Confidence and Severity Scoring: GeoSentinel Reference

Every event in GeoSentinel carries two normalized numeric scores: severity and confidence. Both are FLOAT values constrained to the range [0.0, 10.0] by CHECK constraints on both events_canonical and incidents. Severity measures the real-world impact of the event, derived from source-native physical or humanitarian metrics. Confidence measures how much trust to place in the observation, derived from the independence class of the contributing sources and — for media-derived sources — how many reports likely share the same origin article. These two scores enable API consumers to filter and prioritize incidents without understanding the specifics of each underlying data source.

Severity

Severity is calculated by each source mapper by translating a raw physical or humanitarian metric into the [0.0, 10.0] scale. The goal, as defined in the data model spec, is to enable meaningful comparison of severity across event categories — a severity 7.0 earthquake and a severity 7.0 conflict should represent roughly equivalent levels of impact.

Conflict (ACLED)

ACLED conflict severity is derived from the reported fatality count:

Fatalities	Severity
0	1.0
1 – 5	3.0
6 – 25	5.0
26 – 100	7.0
101 – 500	8.5
> 500	10.0

Conflict (GDELT)

GDELT conflict severity is derived from the Goldstein scale, a numeric measure of the theoretical impact of an event on a country’s stability (range approximately −10 to +10, where more negative = more destabilizing):

Goldstein Scale	Severity
< −8.0	10.0
−8.0 to −6.0	8.5
−6.0 to −4.0	6.5
−4.0 to −2.0	4.5
−2.0 to 0.0	2.5
≥ 0.0	1.0

Earthquake (USGS)

Earthquake severity is mapped from the reported magnitude. The mapper rounds the raw magnitude up to the next integer boundary before lookup, so a magnitude 5.8 is evaluated as 6.0:

Magnitude (ceiling-rounded)	Severity
< 4.0	1.0
4.0 – 4.9	2.0
5.0 – 5.9	4.0
6.0 – 6.9	6.0
7.0 – 7.9	8.0
≥ 8.0	10.0

A magnitude 6.8 earthquake rounds up to 7.0, mapping to severity 8.0. A minor magnitude 3.5 tremor rounds up to 4.0, mapping to severity 2.0.

Wildfire / Thermal Anomaly (FIRMS)

FIRMS provides Fire Radiative Power (FRP) in megawatts, a direct measurement of energy released by the fire. GeoSentinel buckets this into severity levels:

FRP (MW)	Severity
0 – 4 MW	1.0
4 – 8 MW	2.5
8 – 25 MW	5.0
25 – 50 MW	7.5
≥ 50 MW	10.0

A low-intensity detection under 4 MW maps to 1.0. A large, high-intensity fire at or above 50 MW maps to the maximum severity of 10.0.

At the Incident Level

When multiple events are assigned to the same incident, two severity fields are maintained:

Field	Calculation	Meaning
`severity_max`	`max(events.severity)`	The peak severity recorded at any point in the incident’s history. Never decreases.
`severity_latest`	Severity of the most recently linked event	The current intensity of the incident. May decrease if the situation de-escalates.

severity_max is appropriate for alerting and ranking (it captures the worst the incident has been). severity_latest is appropriate for displaying current status in a live dashboard.

Confidence

Confidence measures the reliability and independence of the sources supporting an incident. The fundamental principle — Design Decision D3 — is that two media_derived sources reporting the same event are almost certainly extracting from the same news article and therefore do not constitute independent corroboration. Confidence scoring penalizes this correlation.

Independence Classes

Each source is assigned an independence class that determines its base weight in the confidence calculation:

Class	Weight Factor	Sources
`sensor`	×2.0	FIRMS, USGS, ADS-B, MarineTraffic — physical sensor data
`field_reported`	×1.5	ACLED — field-verified reports from specialized organizations
`media_derived`	×0.5	GDELT, Liveuamap — NLP-extracted from news media

SOURCE_INDEPENDENCE_CLASS = {
    "usgs":         "sensor",
    "firms":        "sensor",
    "acled":        "field_reported",
    "gdelt":        "media_derived",
    "adsb":         "sensor",
    "marinetraffic":"sensor",
    "liveuamap":    "media_derived",
}

INDEPENDENCE_FACTORS = {
    "sensor":         2.0,
    "field_reported": 1.5,
    "media_derived":  0.5,
}

6-Hour News Cycle Penalty

media_derived sources within the same 6-hour window from the same source are assumed to originate from the same news cycle. Additional articles from that window add very little independent information and are penalized with a 0.1× multiplier:

def compute_confidence(events: list[EventsCanonical]) -> float:
    score            = 0.0
    seen_media_cycle = set()

    for event in events:
        source_class = SOURCE_INDEPENDENCE_CLASS.get(event.source, "media_derived")
        factor       = INDEPENDENCE_FACTORS.get(source_class, 0.5)

        if source_class == "media_derived":
            # Round down to the nearest 6-hour boundary
            cycle_window = event.event_time.replace(
                hour     = (event.event_time.hour // 6) * 6,
                minute   = 0,
                second   = 0,
                microsecond = 0,
            )
            cycle_key = f"{event.source}:{cycle_window.isoformat()}"
            if cycle_key in seen_media_cycle:
                factor *= 0.1   # strong redundancy penalty
            seen_media_cycle.add(cycle_key)

        score += factor

    # Normalize: a single sensor observation (factor 2.0) → confidence ~6.7
    # Two sensors → confidence ~10.0 (capped)
    # Ten correlated media articles → well below two field_reported
    max_expected = 3.0
    return min(score * 10 / max_expected, 10.0)

Calibration

The max_expected = 3.0 calibration constant was set so that:

1 sensor source (FIRMS or USGS alone): score 2.0 → confidence 6.7
2 sensor sources (FIRMS + USGS): score 4.0 → confidence 10.0 (capped)
1 field_reported (ACLED alone): score 1.5 → confidence 5.0
10 correlated media articles (GDELT, all same 6-hour window): first gets 0.5, each subsequent gets 0.05 → total ~0.95 → confidence 3.2

This satisfies Decision D3: ten GDELT extractions about the same event produce a lower confidence than two ACLED field reports.

Practical Examples

Earthquake confirmed by two networks

USGS primary report + USGS secondary network report (both sensor, 2.0 each). Score: 4.0 → confidence: 10.0

Conflict with field + satellite

ACLED field report (field_reported, 1.5) + FIRMS thermal anomaly (sensor, 2.0). Score: 3.5 → confidence: 10.0 (capped)

Media-only conflict

5 GDELT extractions in the same 6-hour window. Score: 0.5 + 0.05×4 = 0.7 → confidence: 2.3

Single FIRMS hotspot

One FIRMS satellite detection, no corroborating sources. Score: 2.0 → confidence: 6.7

Using These Fields in Queries

The min_confidence and min_severity query parameters on the incidents endpoint let you filter by these scores directly.

# Only show high-confidence incidents (sensor or field-verified corroboration)
curl "http://localhost:8000/v1/incidents?min_confidence=7.0"

# Severe conflict incidents (26+ fatalities or equivalent)
curl "http://localhost:8000/v1/incidents?category=conflict&min_severity=6.0"

# Major earthquakes with high confidence
curl "http://localhost:8000/v1/incidents?category=disaster_natural&min_severity=7.5&min_confidence=8.0"

# Active incidents in an area with meaningful severity
curl "http://localhost:8000/v1/incidents?status=open&min_severity=3.0&bbox=-10,35,40,70"

fatalities_total on an incident is calculated as MAX(events.fatalities), not SUM. When ACLED, GDELT, and Liveuamap all report the same battle with 47 fatalities, summing them would yield 141. GeoSentinel takes the maximum (47) to represent the single best estimate of the true toll, avoiding inflated figures caused by multi-source reporting of the same event.

Get Started

Data Sources

Core Concepts

Deployment

Confidence and Severity Scoring: GeoSentinel Reference

Severity

Conflict (ACLED)

Conflict (GDELT)

Earthquake (USGS)

Wildfire / Thermal Anomaly (FIRMS)

At the Incident Level

Confidence

Independence Classes

6-Hour News Cycle Penalty

Calibration

Practical Examples

Earthquake confirmed by two networks

Conflict with field + satellite

Media-only conflict

Single FIRMS hotspot

Using These Fields in Queries

Build docs developers (and LLMs) love

Get Started

Data Sources

Core Concepts

Deployment

Documentation Index

​Severity

​Conflict (ACLED)

​Conflict (GDELT)

​Earthquake (USGS)

​Wildfire / Thermal Anomaly (FIRMS)

​At the Incident Level

​Confidence

​Independence Classes

​6-Hour News Cycle Penalty

​Calibration

​Practical Examples

Earthquake confirmed by two networks

Conflict with field + satellite

Media-only conflict

Single FIRMS hotspot

​Using These Fields in Queries

Build docs developers (and LLMs) love

Severity

Conflict (ACLED)

Conflict (GDELT)

Earthquake (USGS)

Wildfire / Thermal Anomaly (FIRMS)

At the Incident Level

Confidence

Independence Classes

6-Hour News Cycle Penalty

Calibration

Practical Examples

Using These Fields in Queries