Layering in the container spec is achieved by overlaying each layer's filesystem (a tarball, I think) over each layer below it. If file "a" is modified in layers 3 and 5, the resulting container will have data for both versions but reading "a" in the container will return version 5.

Docker exploits this to figure out when it can cache a layer, but building a container is different than running one because changing the underlying file system can change what a command outputs. If you're running a container, changing one deeply buried layer doesn't change the layers above it because they're already saved.