# SCF #23 NQG Voting Report

*BlockScience, 27 February 2024*

## Introduction

In 2023, the Stellar Community Fund (SCF) and BlockScience (BSci) [collaborated](https://blog.block.science/introducing-neural-quorum-governance/) on [ideating](https://blog.block.science/the-story-behind-neural-quorum-governance/) and implementing a novel governance mechanism, titled Neural Quorum Governance. Following this initial phase, we are now monitoring and evaluating this mechanism through per-round reports. This allows the community to better inform themselves about the dynamics and effects of this voting mechanism, informing discussions on changes and adaptations.

## Report Summary

On this report, we showcase key statistics along with a breakdown over the votes and the computed power for each project submission. This report shows, that for current parameter selections, the NQG results are mostly (but not perfectly) commensurate to a proxy scenario in which each vote has power +1 or -1 (i.e. one voter one vote).

We also discuss Quorum Delegation, which seems to most affect middle- to lower-tier projects when ranked by `Yes`/`No` votes.

This [spreadsheet](https://docs.google.com/spreadsheets/d/1aM13J8a9qpinzO_wFtXmVio1UisJEMtOQauQBSV12Sk/edit?usp=sharing) allows you to check the delegatee votes and the quorum delegation result. It allows sorting for any `delegating user: submission` combination.

Finally, BlockScience has reviewed the smart contract implementation, which, combined with the backtesting simulation results, led us to recommend several action items for SCF #24.

## Results

### Summary Statistics

During SCF #23, 18 project submissions received 339 direct votes, while 84 were delegated. The direct votes can be split into 258 (76.1%) `Yes` and 81 (23.9%) `No` votes.

A total of 32 unique voters participated. Out of these, 11 voters chose to delegate their vote for at least one project, whose union of quorums comprised 35 distinct voters.

As for the delegated votes, 25% were resolved successfully, with 22.6% `Yes` votes and 2.4% `No` votes, while the remaining 75% were resolved as `Abstain`. In other words, 75% of delegated votes wound up in Delegate Quorums, where no successful agreement was found.

### Votes per Project

In this section, we present the vote and voting power breakdown across project submissions. Project Power seems primarily commensurate with the voting pattern in terms of `Yes` vs. `No,` although the correlation is incomplete. Delegation seems to be more present on middle—and lower-tier projects (in terms of `Yes` vs `No` votes).

---

*Fig [VotesPerSubmissions]: Vote Types for each submission. The number of delegations was relatively uniform, although the higher tier of projects received fewer delegation actions than the middle tier of projects.*

---

*Fig [PowerPerSubmissions]: Assigned vote power per project. Projects with negative voting power are those with more `No` than `Yes` voting power was received.*

---

*Breakdown of vote types across project submissions*

---

### Voting Power Outcomes

This section explores some of the results associated with the NQG-computed power itself.

In table [NQGResults], we can inspect the Voting Power per wallet address (the Power column) and the Effective Power, which is defined as:

$$\text{VoterEffectivePower} := \text{VoterPower} + \frac{\sum \text{DelegatedPower}}{\text{VoterQuorumSize}}$$

We can observe that amongst the voters who did receive delegations, the median potential "bonus" for being a delegate was 250%.

*Table [NQGResults]: [NQG results across wallets](https://docs.google.com/spreadsheets/d/1SCMoa7GwRzzQ3E1Ic4bEleIq0LDoToOi7mnu6yBW4WQ/edit?usp=sharing)* as measured by Power and Effective Power (which also takes into consideration the average power for which the user delegated for). We can observe that **amongst delegates, the median Potential Delegation Bonus was around 250%**.

---

Another analysis that we make is to compare the power results against a simulated 1-person-1-vote scenario. We can observe that the project order is mainly preserved, albeit with some outliers, such as `EasyA` and `Soroban Explorer II.`

*Table [ResultVs1P1V]: Comparison of the project submission rankings between the results and a `what-if` scenario in which each vote would have been assigned as either +1 or -1 (rather than assigning individual voting power). The results are mostly commensurate, although a few projects had distinct numerical results.*


### Delegation Outcomes

A total of 84 delegations were made for quorums to decide on a project vote. Of those, 63 (75.0%) were resolved as `Abstain` votes, 19 (22.6%) were determined as `Yes` and 2 (2.4%) were determined as `No.` The Delegatee Votes and the Quorum Delegation result can be checked with this [spreadsheet](https://docs.google.com/spreadsheets/d/1aM13J8a9qpinzO_wFtXmVio1UisJEMtOQauQBSV12Sk/edit?usp=sharing), which allows sorting for any `delegating user` and `submission` combination. We choose to be transparent on the delegates set so that the community can best evaluate the results as we're in the earliest stages of the NQG implementation.

In terms of voting outcomes, Quorum Delegation was able to reshuffle a substantial portion of the submission rankings to +/-1 place in the ranking. The highest and lowest-scoring project submissions did not show a change in their ranking. This can be a preliminary indication that QD may be essential for submissions that are near the median and intermediary quantiles over the power distribution.

---

*Table [NQGvsNG]: Comparison between the submission rankings for two scenarios: one in which NQG is activated as usual and another one in which all Delegations are mapped to Abstain.*


---

*Fig: Screenshot for SCF #23 Quorum Delegation Results [spreadsheet](https://docs.google.com/spreadsheets/d/1aM13J8a9qpinzO_wFtXmVio1UisJEMtOQauQBSV12Sk/edit?usp=sharing), on which delegating users can be found by their wallet address.*

### Neural Quorum Governance Outcomes

As a first analysis for gaining an insight into how Neural Governance is working, we've built a [spreadsheet](https://docs.google.com/spreadsheets/d/1KHohnttHmZLiFcEqnCx3iA1E_EJOHuLjiiXmfh7K8mU/edit?usp=sharing) that simulates the layering logic in terms of Neurons and Weights. We've assigned the Neuron Values as being the ones originating from the results data, and it is observed that the numbers from the data are in line with what would be expected from those computations within a magnitude of 3 significant digits. The error in the following digits may be related to rounding and truncation errors.

*Fig [NeuronCalc]: A [spreadsheet](https://docs.google.com/spreadsheets/d/1KHohnttHmZLiFcEqnCx3iA1E_EJOHuLjiiXmfh7K8mU/edit?usp=sharing) for manually performing the NQG layer calculations for each voter and comparing them with the numbers from the data. The Neuron Values were retrieved from the results. We can observe that the results are similar within three digits.*

## Upcoming Improvements

During SCF#23, BlockScience completed a comprehensive review of the implemented smart contract code while executing detailed backtesting simulations. As a consequence of those two processes, we recommend the following actions that have the potential to enhance further the NQG performance in reflecting community sentiment. These include:

1. Re-calibrate the Prior Voting History Neuron
- Currently, the implemented PVH Neuron utilizes fixed parametrizations, which have a negligible effect on results. We recommend adopting a dynamical formalism, which introduces a learning curve for new members to start accruing an accelerating bonus until it reaches a saturation point, at which point accrual slows again.
2. Creating an Opt-out for Quorum Delegation
- Right now, users cannot opt-out from being delegates, and therefore their discord IDs can be associated with their decisions between Voting, Abstaining and Not Voting. This could potentially be unacceptable for community members, and therefore implementing a opt-out is a top priority feature for the upcoming rounds.
4. Evaluate making Quorum Delegation more aggressive
- Although QD is working as intended, the high rate of abstains may suggest that we should consider forms of further streamlining Quorum Consensus. One suggestion is to reduce the minimum quorum size - either globally or by giving the choice to the user. Another one is to consider tuning the minimum thresholds. Both could show the effects of more agreement being found within Quorums.
5. Further refinements to the Backtesting simulations
- Our backtesting simulations are currently limited to comparing the tallying vote results. We're continuously improving the cadCAD model capabilities to compare the NQG results more granularly. Scoped improvements include:
1) the ability to compare raw neuron outputs directly,
2) the ability to cross-check quorum delegation directly and
3) being able to visualize the relative impacts of each neuron.
6. Re-calibrate the Trust Bonus Neuron
The Trust Bonus is currently having an undersized effect on the outcome. A re-scaling factor could be introduced to make trust more impactful in the upcoming rounds.
7. Improvements to the NQG layer structure & execution
- Mostly technical improvements for future-proofing the system

We expect the impact on voting power to be substantial, although the relative ranking will still be primarily commensurate. Fig [InterventionCounterfactual] demonstrates the results on the submission ranking when comparing the actual results against a cadCAD simulation over the results that preemptively test those improvements.

*Fig [InterventionCounterfactual]: Comparison between the round results (result) and the expected outcome once the neurons are further calibrated and improved (NQG interv.). We can observe that the results are primarily commensurate, although some outliers are noted*