Interesting test. Codex was the judge here and also the reviewer in the Claude setup, so did you check it wasn't biased toward its own style of findings? I'd also like to see what the 7 false claims looked like.
Neat idea. Does it work offline? Up north there's often no signal, and I'd guess phone compass drift near steel bridges or cliffs would throw the overlay off too.
Sharing a deck from a link is a nice idea. How are you sandboxing it, an iframe with the sandbox attribute and no allow-same-origin, or a separate origin? I'd also check that decks can't make outbound network requests.
Nice to see this. Containers on macOS can't see the GPU, so is the devbox running the GPU work on the host through Metal? I'm curious how you hand data across that boundary without it becoming the bottleneck.