean view. It is a demonstration of what AI assisted monitoring can organize and explain. It is not an official warning system, a crime map, or a measure of public opinion.
Country risk: an experimental screening tool
CaribEye displays a 0–100 structural pressure score. It is not a probability of loss, a credit rating, a travel advisory, or an official warning. Higher numbers indicate more pressure in the selected proxies. The transformations, equal weights and display bands are design choices, not statistically calibrated risk boundaries. No investment, lending, insurance or safety decision should rely on this score alone.
The core is the arithmetic mean of four equally weighted pillars: political stability, climate vulnerability, economic pressure and homicide. All four must have usable observations; otherwise the core is withheld. Credit/fiscal pressure and electricity access are separate diagnostics, never silently substituted into the core. Countries are shown alphabetically rather than ranked. Different reporting years and limitations mean even complete scores are not fully comparable.
| Pillar | Source and transformation | Important limit |
|---|---|---|
| Political | World Bank WGI political stability estimate: 20 × (2.5 − estimate), clipped to 0–100. | Approximate −2.5 to +2.5 scale mapped to a screening scale. Governance measurement uncertainty is not propagated. Not an election forecast. |
| Climate | Notre Dame Global Adaptation Initiative Country Index, 2026 release: vulnerability × 100. | Long-term climate vulnerability, including exposure, sensitivity and adaptive capacity. Not current weather or disaster probability. The source has its own estimation and interpolation methods. |
| Economic | GDP growth: 100 × (5 − growth)/15. Inflation: 100 × |inflation|/20. Clip each to 0–100 and average both. | Both required. Growth at 5% or above maps to zero; contraction of 10% or more maps to 100. Inflation magnitude at 20% maps to 100. These are illustrative policy-neutral screening anchors, not universal economic targets. |
| Safety | Intentional homicides per 100,000: value/60 × 100, clipped. | Does not cover all crime, travel safety or location-specific exposure. National reporting definitions and completeness differ. Sixty is an illustrative saturation threshold. |
| Credit / fiscal | Central-government debt/GDP: value/150 × 100. Total external debt service/exports: value/40 × 100. Clip and average available values. | Different institutional coverage and years. One input means partial evidence. Excludes ratings, maturity profile, reserves and market spreads; cannot estimate sovereign default risk. Excluded from core. |
| Infrastructure | Population with electricity access: 100 minus percentage. | Only an access gap; universal access does not establish reliability or storm resilience. Excluded from core. |
Missing data, freshness and confidence
For each input we take the latest non-null observation, preserve its year and link, and refuse to score observations older than four calendar years. The four-year limit is a design choice and can still admit materially outdated data. A failed collection retains available prior evidence marked stale and excludes it from scoring. Missing does not mean zero. A newly downloaded historical value is still historical: the snapshot time is not its observation date. The current coverage is ten monitored countries and territories; Puerto Rico is treated separately and never inherits United States values.
All scores carry limited confidence because these proxies and transformations have not been calibrated against outcomes. The displayed bands are lower pressure (0–24), moderate (25–49), elevated (50–74) and high (75–100). They are interface labels, not official thresholds. No false precision or statistical confidence interval is claimed. Before operational use, validate outcomes, assess weight sensitivity, broaden regional and language coverage, and obtain expert review.
News watch, separate from the score
The news watch identifies exact country assignments and title keywords in the last 72 hours. It removes identical normalized titles but may retain different reports of the same story. Political, climate, fiscal, economic, violence and infrastructure terms identify candidates for review, not verified adverse events. Country mentions are not proof of event location; negation and context are not fully understood. These headlines do not change structural scores. A quiet news sample does not imply safety. Downloadable country evidence includes the source observations, method version and matched headlines for review.
What the view measures
| LANE | SOURCE | MEANING AND LIMIT |
|---|---|---|
| Weather | Open-Meteo | Model weather at selected city coordinates; not a station observation or national alert. |
| Tropical weather | NOAA NHC | Atlantic feed bulletins; routine bulletins are not warnings for a specific island. |
| Earthquakes | USGS | Reported events in a regional bounding box, subject to revision. |
| Media | GDELT DOC API, News Americas Now, Google News RSS | A limited sample of headlines. Simple title rules assign topics. Counts are not all regional news or public interest. |
| Satellite | NOAA/CIRA GOES-19 | Direct Caribbean GeoColor image. Open the NOAA viewer for observation time and animation. |
| Broadcasts | Publisher YouTube channels and websites | Manually reviewed directory. A Streams page can show past or scheduled video; live status and monetization are not verified. |
| Flights | ADSB.lol | On-demand aircraft positions within sampled 250 nautical mile circles. Receiver coverage is uneven. Not schedules, delays or certified navigation data. ODbL 1.0 attribution. |
| Road traffic | Google Maps traffic layer; Jamaica NWA | External current city maps and official road advisory links. No speeds, incident counts or congestion score is stored by CaribEye. |
| Search interest | Apify Google Trends actor | Country, term and time window index after connection. Separately normalized rows are not comparable as a regional ranking or absolute search counts. |
| Perception | Future survey or validated listening study | No score until sampling, language coverage and validation are documented. |
Freshness and gaps
The dashboard shows the snapshot time and provider status. A failed feed may retain its last records marked stale. The satellite image is served directly by NOAA; check the NOAA viewer for its time. An empty feed does not mean an area is safe. The broadcast directory has a separate manual review date. Flight positions are requested when the Flights panel opens or refreshes. If the browser cannot reach the API, the last scheduled snapshot is shown with its observation time; a regional sample is not complete airspace coverage. Traffic colors are read on the provider map at opening time.
How AI fits
The current prototype combines and classifies structured source records with transparent rules. A later AI layer could cluster duplicate reports, summarize source linked evidence, translate regional languages, detect emerging topics and explain uncertainty, with human review before public risk or perception claims.
Next integrations
- Connect the bounded Apify importer and broaden the country sampling plan.
- Add national meteorological and disaster authorities with source specific alert definitions.
- Verify live video through a licensed platform API before adding a “live now” badge.
- Add source quality checks, deduplication and multilingual coverage.
- Design a representative study before reporting perception.
Open source inspiration: God’s Eye View by Bilawal Sidhu. This interface is written for AI BOSS. The upstream source is MIT licensed; third party data and models have separate terms.