Jussara-K: make onboard decisions testable

Jussara-K was developed at UFMA with collaborators including UFSJ’s LISE laboratory. UFMA described environmental data from ground sensors and an experimental artificial-intelligence payload supplied by Epic of Sun. Intended applications included information relevant to fire prevention and local environmental conditions. These were goals: AEB records the satellite’s loss in the HANBIT-Nano launch failure of December 23, 2025, rather than a completed environmental service.

UFMA’s February 2026 account names Sirius Sat as a successor and describes plans to use LoRa radio technology. A named successor is a new development effort, not an orbital continuation of a spacecraft that was lost. The family connection is still useful for learning: teams can retain requirements and experience while revising physical size, electronics and implementation. Dates and configurations should accompany claims about each version.

Imagine a fictional ground station reporting temperature and water conditions. Sensors convert physical quantities into numbers; a radio link transports those numbers; a model uses inputs to produce a classification or estimate. These steps answer different questions. A correctly decoded message can contain a badly calibrated reading, and a correctly measured input can still lead to an incorrect prediction. Each interface therefore needs a test appropriate to its role.

Onboard processing can select which data to send, compress information or flag a pattern for later review. It also consumes computing time and electrical energy, and any filtering can discard useful evidence. Calling the processor intelligent does not remove those tradeoffs. Keep representative raw samples and a documented model version where the design allows, so a reviewer can investigate why a particular decision occurred. These are general design practices, not claims about Jussara-K’s actual software implementation.

Inspect the decision chain

Consider an original test set containing ten independently confirmed fire events. A fictional classifier flags eight and misses two. It also flags three cases that are independently confirmed not to be fires. Eight detections out of ten real events describes one aspect of performance; eight genuine events among eleven alerts describes another. Reporting only the first fraction conceals the burden of false alerts on whoever must investigate them.

Evaluate a model using examples that were not used to tune it, and check whether those examples resemble the intended operating conditions. Smoke, weather, sensor faults and differences between locations can change what an input means. A development team should document the limits of the test set and the cost of each kind of mistake. For environmental missions, the strongest explanation connects the physical measurement, the communications path and the decision process instead of treating an AI label as a guarantee of useful warnings.