Yes, I understand the signal must be hard to fake. But if the concern is merely about optimizing signal quality, wouldn't it be an even stronger mechanism to noticeably couple your payoff profile to a credible mechanism?

Just as a sketch, find some "punisher" that noticeably imposes disutility (like repurposing the signal faker's means toward paperclip production, since that's such such a terrible outcome, apparently) on you whenever you deviate from your purported decision theory. It's rather trivial to have a publicly-viewable database of who is coupled to the punisher (and by what decision theory), and to make it verifiable that any being with which you are interacting matches a specific database entry.

This has the effect of elevating your signal quality to that of the punisher's. Then, it's just a problem of finding a reliable punisher.

Why not just do that, for example?