How Can We Best Design the Curation Rewards on Blurt?

Yesterday, we looked at the current curation-rewards-curve, how it is supposed to function and how it currently manifests on Blurt. Today, I wish to look at what "curation" actually means, both in theory and practice.

Currently, Blurt users are experiencing near-stable proportional functions for both the rewards and curation. Rewards are proportional to the mana used in a vote and curation is about half of a vote.

This was a fortuitous accident that allows Blurters to experience a slightly different economy to what was inherited. However, we have a choice to make: keep the current behaviour or revert back to something approximating Hive? Both choices require a few code changes. Let us not forget that the vote-dust-threshold is due to be scrapped as soon as the next hardfork is implemented, so things will change anyway.

Firstly, some semantics.

Is Curation Linear or Flat?

Here is an amusing confusion that arises when talking in words without the benefit of a mathematical formula. I have seen LeoFinance describe their new curation curve as flat, meaning that the percentage voters receive is constant. So, if on Blurt, whether you vote 10 BLURT or 110 BLURT, you'll still get 50% of that back as curation rewards.

But notice that this is not a flat curve; it is not identical for everyone. It is actually linear, or more precisely, it is proportional to each vote. I have seen people arguing over this point and it is pure misunderstanding. In this case, a "flat percentage" means a "proportional curve".

Having got that small issue out of the way early, let's see what the curation "curve" actually does.

The Curation Curve

We looked at this in detail previously, so no need to repeat it all here. What I would like to discuss here is whether the encoded curve achieves what it was designed to achieve. If the answer is negative, then we need to ask ourselves firstly, what is the purpose of curation rewards, and secondly, how to encode the behaviour we wish to see.

The fundamental original idea that has been encoded in Steem since inception is that the curation-curve should encourage "content discovery". Does it actually do this?

Does the Content Discovery Curation Curve Work for Actual Curators?

The word "curation" is probably third on my list of most-abused words in crypto - I'll deal with the other two in further posts! That every vote receives curation rewards is slightly laughable as terminology. Votes are financial transactions that are made for a variety of reasons, most of them can be broken down into either financial or content-based or even social-gratitude.

Content discovery remains a human process guided by human perceptions of quality. The fundamental idea behind the curation-curve is that early voters get a larger share of the cream because their curation has been appreciated by later larger voters. The human content curators are thus performing a valuable social function that is rewarded by increasing the curation-rewards from their own (often modest) vote. But this hasn't worked at all!

It hasn't worked because any simple algorithm will have some optimal strategy, and that optimal strategy is in conflict with the original aims. The obvious first strategy is to vote early on every author who has been earning good rewards - that led to every voting bot voting within seconds of a post being published. So a penalty curve was introduced, the so-called "reverse auction" (which it isn't), which has now been whittled down to a mere 5 minutes. But even then, there remained an optimal strategy, but it became a dynamic strategy and so the game of curation-chasing emerged. It's fun for those designing voting bots but we are now a long way away from the human curator - a long way away from fulfilling the aim of a content-discovery-reward-system.

Just as an aside, large curation accounts knew that their curators needed decent rewards and thus paid them directly from their own, much larger, curation rewards. Hence, some level of earnings and prestige could accrue that was commensurate with the time and skill involved in doing a proper job.

So, back to the curation-curve: what is it for?

What Do Curation Rewards Achieve?

As we have just seen, the algorithms designed to promote content discovery in an ecosystem that rewards good content did not actually achieve their aims in the real ecosystem. Instead, those algos promoted algorithmic optimisation at the cost of human curation. Don't get me wrong, algorithmic optimisation can be a good thing in spreading information across an ecosystem. But herein lies the fundamental problem in the current design: the information that the algorithmic voters use to generate their activity is not, "this is a quality post", but rather "this is a rewarding post".

The link between quality and rewards is only in the minds of human members - it has not been established at the code level. This is not to be overly negative, but rather to stress that the current framework has not worked and will not work as it stands. This avoids prolonged and futile discussions that somehow, somewhere, there is a magical curation-curve that will solve this. There isn't. That means seeking solutions elsewhere, such as strengthening a robust curation network that adds what is missing within the pure algorithm: that rewards flow towards quality.

This is also no magic bullet, but it starts to change the balance towards achieving a genuine link between quality content and rewards. This also means that those who still seek financial-curation-rewards will also see that it is within their own optimisation algorithm to vote such content. Blurt has made a start on this with the curation accounts and it will take some time to attract the curators with both the skills and time to make this work smoothly - we also need to see more quality posts too!

However, all this is still predicated on the curation-curve rewarding early discovery. If we move to a linear curation-curve where every vote is paid the same percentage curation-reward back, then this may seem fairer but still avoids the question of how to reward human curation. Such fairness can also lead to agnostic voting - vote randomly and you still get the same curation-rewards. You can probably sense where this is leading now.

Is a proportional share of curation-rewards actually fair? Will it lead to fair behaviour? I have mentioned before about looking at second and third-order effects, not just assuming that first-order cause-and-effect will be the only consequences. Can fair curation-rewards lead to unfair behaviour?

The world of sport is fair in that anybody can compete if they so wish; it is manifestly "unfair" in that there are ultimately winners and losers. So, is quality curation really a competitive sport? As "quality" is only expressed in "rewards" I suggest that the curation-competition has been framed within the wrong metric. Imagine if the winner of a football match was decided by the number of fans in the stadium and not the game itself.

Now, from a human perspective, it does feel much better knowing that when I read and vote on a post I'll be getting a decent share of the curation-rewards. It is a more human way to react. It may not be the best "content discovery" reward-system but it feels like a good "content appreciation" system. But what about the algorithmic curation-optimisers?

The most obvious problem with a flat curation-reward-percentage is that the opposite of "content discovery" is "content apathy". Let us not forget that those algorithmic auto-votes are largely currently voting on real content; they do so precisely because the expectation is that they will be later curated by humans. It is possible to achieve curation-rewards that are higher than the original vote in the current curve. Hence the issue described before, that the auto-voters are making a better return than the human curators whose work they rely upon! But take away that whole dynamic and what do we end up with?

Without turning this into a thesis, I think the door then becomes open to agnostic-voting and content-apathy, insofar as it makes no financial difference what one votes on. Like I said, this is good for humans who appreciate quality content but not so great for those accounts seeking to maximise their returns. Just as a tangent here, much talk about encouraging investors but this is what investors do - they make profits. Indeed, although they probably need no encouragement, those purely financial voting-rings would see proportional curation-rewards as an incentive to carry on, as the alternative of lucrative curation-chasing no longer exists. Such abuse can be dealt with using other methods, so I shan't distract the conversation by veering in that direction. However, agnostic-voting remains a potential issue.

Is there any solution to this? The question itself is baiting me to contradict my previous assertion that no single curation-curve can satisfy all parties. One effect I have not as yet spoken about is the role of huge accounts - yes, the so-called whales. The current reward-curve and curation-curve are heavily skewed in favour of whale-votes. I am not going to be drawn into whether by design or not, but Blurt has been lucky in that it has avoided this issue - albeit due to some small mathematical bugs - and hence the user experience has been manifestly different.

However, if we did revert back to replicating the HF21-experience, those huge votes have a large impact. Firstly, the reward-curve means a whale generates almost double the percentage of rshares of an average vote. Secondly, the curation-curve then yields a very large portion of that 50% curation rewards. If we add the two effects and compare it with what the linear curves would yield, a whale vote can easily make close to a self-vote without self-voting. I suggest this was the fundamental aim of the EIP: to remove funding of the bidbots by giving those accounts an almost-equivalent yield through the machinations of the various curves.

The important point here is not to sentence a corpse to death but to truly understand how it works and what the consequences are. The second-order effects here are that curation-chasers then attempt to front-run such whale votes; the aim of "content discovery" is nowhere to be seen and we are back again to maximising curation-rewards. I cannot stress this enough: the current curation-reward idea leads to an optimisation that ignores its declared original aim of rewarding content discovery.

Summing up

So how do we resolve this? I think we just need to accept that one single function cannot do multiple jobs and that we need to look at a multitude of simple functions that together lead to an experience that Blurters appreciate.

In this case, many have expressed their preference for linear curation rewards so that everybody gets a flat 50% of their vote - there is here a valid option of making curation rewards non-linear but in a more gentle way, and I may write about this more if people express an interest. In terms of content discovery, the development of the curation accounts and their communities will help this in the long term. With the disappearance of curation-maximisation, any attempted increase in abusive voting patterns such as voting-rings will be dealt with using a separate algorithm. We thus have three solutions - two algorithmic and one human - whose combined effects should make everyone content.

Feel free to discuss these issues.

---

Summary

= Changing the curation-curve to linear will lead to the disappearance of curation-maximising voters.
= However, it also leads to content-agnostic voting in that every vote yields the same curation reward independent of content quality.
= Content discovery therefore falls on the shoulders of human curators and community curation accounts.
= Rewards to such human curators will have to come from the larger community accounts and not from the curation-curve.
= Abuse of curation-rewards will be dealt with using the same new algorithm to be developed for general voting abuses.
= Finally, the current curation-curve is due to have a parameter change at the next HF so it can be constructed so as to have a gentler effect yet remain non-linear.