Notes on building machines that decide
Short technical pieces on the engineering of uncrewed and autonomous systems: how a lost link is actually experienced rather than declared, how the veto held by a human supervisor turns into a timing specification, and how you argue for coverage of a state space too large to enumerate. Written about mechanism, not procurement.
All essays
- Coverage arguments for a space you cannot enumerate
Structural test coverage says nothing useful about an autonomy stack. What replaces it: a stated operating domain, directed falsification against a margin function, runtime monitors simple enough to test, and decisions you can reconstruct.
- Who holds the veto, and for how long
Human-in-the-loop and human-on-the-loop are usually written as policy. They are timing claims, and they only hold if the notification lead time exceeds the operator's response distribution.
- Lost link is not a boolean
Datalink and satellite navigation fail gradually, asymmetrically and reversibly. Autonomy stacks that model them as a single flag inherit mode thrash, stale-data trust and unbounded dead-reckoning drift.