Edge inference after the demo

Measure on target

Edge inference competes with flight software for power, memory, thermal headroom and update authority. Average benchmark speed hides worst-case mission cost.

Target protocol

Measure cold start, peak memory, sustained power, thermal throttling and corrupted-input behavior on target hardware. Exercise signed rollback and a deterministic fallback.

Release gate

State when inference abstains and what the mission does next.

Challenge the conclusion

Some missions justify aggressive onboard intelligence. The burden of proof rises with irreversibility.

Use this in a working session

Run the worst input at the worst thermal condition while the rest of the mission workload is active. Review resource contention, not only inference speed.