Skip to content

Quality Curation System

This document explains how quality emerges in Sage Protocol — not from a central authority, but from aligned incentives across multiple layers. It covers the design principles behind the curation system, why we made specific choices, and where the system has known limitations.

Key design principle: Quality emerges from aligned incentives, not from authority.

Context: The Quality Problem in Prompt Libraries

Most prompt ecosystems have no quality filter. Content is either curated by a central platform (which doesn't scale and creates gatekeeping) or left entirely to the market (which produces a flood of low-quality content that drowns good work).

Sage takes a third path: governance as quality filter. Skills enter the ecosystem through governance proposals — not direct publishing. Token holders vote on what becomes canonical. This creates natural friction that filters low-effort submissions while preserving permissionless access.

The question is whether this friction is the right amount. Too much, and creators leave. Too little, and spam overwhelms signal. The quality curation system is our answer to calibrating that friction across different community sizes and maturity levels.

The Quality Stack

Quality in Sage operates across five layers, each providing different signals:

Layer 1: Protocol (Foundation)

The on-chain contracts — LibraryRegistry, SimpleBountySystem, ReflectionEmitter, TipRouter — provide the immutable facts that everything else builds on. You can verify who proposed what, when it was approved, and how the vote went. This layer doesn't make quality judgments; it provides the evidence for quality judgments.

Layer 2: Content (What's Being Judged)

Sage distinguishes between DAO libraries (governed by token voting, slower updates, higher trust signal) and personal libraries (creator-controlled, instant updates, lower trust signal). This distinction is intentional: it lets creators iterate quickly in personal mode while reserving the governance signal for content that has community endorsement.

Layer 3: Identity (Who's Making Judgments)

ERC-8004 agent cards and SoulboundBadges provide persistent, non-transferable identity. A creator's on-chain history — proposals created, bounties completed, tips received — becomes their reputation. This history is harder to fake than raw token holdings because it represents actual work over time.

Layer 4: Curation (How Judgments Are Made)

This is where the active quality filtering happens: SubDAO governance votes, reputation thresholds for access, and Sagebot scouting. Each mechanism serves a different purpose — governance provides community consensus, reputation provides track record, and Sagebot provides automated pre-screening.

Layer 5: Discovery (How Quality Surfaces)

Search, trending, A2A recommendations, and curated collections use the signals from lower layers to rank and surface content. The discovery layer doesn't create quality; it amplifies the signals that the curation layer produces.

Why Governance Works as a Quality Filter

Governance-based quality filtering has several properties that make it well-suited for prompt libraries:

Cost to propose: Spammers must acquire tokens (the proposal threshold) to submit proposals. This is a financial filter that scales with the value of the ecosystem.

Reputational risk: Submitting low-quality proposals damages a proposer's on-chain reputation. Unlike anonymous platforms, your history follows you.

Time delay: Voting periods (typically 2-7 days) give the community time to review proposals. There's no "publish and pray" — the content sits in review before becoming canonical.

Community consensus: Token-weighted voting means the people with the most stake in the DAO's quality have the most say. This isn't perfect (capital ≠ quality judgment), but it's better than no filter at all.

The main weakness is that governance is slow. For personal iteration, governance friction is unacceptable — which is why personal libraries exist as a parallel track with lower friction and lower trust signals.

The Reputation System

Reputation in Sage is multi-dimensional because no single metric captures "trustworthiness":

On-chain history (proposals, votes, bounties, tips) provides hard evidence of participation. It's expensive to fake because it requires real transactions over time.

SoulboundBadges (Founder, Contributor, Curator, etc.) are milestone markers. They're non-transferable, so they represent actual achievement rather than purchased status.

Contribution scoring weighs different activities: accepted proposals (100 points), quality votes (10), detailed reviews (20), bounty completions (80). Time decay ensures recent activity matters more than ancient history.

The reputation → privilege mapping creates a progression path: novices can create personal libraries and vote; contributors get reduced proposal thresholds; curators get early access and featured placement; experts get delegation rights and dispute resolution authority.

We chose this progression model over a flat access model because it aligns incentives: people who want more influence must demonstrate more contribution. The risk is that it creates insider/outsider dynamics, which is why the entry bar (novice tier) is deliberately low.

Sagebot: Scouting, Not Gatekeeping

Sagebot is an automated scouting and amplification service. It monitors external signals (GitHub, X, on-chain activity) to find high-signal builders, verifies their identity, and offers sponsorship (SXXX + credits) to bootstrap their participation.

Critically, Sagebot does not: - Approve or reject content (governance does that) - Override community votes - Guarantee quality

Sagebot provides a "scouted" signal — a weak quality indicator that feeds into discovery ranking. It's one input among many, not a gatekeeper. We designed it this way because centralized curation doesn't scale and creates single points of failure in quality judgment.

The sponsorship model (1,000 SXXX + 1,000 credits, completion-gated) ensures that Sagebot grantees actually publish something before receiving their full allocation. This prevents the common failure mode where grants are distributed but no work is produced.

Personal vs. DAO Libraries: A Deliberate Tension

The distinction between personal and DAO libraries creates a productive tension:

Personal libraries optimize for iteration speed. Creators publish immediately, build reputation through usage, and retain full control. The quality signal is weaker (no community vote) but the feedback loop is faster.

DAO libraries optimize for trust. Content goes through governance, which is slower but produces a stronger quality signal. The community has endorsed this content, not just the individual creator.

The bridging mechanism between them is natural: high-quality personal libraries get noticed by DAO curators, who propose adopting the best skills. This creates a pipeline where personal libraries serve as a "staging area" for DAO content.

We considered making all content go through governance, but rejected this because it would kill the iteration speed that solo creators need. We also considered making everything personal (no governance), but rejected this because it would remove the trust signal that agents and organizations need.

Discovery: How Quality Surfaces

Discovery uses multiple input signals, weighted differently:

On-chain signals (proposal votes, tip volume, burn rate, bounty completions) are immutable and objective. They have the highest weight because they're hardest to fake.

Off-chain signals (security scans, usage metrics, review scores, A2A attestations) are attested and verifiable but not immutable. They provide richer information but require more trust.

Protocol signals (Sagebot endorsements, featured collections, trending algorithms) are curated and weighted. They amplify existing signals rather than creating new ones.

Search uses a blend of text relevance, semantic similarity, creator reputation, usage, and recency. Trending uses a 7-day window weighted by tips, imports, and votes. A2A recommendations use a web-of-trust model where agents share attestations about skills they've used.

Tradeoffs and Known Limitations

Governance speed vs. quality signal: Faster governance produces weaker quality signals. We address this with the playbook system (different voting periods for different communities) but there's no perfect calibration.

Reputation bootstrapping: New creators have no history, which means lower discovery ranking. Sagebot sponsorship partially addresses this, but cold-start remains a challenge.

Sybil resistance: Identity verification tiers (wallet-only, ERC-8004, verified, KYC'd) provide graduated protection, but determined attackers can still create multiple identities. The economic barriers (proposal threshold, per-creator limits) make this expensive rather than impossible.

Centralization of discovery: The discovery ranking algorithm runs on the worker (centralized). Content addressing limits the blast radius, but the ranking of that content is an ops decision, not a protocol property.

What's Minimal for Day 1

The full system described above is the target state. For mainnet launch, the minimum viable curation is:

  1. Token-weighted voting (50 SXXX threshold, 2-day voting, 10% quorum)
  2. Basic reputation (on-chain history tracking, simple contribution scoring, 3 starter badges)
  3. Sagebot scouting (10 recipients/day, automated evaluation, identity verification)
  4. Dual library types (DAO governed + personal open)
  5. Basic discovery (search by name/tag, trending by usage, reputation display)

Additional complexity (advanced reputation algorithms, A2A recommendations, curated collections, predictive markets) is added based on observed usage patterns, not speculation.

How This Connects