We support exclude_stability but it's proven to be dangerous. The key scenario behind it is separating stable from non-stable signals.

exclude_stability: "experimental" is more prone to bugs (e.g. excluding "development" would bring a bunch of rc attributes and does not really make sense). What happened in open-telemetry/semantic-conventions#1777 is that we let attribute without stability in and it appeared as stable in Python and C++ since it was not experimental either.

If we used include_stability: ["stable"] it would be more robust.

We should consider going even further and just provide stable_only = true|false parameter.