As for theory, IMO the biggest concerns are:

1. people doing the work and not being paid
2. people not doing the work, or doing poor or partial work, and expecting to be paid

For (1), SN attempts to make a bounty-poster's past bounties visible so that a bounty-doer can determine if the bounty-poster makes good on their bounties. Notes:

1. We could improve this by giving a bounty-poster an aggregate score like "the pay X% of their bounties"
2. We take a sybil fee of 10% on paid bounties. While we could remove this, it biases bounty-posters away from doing their own bounties and paying themselves for free (ie cheating and maintaining reputation has a non-zero cost at least).
3. We do not have a good solution for many people completing the bounty, so we leave it up to the bounty-poster to solve. If we were to help with this, the best top-of-dome idea is letting someone elect themselves as the sole bounty-doer for a time period and have them put up a bond for the right.
4. We could also have the bounty-poster put up a bond for the right to post a bounty. Then, if they don't elect a winner after enough time has passed, the bond is divided among everyone that attempted the bounty.

For (2), whether the bounty work is done or not is not computer-decidable, because if a computer could determine if work has been done well, the computer would be able to do the work in most cases. Notes:

1. Our system allows for partial work to be paid, by normal tipping, so that seems okay.
2. IME This is best solved by being very explicit about the bounty requirements.

The best bounty system I've ever used, which might also be the only one I've used, is the one employed by 99designs. I once had a logo designed with them and IIRC:

1. they call it a contest and a contest has several rounds
2. you put up the contest requirements and give the platform the contest prize money in advance
3. in every round you whittle down the contest participants and give feedback
4. in the final round, you pick a single or many winners, else all contest participants split the contest prize equally

Given it's a contest, you do tend to get a lot of derivative, low effort work, because contest participants are incentivized to enter multiple contests and not over-index on a single contest. Getting great work done requires a lot of effort specifying requirements and giving very specific feedback at each round.