Systems you can trust, not just demo.

I authored and presented CETA (Cyber Evaluation Test Adequacy), a formal framework that recasts cyber test-and-evaluation sufficiency as a decidable question over an OWL 2 knowledge graph. Coverage in DoD cyber developmental test & evaluation is currently enumerative — techniques ticked off a matrix, requirements checked against a spreadsheet — and those counts break whenever the system model changes. The artifacts that would make traceability mechanical sit in vocabularies that don't map to each other: SysML/UAF for architecture, MITRE ATT&CK for adversary behavior, CWE/CVE for weaknesses, and criticality analyses for missions.

The framework aligns four modular OWL 2 sub-ontologies — system architecture, threat, weakness, and mission — with SWRL-style inference rules that derive every admissible attack path ending at a mission-essential function. Under its “semantic test adequacy” criterion, a test suite is adequate exactly when every such path above a criticality threshold is covered by an executable test; remaining risk becomes an explicitly enumerated gap set instead of an unexamined remainder. For gap closure, the paper proposes a retrieval-augmented LLM that suggests candidate tests but can only emit instances of declared ontology classes, so the same reasoner that found a gap checks every proposal, and malformed proposals are rejected before a human reviews them.

The paper defines a metric suite — semantic coverage ratio, gap-set cardinality, false-test rate, gap-closure rate, re-materialization time after model revision, and traceability completeness — and a staged evaluation design over open reference architectures, with four stated hypotheses and their falsification conditions. It draws on 18 references, builds on current DoD/DoW policy (Cyber DT&E Guidebook v3.0, DoW Manual 5000.103, DoD Cyber Table Top Guide v3.0), and addresses a gap DoD acknowledges: its own documents describe the existing DoD ontology, OACRA, as low technology readiness level. As the workshop's call requested, it is a concept-and-methodology paper — no empirical results are claimed.

The lab's question is the one that matters most for autonomy: not whether a robot can complete a task, but whether you can trust that it will. My work sits in the simulation-first validation loop — building perception and control pipelines for ROS-based autonomous robots and systematically breaking them before they ever touch hardware.

The workflow runs on the Duckietown framework: an image-processing pipeline finds the lane, odometry and IMU data estimate where the robot actually is, and PID control keeps it tracking. The research contribution is in the validation methodology — designing simulation environments that stress the system across dynamic scenarios and identify failure modes early.

Formal methods are the other half of trustworthy autonomy. DESops is a University of Michigan Python library for discrete event systems — finite-state automata, parallel and product compositions, observer computation, supervisory control, and opacity enforcement. I made minor contributions to the library and used it for academic research in formal methods and the control theory of autonomous systems.

Where the CASA-Goes work validates behavior empirically, discrete event systems let you prove properties about it: what a supervisor can permit, what an observer can infer, what an outside party can or cannot deduce about system state. The combination — empirical robustness testing plus formal guarantees — is the direction I find most interesting.