TOPLOC attempts to detect model substitution, i.e. responses being generated by a different model than requested, it comes with certain caveats, as far as I can

TOPLOC attempts to detect model substitution, i.e. responses being generated by a different model than requested, it comes with certain caveats, as far as I can tell the TOPLOC paper considers verifiable learning / training as out of scope.