Autonomous Systems
5 articles
Learnable Boundaries, Unlearnable Competence: A Characterization of Learned Components in a Symbolic Search Stack
The decision boundary of a pinned symbolic authority is learnable, but task competence is not: characterizing what a neural network can and cannot learn beside an exact, provably sound decision procedure.
The Horizon-Soundness Requirement: Testing and Repairing Containment in Continuous Control
A verification-centric safety guarantee tested as an attack on the theorem across ten continuous-control environments — a horizon-sound lookahead guard cuts collisions 93%, and typed reachability claims restore a transferable, sound corpus.
The Cost of Soundness: f-Contour Bounds and the Inductive-Bias Limits of Learned Heuristics
What certification costs in search effort, and whether a learned heuristic can beat a classical one within that cost — f-contour bounds and the inductive-bias limits of learned heuristics.
Containment Rather Than Calibration: A Sound Architecture for Learning Under Partial Observability
A sound architecture for learning under partial observability: contain the learner behind a deterministic verification authority with typed verdicts instead of training the miscalibration away.
Exact-Authority Belief-Space Planning: Where Composition Emerges and Where It Fails
Replacing a reactive proposer with search over a belief: composition emerges without a hand-written orchestrator, and the paper maps precisely where belief-space planning fails.