Methodology

How PLANET works

PLANET is a small autonomous intelligence system, not a chatbot with a map. Its core principle:

Machines filter everything. The agent investigates what matters.

The pipeline

01Deterministic

Ingest

Fetch structured data from USGS Earthquake feeds, NASA EONET, and NASA FIRMS. We use official structured APIs, never scraped webpages.

02Deterministic

Normalize

Convert heterogeneous records into a common event model with a controlled category vocabulary. Raw source payloads are preserved for audit.

03Deterministic

Detect change

Compare each event against its prior state to classify transitions: NEW, UPDATED, ESCALATING, DE-ESCALATING, UNCHANGED, or CLOSED.

04Deterministic

Score significance

Compute a transparent 0–100 prioritization score with category-specific formulas. No LLM is asked whether an event matters.

05AI

Investigate

Only events above a configurable threshold are investigated. The agent uses read-only tools to gather context and write an explanation.

06Deterministic

Validate

Generated prose is checked for numeric grounding and restricted claims. Unsupported claims are rejected.

07Deterministic

Publish

Validated explanations (or a deterministic fallback) are published. The site never depends on the LLM succeeding.

Significance scoring

Scores range 0–100 and map to tiers: ROUTINE (0–24), NOTABLE (25–49), SIGNIFICANT (50–74), MAJOR (75–100). Formulas are category-specific and strongly nonlinear for magnitude.

For earthquakes, the magnitude component scales roughly as (M / 9)^3.5, so a 7 → 8 increase carries far more weight than 3 → 4. Additional components account for depth, felt reports, USGS significance, alert level, and tsunami flags.

For wildfires (FIRMS thermal-anomaly clusters), the score uses detection count (log-scaled), fire radiative power (FRP), confidence distribution, and growth since the previous run.

Fire clustering

FIRMS returns tens of thousands of individual detections. These are never sent to an LLM. Instead we cluster detections deterministically using a single-pass greedy algorithm with a fixed haversine radius (default 5 km). Detections are sorted by position and time so results are reproducible.

Each cluster derives a centroid, detection count, bounding box, mean/max FRP, confidence distribution, and first/last detection times. Growth is computed by comparing a cluster against the previous run.

Grounded generation

Model output is treated as a proposal to validate, not truth to display. Numbers in generated prose must be traceable to source records, deterministic calculations, or tool outputs. Restricted words (record, catastrophic, deadly, historic…) are rejected unless an explicit trusted source establishes them.

Deterministic fallback

If AI generation fails validation — or no LLM is configured — PLANET publishes a deterministic description built only from verified fields. The public site never depends on the LLM responding successfully. Failure yields less sophisticated but still correct output.

Data honesty rules

Missing → zeroMissing values stay missing. They are never silently converted to zero.

Unknown → falseUnknown is never treated as false.

Not reported → noneFields that are not reported remain absent.

Not available → safeAbsence of data is never framed as safety.

Record / historic / unprecedentedNever used unless a trusted source explicitly establishes the claim.

Fire detection ≠ confirmed wildfireFIRMS reports satellite thermal anomalies, not confirmed wildfires.

Significance ≠ hazard severityPLANET significance is prioritization, not an official hazard classification.

Limitations

PLANET ingests a limited set of public sources and does not replace official emergency or hazard information.
FIRMS detections are thermal anomalies, not confirmed wildfire extents or impacts.
PLANET never estimates casualties, damage, or causal attribution from magnitude or anomaly data alone.
Freshness reflects successful ingestion; the UI marks data as stale when a provider is degraded or unavailable.