A Retrospective on SCF’s Reputation-Based Governance Mechanism
Executive Summary
Since October 2023, the Stellar Community Fund (SCF) has used Neural Quorum Governance (NQG) to signal community sentiment in award allocation for the SCF Build Award. Co-designed by BlockScience and the Stellar Development Foundation (SDF), NQG was designed to provide an open governance system that highlights trusted and expert-driven interactions, while allowing newcomers to climb the ranks.
Over 23 rounds (starting at SCF #22, with #44 being the latest), SCF has leveraged NQG to distribute over $38M to over 500 submissions contributing to building Stellar’s ecosystem of financial services applications, protocols, tooling and infrastructure.
This retrospective traces the journey: the foundational work with BlockScience, the open-sourcing of the smart contracts on Stellar, the introduction of dynamic voting weights based on natural learning curves, and the innovations in delegation and measurement that strengthened the system over time.
New: Analytics Dashboard
NQG has matured from an early-stage implementation into a proven governance signal. As part of the retrospective, we invite readers to also check out our latest dashboard (currently in beta, requires sign-up to access) which shows relevant insights from the 13 latest rounds and offers an animated explanation of how Neural Quorum Governance works in SCF. Across 832 submissions and 256 awards analyzed, NQG’s community ranking aligned with delegate panel award decisions 89% of the time, reaching 100% in the four most recent rounds (#41-#44), signaling a path towards maturity. Along the way, the mechanism’s neuron architecture, delegation model, and measurement methodology have each been refined based on real-world data from every round.
What is Neural Quorum Governance?
NQG is a novel voting-power attribution and delegation mechanism. Described in detail in the SCF Handbook, NQG combines two subsystems: Neural Governance, which computes individualized voting power through layered voting neurons, and Quorum Delegation, a group delegation mechanism inspired by the Stellar Consensus Protocol.
Neural Governance
A modular, layered architecture for computing a member’s voting power using Voting Neurons. Each neuron represents a dimension of a user’s contribution or reputation and can reward expertise.
Quorum Delegation
Rather than delegating to a single individual, users select a number of ‘delegates’ they trust to be part of their Quorum (anonymous) for the decision they’d like to delegate. If enough of those delegates actively vote and reach agreement, the delegating member’s weighted vote (through Neural Governance) is cast automatically based on the consensus of the Quorum. No voting power is transferred directly to a delegate — the decision to vote Yes or No is decided by the delegator’s selected quorum. This mechanism provides flexibility of delegation without single-delegate concentration risk, inspired by the Stellar Consensus Protocol that powers the Stellar network itself.
SCF’s Implementation of NQG
The current implementation of Neural Governance in SCF uses 5 Neurons: Trust Graph, Trust Loss, Assigned Reputation, Prior Voting History and Retro Vote Quality. Several of the neurons have weight attribution methods that are designed to mimic natural reputation and learning curves — making it accessible for newcomers to gain voting power, while preventing indefinite accumulation by long-standing members. For current parameter definitions, see the Components & Parameters page.
As for the implementation of Quorum Delegation, SCF currently allows community members with a Pathfinder, Navigator or Pilot tier to delegate their voting power to a quorum of 7 to 15 delegates (elected on a quarterly basis), with a threshold of 66.7%. Consensus can be reached from a minimum of 7, but the system will always take the maximum number of delegates to reach consensus from.
The Stellar smart contracts used to power SCF Community Vote are open-sourced on GitHub. To balance voter privacy with transparency needed to understand the system, NQG in SCF is implemented as a hybrid on-chain/off-chain architecture, where all data is published in the SCF Handbook and can be verified on through its testnet contract (note, Stellar testnet resets quarterly).
Phase 1: Foundations (SCF #20–#29)
October 2023 — July 2024
NQG was first activated in SCF #20 and the first full voting report (requires log-in to view) was produced for SCF #22. During these early rounds, BlockScience produced detailed HackMD voting reports and worked alongside SDF’s engineering team to calibrate the implementation against the formal specification.
Building the Foundation
The initial rounds served as a critical calibration period. With a voter pool of 20–41 participants across SCF #22–#27, the team established baselines for how each neuron contributed to overall voting power, identified areas where the implementation could be improved, and began iterating on both neuron parameters and the delegation model. Key outcomes include the first evidence of NQG differentiating project rankings compared to simple one-person-one-vote, and the groundwork for the logistic weighting functions introduced in later rounds.
Open-Sourcing the Contracts
By SCF #27, the team validated that on-chain and off-chain calculations matched within rounding error: an important milestone that set the stage for the open-source release. SCF #28 marked the open-sourcing of the NQG Stellar smart contracts, documented in the SCF’s Implementation of NQG — Open-source! blog post. The contract source code was made publicly available, allowing anyone to inspect the governance logic, neuron aggregation, and vote tallying — or deploy their own instance using the included getting-started tutorial.
Check out the SCF NQG Github Repo!
Dynamic Voting Weights
One of the most significant advances during this period was the shift from static to dynamic voting weights. The Prior Voting History neuron moved from a linear bonus model to a logistic function where recent active participation is weighted more heavily, and the bonus saturates over time. This design philosophy (described in The Road Ahead) mirrors how trust and expertise work in real communities: it’s relatively easy to level up once you pass an initial hurdle, but the more you level up the harder each level is to achieve.
Early Participation Trends
Phase 2: Experimentation (SCF #30–#37)
August 2024 — July 2025
SCF v6.0 launched with SCF #30, introducing category delegated panels. With the foundation built, this phase focused on open experimentation to improve voting outcome and experience. During this period the voter pool stabilized between 30–50 participants, while the average number of submissions increased from 20 to 65.
Category-Specific Quorum Delegation
To improve the abstain rate and further engage delegates, SCF v6.0 introduced Delegate Panels for specific project categories of the SCF Build Award: Applications, Financial Protocols, Infrastructure & Services, and Developer Tooling. Each quarter, the top 6–10 Navigators and Pilots in Quorum Delegation became part of that category’s Delegate Panel. These ‘Panels’, consisting of subject matter experts, provided targeted input for submission criteria and guidelines, as well as provide extensive feedback to projects.
Rise of Trust Graph Neuron
By mid-Phase 2, the Trust Graph neuron was NQG’s dominant driver of voting power. As its “highly trusted” multiplier sat on a narrow cohort, there was no mechanism to react when a previously-trusted member’s standing slipped. SCF #35 fixed both issues.
The Highly Trusted Bonus (HTB) widened the cohort from the top 1% to the top 10% of trust-graph scorers, with each contributing a 15% bonus to the VP they direct outward. This broadened influence without letting any single member dominate. A Trust History safeguard was added alongside, penalizing sudden drops in trust score. The Trust Graph Neuron now bundles PageRank + HTB + Trust History safeguard into one final VP contribution. The Trust Graph Neuron remained the largest VP driver through SCF #38, until Discord role bonuses rebalanced the system in #39 (see Phase 3).
Phase 2 at a Glance
Phase 3: Scaling & Refinement (SCF #38–#40)
August 2025 — January 2026
SCF #38 through #40 saw the biggest shifts in NQG’s evolution. Voter participation surged, new neurons were introduced, and the team refined both the delegation model and how results are interpreted — culminating in the canonical quorum algorithm rewrite that has been running ever since.
A Participation Surge
SCF #38 saw the voter pool almost triple, from 44 to 104 voters, primary new pathfinders delegating their vote. SCF #39 continued the trend to 120 and has found balance around the 100 voter mark to date. This growth was a positive sign of deepening community engagement with governance.
Refining the Delegation Model
The growing participation introduced a very high abstain rate where it was difficult to find consensus on results with prior methods. In SCF #38, the issue was temporarily mitigated by a lower quorum threshold allowing more delegations to succeed, but the real issue in delegation quorum logic was identified and fixed in SCF #39, shipped alongside a more nuanced measurement approach. Rather than ranking projects purely by total NQG score, the new methodology separated two questions: (1) did enough voters weigh in to make the signal meaningful, and (2) of those who did, what was the sentiment? This two-step approach (detailed in the SCF #39 Round Recap) produced results that better reflected actual community sentiment.
The quorum multi-layer issue in SCF #39 surfaced after initial result generation and was resolved with a post-tally hot-fix that lifted resolution from 61.6% (#38) to 100%. SCF #40, documented in the SCF #40 Round Recap, formalized the redesigned quorum algorithm representing the culmination of learnings from SCF #39. Key solutions included simplifying the quorum logic which requires 66.7% approval rate for a delegated ‘Yes’ vote, requiring a minimum of 7 delegates per voter, favoring the maximum possible quorum (all available delegates), and enforcing that all delegates vote before a round closes. These changes, combined with operational enforcement, resolved the delegation abstain challenge that had been present since the earliest rounds. The system operated without requiring mid-round adjustments for the first time in several rounds.
Neuron Updates
SCF #39 introduced two changes to the neuron components: the introduction of Retro Vote Quality Neuron, and expansion of Verified Tier Neuron.
Introduction of Retro Vote Quality Neuron
Awards a bonus (or penalty) based on how well a voter’s previous votes predicted project success. Voters who supported projects that reached mainnet (or another tranche) within 6 months receive a bonus relevant to their tranche progression (0.1 to 0.3); those who supported projects that stalled receive a small penalty (-0.1 to -0.3). This creates a forward-looking quality signal grounded in actual outcomes.
Expansion of Verified Tier Neuron
Broadened recognition for community roles including Ambassadors, Public Goods Contributors, and other engaged members — expanding the reputation inputs beyond the core Pathfinder → Navigator → Pilot progression.
The Verified Tier bonuses not only expanded in recognition, but materially rebalanced the composition of voting power. Reputation’s share of total voting power jumped from ~20% pre-#39 to 75.3% in #39 itself and has hovered between 66% and 76% in every round since. The Trust Graph neuron, which had been the dominant contributor through #38, fell to 15% in #39 and as low as 3.2% in #41.
Despite this composition shift, voter-level inequality dropped. The Theil index fell from 0.923 (#38) to 0.319 (#39) — the most equitable distribution in NQG’s history at the time. The share held by the top 5 voters fell from 43.6% (#37) to 18.4% (#39). Flat per-voter role bonuses effectively leveled the per-voter distribution that the Trust Graph had concentrated.
Reason for change
Even after two years of running the mechanism, the voting results could not yet be used as determinative, and each round still required a dedicated discussion period with the delegates after the voting concluded to confirm the final awarded projects. This was especially noticeable in SCF #40, where the NQG-panel alignment (how well the community’s NQG ranking matches the delegate panel’s actual award decisions) was at its lowest.
The root cause wasn’t in the NQG mechanism itself, but in the overall award structure: the growing number of submissions to review (~70 each round) introduced voter apathy where less attention was spent on each submission. Along with the increasing amount of AI-generated submissions, and lack of signal from referrers, it became a lot harder to distinguish between grantfarmers and high-quality teams providing real ecosystem value.
Phase 3 at a Glance
Phase 4: Maturity (SCF #41–#44)
February 2026 — July 2026
SCF 7.0: Entering a New Era
SCF #41 marked the transition to SCF 7.0, restructuring the program into three tracks: Open Track (novel on-chain use cases), Integration Track (building on ecosystem tools), and RFP Track (targeted developer tooling). Community Vote now applied to Open Track only, with Integration- and RFP tracks evaluated by the delegate panel directly. The category-specific panels were subsequently also simplified to one delegate panel. The reduced submission count (33 in SCF #41) is a definitional consequence of this restructure rather than a contraction signal: only Open Track submissions feed into the community vote.
Most significantly, SCF #41 produced the first round of 100% NQG panel alignment in NQG’s history: every awarded project also cleared the >82%-approved-NQG-score bar. This kicked off what we now call the <82%-approved-NQG-score selection trend — the panel using NQG score as an explicit quality threshold for award eligibility.
SCF #42-#44: Steady-State Maturity Rounds
SCF #42–#44 operated as the SCF 7.0 steady state and extended the 100% NQG–panel alignment streak that began in #41 to four consecutive rounds. Across the three rounds, 70 Open Track submissions reached the community vote and 29 were awarded — every one of them matching NQG’s top-N ranking exactly and clearing the 82%-approved-NQG-score bar the panel adopted in #41.
Participation grew steadily: #42 recorded 106 voters (a record at the time), #43 held at 100, and #44 set a new all-time high of 128 voters against a community that grew from 585 to 659 verified members. Turnout was 18.1%, 16.3%, and 19.4% respectively, with voter retention from the immediately prior round holding above two-thirds throughout (81.9% into #42, 66.0% into #43, 79.0% into #44).
Delegation continued to be the dominant mode of participation. All three rounds resolved 100% of delegated vote-slots — extending the perfect-resolution streak to four consecutive rounds — and the delegate slot share climbed to an all-time high of 80% in #44 (up from 75% in #42 and 74% in #43). The score gap between the lowest-awarded and highest non-awarded projects also widened over the three rounds, reaching its widest measured value in #44 (+424 vs +364), giving the panel a clean and undisputed community shortlist to work from.
Neuron weight rebalancing
Over the course of 2026 rounds, we noticed relative weight rapidly accumulating in certain neurons over time — the adjustments made in SCF #39 had skewed the weights significantly towards the Reputation neuron. The Vote Quality neuron, introduced at 4.1% of total VP in #39, reached 24.0% of VP in #42 — roughly a 6× share increase in three rounds. As the initial weighting function was a simple sum, the more Stellar projects ship and more retroactive signal accumulates, the neuron’s weight has compounded. This made the impact of the Trust- and Voting History neurons nearly negligible.
Analysis to understand the ‘ideal ratio’ of the neuron weights for optimal performance on Community Vote is challenging given the number of variables that influence it. From a design perspective, the desire is to balance the Neuron contributions to a certain extent and prevent extremities, so each plays a role in one’s voting power.
In #43, the weighting function of the Vote Quality neuron was adjusted to a logistic function similar to the Vote History neuron, to prevent indefinite voting power accumulation over time. In #44, the Voting History and Trust Neuron range were re-scaled from 0.0–1.0 to 0.0–3.0. We also noticed issues with the Trust loss function inside of the Trust neuron and decided to move it into a separate neuron. The Trust Loss neuron is a safeguard triggered when a significant number of users and voting power remove someone from their trusted list of users.
Phase 4 at a Glance
The Million-Dollar Question: How Well Does NQG Actually Predict Award Decisions?
Because SCF’s Community Vote is non-determinative (the Category Delegates panel makes the final award allocation) we can directly compare NQG’s community signal against actual outcomes. Across 15 rounds with NQG data (SCF #30–#44), the results are compelling.
Methodology
For each round, we rank all submissions by NQG score and compare the top n (where n = the number of projects actually awarded) against the panel’s decisions. Historically, SCF has used a threshold-based interpretation: ≥75% of total NQG voting weight directed at a submission signals community approval, 50–75% signals ‘to be discussed,’ and below 50% signals a no. This was adjusted in SCF #41 with the restructured program changes of SCF v7.0, which introduced an explicit ≥82%-approved-NQG-score bar for award eligibility. The rank-based comparison below captures the overall alignment regardless of threshold.
Results
When Community Signal and Delegate Panel Diverge
In SCF #40, several projects were awarded despite lower NQG rankings because panel reviewers recognized specialized ecosystem value. For example, one privacy infrastructure project was described by reviewers as a “strong use case with clear potential impact” by a team that had “already built similar infrastructure elsewhere.” Another, a security service provider, was noted for offering “free risk APIs for any wallet in the ecosystem” — a value proposition that resonated more with expert reviewers familiar with ecosystem gaps than with the broader voter base.
Conversely, projects that scored well in the community vote but weren’t awarded typically received panel feedback about execution readiness. One highly-ranked project was told its enterprise partnerships were unclear and its funding rationale needed clarification alongside existing venture backing. Another was encouraged to “share more about how the value capture can occur” on Stellar specifically. These are exactly the types of assessment where expert panel review complements the broader community signal.
The four latest consecutive rounds of perfect alignment in SCF #41 to #44 — the first in NQG’s history — suggest that the combination of the canonical quorum algorithm, the SCF 7.0 structure changes, the explicit ≥82%-approved-NQG-score eligibility bar, a consistent ‘sweet-spot’ voting power concentration, and the matured neuron architecture is producing a reliable and actionable community signal.
Key Insights
1. Open-entry, reputation-based governance is possible at scale.
Open-entry governance has historically defaulted to token-weighted mechanics: one dollar, one vote. That works for a narrow set of outcomes but reliably suffers plutocratic capture and rewards passive holding over engaged contribution. Traditional reputation-based governance sidesteps that by trusting people rather than capital, but the trust has to be either informal (capped by Dunbar’s ~150-relationship limit) or explicitly whitelisted (a management burden that doesn’t scale across the internet). NQG threads that by encoding reputation formally on-chain (through PageRank on the peer trust graph, tier assignment, historical participation, and retroactive vote quality) while keeping the entry gate open to anyone who joins the community and gets verified. Combining this with flexible, quorum-based delegation, it creates system that’s scalable, led by an emergent steering core which makes decisions representative of the wider community.
Across the 15 rounds analyzed, about 22 members vote in nearly every round and hold roughly 3x the mean voting power of members who appear in quorums but rarely vote directly (about 94 of them), and 5x that of newcomers in the most recent round. The result is closer to an earned board ratified by delegated members than to a flat democracy.
2. The right onramp, gain and decay functions of voting power are critical for systems like these to work.
NQG’s evolution folds in basic verification criteria as an entrypoint into the system, after which the user can grow their influence by consistently voting (well) and gaining trust with other community members. Logistic curves reward consistency and prediction accuracy up to a bounded ceiling so accumulated influence saturates rather than compound indefinitely, making day-one voters competitive immediately.
The same curves decay when a voter stops showing up: the Voting History weighs recent activity more heavily than older rounds, so missed rounds gradually erode the accumulated multiplier. Reputation has to be maintained, not just earned. Where the logistic decay handles gradual disengagement, the Trust Loss safeguard neuron is fast-acting: it applies an immediate penalty when a voter’s peer-trust score drops sharply between rounds. This way, accumulated influence should not survive an abrupt loss of community standing.
3. Non-transferable voting power is a security feature and a structural cost.
Reputation-based governance is naturally resistant to a class of attacks that plague token-weighted systems: vote-buying markets, whale accumulation via secondary markets, and Sybil-scaled purchases. Because VP cannot be sold, an attacker cannot simply buy influence. They have to earn it through the same slow reputation loop as everyone else, which the logistic curves deliberately meter.
The same non-transferability, however, means disengaged contributors do not exit — they just stop showing up. The Drop-offs cohort (n=11) left between rounds #40 and #41, exactly when their accumulated VP had quadrupled. In a token-weighted system, that VP would have carried a resale value and given them a clean exit; in NQG it simply lapses back into the pool.
4. NQG is inherently complex and requires continuous evaluation.
The main downside with more complex systems such as NQG over a simpler system such as one-token/person-one-vote is the difficulty to fully understand how voting weight is accumulated for its members, as well as evaluate what parameters and what voting weight functions influence the desired outcome.
The non-determinative nature of the community vote adds another layer: because the delegate panel makes the final call, understanding a round requires reading both the NQG signal and the panel’s rationale. The latest four rounds with the SCF 7.0 restructure has proven the voting outcomes aligns with the panel indicating the direction is right, but the ‘ideal’ voting weight recipe for the highest impact remains somewhat speculative, given the number of variables that influence it.
The need for large data processing does pose a natural fit for AI-assisted explanation and audit. Round recaps that used to take days now assemble from the data in minutes, and interactive tools like the Analytics Dashboard let anyone drill from a headline number into voter-level attribution.
Looking Ahead
The next phase is about extending it to new contexts, making it verifiable without exposing voters, and moving the framework toward community control.
In development
- Improved UX for Trust Neuron: The Trust Graph mechanism works, but maintaining a personal trust list gets harder as the verified SCF community grows past 600 members. Goals are to create a more discoverable trust-list interface, batch attestation (so vouching for several people isn’t several separate actions), and add visible feedback on what an attestation does, so members can see how their trust translates into voting weight for the people they vouched for. With an improved UX, it makes sense to increase the relative weight closer to earlier levels.
- Improved Assigned Reputation Neuron flow: As you can see in the dashboard, the amount of voters with the Navigator role is small compared to the other two, which is likely caused by a friction between Pathfinder to Navigator throughflow, and a limited amount of benefits/responsibilities connected to the Navigator role. Improving both will likely result in a more balanced Assigned Reputation Neuron.
- NQG in other funding contexts to stress the design in other conditions, with two implementations are underway. The Stellar Public Goods Award, hosted on Tansu, adapts the NQG design for ongoing public-goods funding, and the SCF-funded Nouns Builder project is exploring how quorum delegation complements DAO-native governance. There’s also an open invitation here: teams working on reputation-based governance, trust graphs, attestations and delegation are invited to collaborate — comparing notes would speed up design work on all sides.
Next up to explore
- Privacy-preserving verification. NQG runs as a hybrid architecture: the Stellar contracts are open-source and testnet verifiable, but neuron inputs run off-chain to protect individual voter privacy. The goal is to close that using zero-knowledge proofs, verifying neuron calculations end to end without exposing voter identity.
- Simulation tooling for parameter changes. BlockScience’s cadCAD model gave us a working simulation environment for the first few rounds, but required significant effort to operate. Round-to-round iteration needs something more accessible: propose a parameter change, replay it against the last n rounds of real data, and see how the tally, distribution and alignment would have shifted. With increasingly capable AI tooling it becomes more feasible to give not just the maintainers, but also the broader community a shared artifact to point at when asking why a parameter was tuned.
- New and improved neurons, including: Technical expertise, drawing on verifiable GitHub contributions to Stellar (ecosystem) repositories and code review activity. Longer-window vote quality, tracking impact, not just in-time delivery — whether a supported project retained users, generated volume, or hit measurable on-chain milestones after being awarded. It’s the most interesting but also the hardest.
- Community-governed parameters. NQG today is designed, iterated and operated by SDF on behalf of the SCF community. The intent is for the framework itself to be governed by that community, with SDF in a stewardship role rather than an operator role.
Congratulations for making it to the end of this near 5K word retrospective! If you’re hungry for an even more in-depth review on this report, check out this follow-up article in my blog Quorum & Consequence.
Resources & References
Core Documentation
- SCF Handbook: NQG Overview
- SCF Handbook: Components & Parameters
- SCF Handbook: History of SCF
BlockScience Publications
- Introducing Neural Quorum Governance
- The Story Behind Neural Quorum Governance
Implementation & Code
- GitHub: stellar/stellar-community-fund-contracts
- GitHub: BlockScience/neural-quorum-governance
- Contract on Testnet (run for SCF #44, but will be temporarily as Stellar testnet resets every quarter)
Medium Articles
- The Road Ahead: SCF’s Implementation of NQG
- SCF’s Implementation of NQG — Open-source!
- SCF #39 Round Recap
- SCF #40 Round Recap
- SCF #41 Round Recap
- SCF #42 Round Recap
- SCF #43 Round Recap
- SCF #44 Round Recap
Voting Reports
Reports for SCF #22–#27 as HackMD articles by BlockScience, reports from SCF #28–38 are listed as spreadsheets, and reports for SCF #39 — #44 are listed inside the entire round recaps on the Community Medium, all listed on the SCF Handbook.
Contributions and review provided by Bartłomiej Chaliński from Software Mansion, as well as SCF Verified Members. Special shout-out to Matias, whom specifically requested more visualization, resulting in a visual for every ~180 words.
About Stellar Community Fund
The Stellar Community Fund (SCF) is an open-application awards program managed by the Stellar Development Foundation to support developers and startups building on Stellar and Soroban. SCF relies on input from verified members in the Stellar community for award allocation, governance and project support to keep the fund running and developing. Review rounds run monthly and projects can request up to $100K worth of XLM at a time.
About Stellar Development Foundation
The Stellar Development Foundation (SDF) is a non-profit organization that supports the development and growth of Stellar, an open-source network that connects the world’s financial infrastructure. Founded in 2014, the Foundation helps maintain Stellar’s codebase, supports the technical and business communities building on the network, and serves as a voice to regulators and institutions. The Foundation seeks to create equitable access to the global financial system, using the Stellar network to unlock the world’s economic potential through blockchain technology. For more information, visit stellar.org/foundation.
About BlockScience
BlockScience® is a complex systems engineering, R&D, and analytics firm. Our goal is to combine academic-grade research with advanced mathematical and computational engineering to design safe and resilient socio-technical systems. We provide engineering, design, and analytics services to a wide range of clients, including for-profit, non-profit, academic, and government organizations, and contribute to open-source research and software development.
2+ Years of Neural Quorum Governance was originally published in Stellar Community on Medium, where people are continuing the conversation by highlighting and responding to this story.