As long as the model itself understands that when it says "foo bar" in its own reasoning trace it means "pursue theory foo; no, this didn't lead anywhere, let's backtrack and pursue theory bar", where's the loss in clarity?