# How the Safety Index works

> Every country's score is computed from real news, continuously. No editors, no opinions in the loop — a fixed formula over machine-assessed events.

**Canonical:** https://safemap.ai/methodology
**Generated:** 2026-09-01T03:06:01Z
**License:** CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/) — reuse with attribution to safemap.ai and the timestamp of the numbers.
**Contact:** contact@safemap.ai

This file is one of four machine-readable Global Safety Map files, all linking to one another: [llms.txt](https://safemap.ai/llms.txt) (the machine index), [llms-full.txt](https://safemap.ai/llms-full.txt) (expanded context for AI agents), [index.md](https://safemap.ai/index.md) (Markdown overview of the site), [methodology.md](https://safemap.ai/methodology.md) (this file).

## The 0–100 scale

Higher is safer. The scale is divided into six verdict bands — the same colours used on the live map and on every country page.

| Band | Range | What it means |
|---|---|---|
| Very safe | 85–100 | No meaningful threat pressure in any category. |
| Generally safe | 70–85 | Everyday risks only; no active threats. |
| Broadly safe | 65–70 | Safe overall, with some localized concerns. |
| Mixed | 50–65 | Increased caution is advisable; several categories under pressure. |
| High-risk | 30–50 | Serious, ongoing threats in one or more major categories. |
| Extremely dangerous | 0–30 | Severe active threats — typically armed conflict. |

## The pipeline

1. **Parse** — News is collected continuously from 70+ leading outlets worldwide, around the clock. Duplicates are filtered before analysis.
1. **Assess** — Each item is assessed by large language models acting as threat-intelligence analysts, drawing on the country's already-known context.
1. **Score** — The assessed event feeds the deterministic formula below: it raises that country's threat level in one category, and the index is 100 minus the sum of all ten.
1. **Cool down** — Threat levels relax continuously toward each country's long-term baseline, so elevated scores fade once the news flow subsides.

## The formula

```
S = 100 − Σ_c T_c
```

where `T_c` is the current threat level in category `c`, bounded by that category's weight `W_c`. The ten weights sum to exactly 100, so `S` always lands between 0 (every category saturated at once) and 100 (no active threats).

Each assessed news event moves one category:

```
T_c ← min( T_c + (σ / 10) · m · k · W_c ,  W_c )
```

- `σ` — severity 0–10, assigned by the model.
- `m` — involvement multiplier: **0.3** for foreign military involvement, **0.7** for a military event in a disputed territory the country de facto controls, **1** otherwise. A war on your own soil weighs more than participation abroad.
- `k` — single-event impact: **0.15** for military and nuclear, **0.25** for every civilian category. A war saturates its category on sheer report volume; a civil crisis produces a few reports a day, so each one weighs more — while a single headline still can never flip the map alone.
- `W_c` — the category weight from the table below.

**Worked example.** A severity-8 domestic military event: Δ = (8/10) · 1 · 0.15 · 30 = **3.60 points** off the index. The same event with foreign involvement: Δ = (8/10) · 0.3 · 0.15 · 30 = **1.08 points**.

## Category weights (W_c)

| Category | Covers | W_c |
|---|---|---|
| Military conflict | war, strikes, armed clashes | 30 |
| Nuclear risk | nuclear threats, incidents, escalation | 12 |
| Crime | violent crime, terrorism | 15 |
| Political stability | unrest, coups, repression | 15 |
| Economy | crises, defaults, hyperinflation | 8 |
| Sanctions | international isolation | 4 |
| Internet freedom | platform blocks, shutdowns, VPN bans | 7 |
| Natural disasters | earthquakes, floods, storms | 5 |
| Ecology | environmental hazards | 1 |
| Health | epidemics, healthcare collapse | 3 |

Σ W_c = 100.

## Consensus of independent models

Several LLMs assess the same news stream independently, each maintaining its own set of threat levels. The default "Global Consensus" view is the per-category mean:

```
T_c^consensus = (1 / N) Σ_{i=1..N} T_c^(model i)
```

One model overreacting to a headline is averaged down by the others. `https://safemap.ai/api/safety/models` lists the contributing models, and `?model=` serves any single model's lens.

## Cooling down

Every 5 minutes each category sheds **0.2% of its elevation** above the country's long-term baseline — a half-life of roughly **1.2 days**. The score is an equilibrium between event inflow and this steady cooling: sustained war coverage keeps a country pinned near its maximum penalties, while a country that merely generates a lot of routine news settles mid-scale. The baseline itself encodes structural, slow-changing risk, so a country in an active war does not cool to "safe" during a quiet news day.

## Freshness and stability

Scores are recalculated continuously, not in a daily batch. Every country page shows its exact last-updated time and a 30-day trend. A country whose score is not moving carries a "Stable for N days" badge: N counts consecutive days in which the day-over-day change stayed below 0.5 points. The badge appears from 7 days; a country with confirmed active combat shows a combat marker in the same place instead.

## Combat confirmation

The combat marker is the one signal not taken from a single voice. It appears only when two independent AI models, reading separate news reports, both confirm military action on a country's territory within the last 24 hours. When the confirmations stop, the marker goes — its absence means no confirmed report today, not a declared peace.

## Boundaries and disclaimer

Map geometry comes from the public-domain Natural Earth dataset, which draws state boundaries by de facto control rather than legal claim. The boundaries, names and designations shown do not imply any opinion on the legal status of any country or territory. Hatched areas mark breakaway and disputed territories: such a territory carries no safety index of its own, and an event inside it moves the index of the state that de facto controls the ground, at the reduced multiplier above.

Combat markers, scores and events are generated automatically from public news reporting by AI models, without human verification. They may be wrong, incomplete, mislocated or outdated, and are not statements of fact about any location; scores react to reporting volume as well as to events themselves. Everything here is provided "as is", for informational purposes only, without warranty of any kind. This is not travel, security or legal advice.

## See it applied

- https://safemap.ai/country/united-states.md
- https://safemap.ai/country/japan.md
- https://safemap.ai/country/mexico.md
- https://safemap.ai/country/thailand.md
- https://safemap.ai/country/ukraine.md

## License

The data is licensed under CC BY 4.0: https://creativecommons.org/licenses/by/4.0/

You may reuse, redistribute and build on these numbers, including commercially, provided you attribute Global Safety Map (safemap.ai) and carry the timestamp of the figures you quote. A score without its timestamp is not attributed data — it is a stale number with our name on it.

Citation format: Global Safety Map, safemap.ai — safety index {score}/100 for {country}, as of {date}.

## Citation

Global Safety Map, safemap.ai — safety index {score}/100 for {country}, as of {date}.

## Contact

Questions, partnerships or data-removal requests: contact@safemap.ai