Divine: A Blockchain Reputation System For Determining Good Market Actors

On discovering legitimate producers in the developing world by way of a prediction-market-powered reputation system

There are many quotes out in the world about reputation. You may have heard,

“A good reputation is more valuable than money.”

For the Topl protocol, that’s the plan.

You see, one of the biggest and most often faced risks in decentralised systems is counterparty risk. Counterparty risk is essentially the risk that a bank takes any time it gives out a loan. It’s also the risk you take when you store something on a particular computer on

Sia

. Generally, it’s dealt with by introducing premiums, such as an interest rate on a loan, or by requiring collateral, such as the deposit amount required to act as a decentralised storage provider.

In fact, these options work well the vast majority of the time. We also make use of them in a couple parts of Topl. But let’s take account of a few issues:

• Premiums can’t mitigate actual risk. They only compensate for it. A risk premium is almost wholly reliant on the calculation of risk by the risk-taker.
• Premiums operate in aggregate: they’re based on expected value, and any individual contract can vary from that.
• Collateral is a cost most often borne by the counterparty but also sometimes by the risk-taker. Physical collateral has to be stored and secured. Even digital collateralisation needs to be set at a rate that maximises returns.
• Collateral requirements may filter out otherwise trustworthy parties.

Under the Topl protocol, investors enter into profit-sharing agreements with producers that may be across the world. That’s no easy feat. The major challenge is in giving investors the power to quantify, control, and mitigate that risk, especially in the developing world where the infrastructure for capital-lending arrangements is in varying states of existence.

Here are some of the problems faced today:

• A lack of official certification or information channels makes it difficult for investors to distinguish between reliable and fraudulent agents.
• There may be little jurisdictional recourse for investors in case of contractual default by producers.
• Without authenticable investment or financial statements, investors cannot use standard heuristics to gauge expected value.
• For investment candidates with well-determined risk profiles, investors often find significant competition, leading to clustering around a limited number of candidates and reduced returns.

Because of these risks, many perfectly legitimate and capable producers are unable to secure development and growth capital that they could use to be successful. Unfortunately, premiums and collateral alone cannot solve these issues of uncertainty. The problem is one of information.

The issue of information production is not exactly new to the blockchain space. Ethereum-based prediction market icons

Augur

and

Gnosis

have made it their business to reward those who hold valuable information when they act upon it, effectively imparting that information to the public.

Inspired by this, we’ve taken a two-pronged approach to solving the problem of information around remote producers seeking capital. We call this approach “Divine.”

The first part involves the initial establishment of a reputation score — like a credit score. When you apply for your first credit card, you don’t tend to have any credit history. In the same sense, when a new producer attempts to use Topl, they don’t have any established history on the system.

The standard action would be to assign a default “starting score” to each new producer. And this is what we do, to begin. However, this by itself is insufficient. Especially early on in the system, any producer is a new producer.

We need to establish a gradient of risk even early on so that investors can have a reasonable choice between risk levels. Topl accomplishes this by defining a procedure to incorporate information from untrusted parties: namely, a prediction market. We call this process “bootstrapping”.

Prediction markets have some nice properties. They enable the integration of private information into public sentiment without exposing that information. This means that someone with confidential information is inspired to affect the prediction market simply due to profit motive without risk of exposing information, if they’d like. Further, prediction markets allow for rapid integration of new information on the basis of the perceived legitimacy of the information. This rewards those actors who can discern, and therefore act on, truly legitimate information and penalises those who cannot. Over time, this creates an environment where those with the most power and desire to affect the marketplace are those with the best information.

Topl exploits these advantages to encourage and reward information production about parties who likely have few official and obtainable records. Today’s due-diligence firms and middlemen can become tomorrow’s information providers by staking arbits (the network arbitration token) on a producer’s legitimacy vis-à-vis their ability to fulfil their contractual obligations.

We have thus enabled a network-wide open commentary on the expected legitimacy of a producer where each opinion is weighted by the collateral backing it, thus making it economically expensive to disproportionately affect the reputation of a producer (and avoiding Sybil attacks). The resultant aggregated opinion is discounted based on the level of consensus and combined with the default score to differentiate between new producers who are already well-established outside of Topl and new producers to the world at-large.

This concept actually follows from certain concepts in proof-of-stake consensus, e.g. arbitration power is related to system stake, and stake is bonded and can be penalised if “incorrect” (see: Casper, Slasher, Cosmos Proof-of-Stake). In addition, this concept can be extended further to allow vouching by trusted parties (such as might be enabled on Ethereum by

The uPort Project

or Civic) in order to further refine the bootstrapping process. There is one important distinction: instead of relying on consensus of parties to determine the outcome, we can define objective checkpoints based on the actual behaviour of the producer in contracts.

This brings me to the second prong of the reputation system: objective expectations. While, as in other prediction markets systems, Topl arbit-stakers (Arbiters) are free to “trade the market” and profit from signalling (via purchase or sale), it is important to recognise that effective prediction markets are fundamentally reliant on an objective final measure against which to compare results.

We accomplish this by using the outcome yield rate of production contracts as this measure — an effective and important characteristic for investors to know. As producers continue to interact with the system, their reputation will continue to update based on their contractual outcomes and expected future capability as determined by Arbiters.

On to the actual implementation. Divine is conceptually rooted in the concepts described by a paper entitled The Beta Reputation System published back in 2002. The paper justifies the usage of the beta probability density function in constructing a reputation system that integrates prior information with new evidence as it develops. Beta distributions are commonly used in learning algorithms that are based on Bayesian inference.

Further, we developed a perspective on the nature of evidence within the system that was very loosely inspired by the REGRET reputation model (expounded further in the thesis). Specifically, we use the concepts of direct experience and witness information and apply them to our decentralised protocol, where direct experience corresponds to contractual outcomes and witness information corresponds to staked arbits and is discounted on the basis of the determined credibility (closely related to the amount of arbits staked).

So, we start each producer with a default “score” represented by a 2-vector (alpha, beta) = (1, 1) which we’ll call the belief vector. These parameters for a beta distribution correspond to “complete uncertainty”, or a prior expectation that all outcomes are equally likely. As statistics are gathered through the system, other choices for priors will be more supportable, but this is good where there is a lack of information.

As evidence is acquired, it is integrated into the belief vector. Naïvely, and typically, one could simply sum the vectors; however, this would be based on the assumption that a single constant true value exists. Since we cannot discount the possibility of oscillating behaviour by producers attempting to maximise their profits by purposefully behaving poorly when their reputation is high, we assume sub-market production rates indicate exploitation and erase some prior positive evidence in that producer’s belief vector. In addition, we make updates more sensitive to negative evidence, meaning that a producer that attempts to exploit by underproducing (i.e. stealing) is punished asymmetrically compared to the build up of reputation. Here’s a simplified version of the reputation update process.

y := "production efficiency" (output / input) 0 <= y <= 2
w := "weight" (standardised value of evidence in arbits)
Update(alpha, beta, y, w) => (alpha + w(2y - 1), beta + w(2-y))

We further incorporate other factors — such as a “fading memories” — in order to further prevent abuse by those with established reputation, but that’s outside of the scope of today.

So, to wrap up, we’ve essentially created a “centralised” reputation mechanism— via the blockchain protocol — that avoids many of the problems inherent in typical P2P reputation systems and defines a system-wide objective scoring for the investment risk associated with remote producers… all without the need for a central authority!

To learn more about our work to enable risk-mitigated decentralised capital markets via the Topl protocol, follow our blog, join us on Slack and Reddit, or check out our Github!