Sensor-to-shooter architectures are usually drawn as a diagram and argued about as a latency figure. The harder question is what happens to accountability once the loop is short enough that nobody has time to think.

The mission threads that drive the design are the ones with the least slack: counter-UAS, suppression of enemy air defences, maritime interdiction. For a time-sensitive target you are working to roughly a second end to end, with perhaps 50 to 200 milliseconds of that available for inference at the edge. Almost every other choice in the architecture follows from that budget.

Detection belongs on the sensor

EO, IR and RF payloads run pruned and quantised models locally. That keeps the link clear, and more importantly it keeps the system useful when the link has gone. Published edge inference benchmarks such as MLPerf Inference: Edge are a reasonable sanity check on whether a claimed figure is achievable on the hardware you can actually fly.

A fabric, not a pipeline

Publish and subscribe with deterministic quality of service, attribute labels on every object, and alignment to the relevant NATO STANAGs so that full-motion video and ISR holdings are ordinary traffic rather than special cases to be engineered around later.

The decision layer is where the argument actually is

Multi-model fusion and Bayesian tracking produce proposals, not orders. Each one carries a confidence band and is checked against the rules of engagement in force at that moment. A human stays on the loop, and the loop is logged in a form that survives scrutiny: tamper-evident records, risk controls drawn from the NIST AI Risk Management Framework, and the responsibility principles set out in the UK Defence AI Strategy.

Cryptography outlives the platform

Anything fielded this decade will still be flying when today’s assumptions have expired, so the stack is built to swap algorithms rather than to trust one. Post-quantum migration follows NIST PQC, and access control follows NCSC zero trust rather than network position. Coalition readiness is proved the same way, through NATO Federated Mission Networking spiral profiles tested in the build pipeline instead of asserted in a document.

What it buys

Sub-second proposals with stated confidence. A decision trail that stands up to legal review months later. Enough autonomy at the edge to keep operating through jamming and node loss. None of it is exotic technology; it is a set of decisions about where computation sits and who remains answerable for the result.