The subject matter changed. The unit of curiosity did not: what does this abstraction actually prove, and what test would expose the gap?
Early security and protocol work taught me to separate the property a system claims from the property its mechanism actually enforces. A signature proves authorization, not truth. A collateral invariant proves local accounting, not a remote-chain event. That distinction became the seed of everything after it.
BRIDGE-bench made the failure mode explicit: an evaluation can look quantitative while measuring metadata mixed into its own prompt. Discovering severe leakage in 13 of 24 examples forced me to weaken a result I had published. The important artifact became the correction and the controls, not the flattering score.
Materials screening reinforced the distinction between prediction and confidence. The pretrained mean was useful; the tested MC-dropout uncertainty estimate was anti-correlated with error. A number named “uncertainty” only earns that interpretation after calibration.
roce-preflight had 197 passing tests before real RDMA execution found four defects and rejected the first attempted fix. That changed how I think about mocks: they are excellent tools for logic, but weak independent witnesses when the implementation and fixture share the same mistaken model of an external system.
Aiur is intentionally reduced to one risky physical interaction—recovery—because a rendering cannot validate a docking architecture. Design, CAD, simulation, and controller logic are useful intermediate artifacts; telemetry is what will decide whether the architecture survives contact with reality.
Through-line
Protocol invariant → evaluation validity → calibration → hardware boundary → physical falsification. The portfolio is broad by domain, but narrow by question.
Start with the evidence → · see what changed when evidence disagreed →