Thanks for the feedback!

The thing that I really like about this framework is that it can be implemented in increments and by multiple independent developers. Basically, I see an implementation roadmap that looks something like this:

1. Define/publish the custom_json protocol.
2. Developers start creating games for the surveyors - and multiple models are already available for this: solitaire vs. dual player from my post, or tournaments as you suggest here. I'm sure there are many other possibilities, too. Your tournament idea might be the best way to "bootstrap" it, since beneficiary rewards from analysts aren't available yet at this point. Or maybe just a daily post with a summary/analysis of a single surveyor's findings (top-5 overvalued, top-5 undervalued, something like that...)
3. As the various games achieve adoption, developers can create additional tools for analysts to aggregate, summarize, and report the findings from multiple surveyors, and to share beneficiary rewards with the surveyors who they team with.
4. As the analyst reporting quality improves, top tier stakeholders can begin using that to inform their voting. (again, possibly with the aid of new tools from developers)
5. Once the whole system is in place (and before), developers continuously tune the reward structures that were put in place during the earlier phases. As stated above, the goal here is to align rewards with the players' contributions towards returning rewards from overvalued posts back to the rewards pool.

It might take years to fully implement, but it's fairly easy to get started and to make frequent incremental improvements.

The weak link of the idea is the implementation of the final decision, that is, there must be some powerful account that will give the downvote.

This is true, but hopefully not insurmountable. And even without the downvoting accounts, at least we can get a better understanding of the scope and scale of the problem. This understanding might lead to other possible solutions that we can't be aware of yet.

Not all Steemit users are good authors, but they have to post something because it's the only way for them to "mine" Steem. I think such a category of users would be happy to compete for rewards by evaluating other people's content.

This is a really good point that hadn't occurred to me, but I definitely think that you're correct. It's in line with curation rewards, but you don't have to start with a large stake in order to collect rewards.

Undoubtedly, the tournament table should be compiled not by indicators of activity, but by indicators of high-quality work. How to assess quality? It should probably be the deviation of the ratings given to a post by a particular user from the average rating of the post, or something similar.

This is also a good point, and your suggestion is in line with my thoughts on the topic. Also, even if we don't know an answer right now, with continuous experimentation and improvement we can discover a variety of good solutions over time.