Separate the functions
| Layer | Question | Typical machine role |
|---|---|---|
| Perception | What signals or objects are present? | Detection, classification, anomaly screening |
| State estimation | Where are they and how uncertain is the track? | Filtering, association, covariance, track maintenance |
| Interpretation | What activity is consistent with the evidence? | Competing hypotheses and prediction |
| Decision support | Which authorized options appear feasible? | Constraint screening and ranked alternatives |
| Execution support | How does an approved task proceed? | Navigation, routing, formation, deconfliction, health management |
| Force application | Who selects the specific target and initiates force? | Requires explicit policy, authority, legal, and safety design |
| Governance | Who changes models, thresholds, data, and permissions? | Human institutional responsibility across the lifecycle |
“Best capability” is a constrained recommendation problem
The system should first exclude options that fail positive identification, rules, geographic or temporal bounds, authority, civilian-protection constraints, data quality, system health, or safety. Only surviving options may be ranked by mission value, timeliness, availability, confidence, reversibility, and opportunity cost.
Hard legal and command constraints should not be soft penalties that a learned optimizer can trade away.
Uncertainty multiplies across the pipeline
A moderately uncertain detector feeds a tracker; the tracker feeds a behavior model; that model feeds a resource-allocation service. If the interface compresses the result into one precise-looking score, operators may see certainty that the evidence never supported.
Displays should keep detector confidence, track existence, positional uncertainty, identity confidence, data age, model version, source provenance, and policy status separate.
- Show alternative hypotheses and dissenting sensors.
- Expire recommendations when evidence becomes stale.
- Detect out-of-distribution conditions and abstain conservatively.
- Record model, dataset, configuration, policy, and human actions.
- Test the combined human-machine system under deception, overload, and communications loss.
Autonomy is function-specific
A system may navigate, maintain formation, route messages, or return to base autonomously while a human retains force decisions. Another system may be stationary but select and engage after activation. The meaningful question is which function is delegated, within what bounds, with what evidence, and with what opportunity for accountable human intervention.
Human-in-the-loop and human-on-the-loop labels are not guarantees. Control becomes nominal when the person lacks time, evidence, understanding, authority to reject, or a technically reachable intervention path.
Autonomous edge nodes
Distributed autonomous systems can perform local sensing, networking, navigation, task allocation, and health management without streaming all raw data to one headquarters. Their value comes from coverage, reduced bandwidth demand, and graceful degradation.
Their risks include emergent behavior, compromised nodes, inconsistent world models, network partitions, duplicate tasking, and loss of supervisory control. The public research corpus does not establish that unrestricted autonomous target engagement is mature or lawful.
Research basis: KW-RPT-006 and KW-RPT-012.