When a satellite-AI result is an alert—and when it is evidence

An alert is a prompt for action or inspection

An alert says that a model found a pattern meeting a threshold: a possible burn scar, flooded area, unusual surface change or image worth transmitting. It can be useful before anyone knows whether the event is real. Its quality is measured with operational questions such as how many true events it catches, how often it cries wolf, how quickly it arrives and what happens after it fires.

A representation model such as TESSERA does something even earlier in the chain. It compresses a year of Sentinel-1 and Sentinel-2 observations into a numerical embedding for each location. A task-specific model and labels are still needed to turn that representation into a map or alert.

Sources: TESSERA: Temporal Embeddings of Surface Spectra for Earth Representation and Analysis · TESSERA project documentation

Evidence supports a specified claim

Evidence begins when the output is tied to a clearly defined claim and checked against an appropriate reference: field measurements, expert annotations, another instrument, later observations or a held-out labeled dataset. The test population, geography, time period and decision threshold matter. A strong benchmark in one region does not establish the same performance elsewhere.

Independent observations can confirm an event without proving the model caused a better outcome. Likewise, a visually plausible colored embedding is not a land-cover measurement. It is a learned representation whose meaning depends on the downstream test.

Sources: TESSERA: Temporal Embeddings of Surface Spectra for Earth Representation and Analysis

The same output can move through both roles

A model result may begin as an alert, guide a person toward the relevant imagery, and later become one input to a documented evidence record. Bright keeps those stages separate: signal, verification, interpretation and consequence. Speed is valuable; provenance is what lets someone audit what the speed produced.

Sources: TESSERA project documentation

What this guide does not establish

• There is no universal threshold that turns an alert into evidence; validation must fit the claim and intended decision.
• Reference labels can themselves be incomplete, delayed or wrong.
• TESSERA embeddings are inputs to downstream analysis, not direct proof of a fire, flood, crop type or land-cover change.

Primary sources

Source-reviewed 2026-09-19. Source review means Bright compared this explanation with the linked records; it is not independent replication or expert peer review by Bright.

Continue into the evidence