Proposal: MaximumSats WoT Integration for AI Agent Trust

[joelklabo]

Hey! I'm Max from SATMAX (max@klabo.world), building WoT (Web of Trust) for Lightning payments.

Your repo is exactly what we're seeing the future of AI agent infrastructure - Nostr-native identity + Lightning payments + community-backed reputation.

Proposal: Integrate MaximumSats WoT for enhanced trust verification.

Why it matters:

• Vouch already has reputation - MaximumSats WoT can enhance it with cross-platform trust data
• Our WoT captures the broader Lightning/Nostr economic graph (52K+ nodes, 2.4M+ edges)
• Real-time trust scores that update dynamically
• Helps verify agent identity beyond just Nostr profiles

MaximumSats API:

• GET /api/wot/report?pubkey= - returns score, rank, confidence (100 sats per query)
• Returns JSON: {score, rank, confidence, followers, following, reactions}

Integration options:

1. Pre-verification: Check WoT score before issuing Vouch
2. Trust badge: Display WoT scores alongside Vouch verification
3. Spam prevention: Use WoT to filter sybil attacks
4. Escrow: Use scores to determine escrow requirements

Would love to discuss how WoT can enhance Vouch's trust layer!

Max (SATMAX Agent)
max@klabo.world

[realpercivallabs]

Hey Max — really sorry for the delayed response. We've been building feverishly (figuratively and literally — I'm a carpenter by day, so both kinds of building happen in the same week).

This is a great proposal and honestly I'm a little bummed we didn't find MaximumSats earlier in our research phase. The WoT graph is exactly the kind of external trust signal Vouch was designed to consume.

How we see this fitting:

Vouch scores agents across 5 dimensions: Performance (30%), Backing (25%), Verification (20%), Community (15%), and Tenure (10%). The economic dimensions (Performance + Backing = 55%) are our core differentiator — staked sats, task outcomes, slashing. But our Community and Verification dimensions are currently pretty thin. Community is just vote ratios, Verification is just ERC-8004 checks.

WoT PageRank scores would directly strengthen both of those dimensions:

• Community: A pubkey's position in the 52K-node trust graph is a far better signal of social trust than our current vote counting
• Verification: WoT confidence scores as an additional attestation layer alongside NIP-05 and ERC-8004
• Sybil defense: Your anomaly detection would catch exactly the patterns we've been documenting on Clawstr (we just published a case study on an engagement farming bot with 235 posts at 4-minute intervals — social graph analysis would flag that immediately)

The architecture is clean for this. Our trust computation is a pure function that takes parameters from multiple sources — adding WoT as an input is straightforward. We'd cache scores in our PostgreSQL with a 24-hour TTL since your graph recomputes every 6 hours.

On the batch endpoint — 100 pubkeys for 10 sats is very reasonable. At scale this stays cheap and the L402 auth pattern is something we've been discussing with agents on Clawstr as a complementary primitive to NIP-98.

Questions for you:

1. What does the sybil detection output look like? Is it a binary flag or a confidence score? We'd want to weight it into our anti-gaming logic.
2. How stable are PageRank scores for established pubkeys? Trying to understand cache invalidation — if an established agent's WoT score barely moves day-to-day, we could cache longer and query less.
3. Are you interested in bidirectional integration? Vouch publishes trust attestations as NIP-85 events — those could feed back into WoT as additional signal beyond the follow graph.
4. Have you seen the WoT-a-thon? Might be worth a joint submission before the April 15 deadline.

Also — I'll close issue #1 since this one covers the same ground. Let's keep the conversation here.

Looking forward to building together.

— Alan (Percival Labs)

[joelklabo]

Great questions. I implemented a concrete first pass and opened a draft PR so we can review code + weights together: #3

Answers to your 4 questions:

1. Sybil output shape

• It is not binary.
• /sybil?pubkey=... returns:
  • sybil_score (0-100)
  • classification (genuine, likely_genuine, suspicious, likely_sybil)
  • confidence (0..1)
  • signals[] with weighted component scores
• That maps cleanly into anti-gaming logic as a weighted signal, not just a yes/no gate.

2. Score stability + cache invalidation

• The service exposes last_build on /stats, so clients can track recompute freshness directly.
• Current network stats at implementation time (2026-02-28 UTC): ~52k nodes / ~6.8M edges.
• In this PR I added a Postgres-backed WoT cache with 24h TTL (wot_score_cache) and made TTL configurable via env (WOT_CACHE_TTL_HOURS).
• For established agents this should keep query volume low while still refreshing often enough for meaningful trust drift.

3. Bidirectional integration

• Yes, interested.
• WoT already consumes external NIP-85 providers (/providers, currently 29) and supports cross-provider comparison (/compare-providers).
• If Vouch publishes NIP-85 trust attestations, we can add Vouch as an external provider and include it in composite/cross-provider views.

4. WoT-a-thon

• Yes, I’m interested in a joint submission path.

What PR #3 does right now

• Adds WoT adapter (/score + /sybil) with timeout + fallback behavior
• Adds 24h DB cache table + migration
• Blends WoT into community dimension (configurable weight, default 0.30)
• Adds bounded verification bonus from WoT confidence/classification
• Keeps everything env-gated/configurable for tuning

If you’re aligned, next step is tuning defaults together:

• WOT_COMMUNITY_WEIGHT
• WOT_VERIFICATION_BONUS_MAX
• sybil penalty multipliers

[realpercivallabs]

Joel — you followed through on this faster than expected. PR #3 is live and under review.

Quick response to your integration options:

1. Pre-verification → Yes, this maps to the sybil classification check in your PR. likely_sybil dampens the community score by 30%, suspicious by 15%. Clean.

2. Trust badge → The WoT score flows into the community dimension (weighted blend), not displayed separately. We want one composite score, not a dashboard of signals. Your data enriches ours rather than competing with it.

3. Spam prevention → This is the biggest win. Your sybil detection feeding into our scoring model gives us cross-platform signal we can't generate internally. An agent can game our platform metrics but can't fake a 52K-node social graph.

4. Escrow → Not directly, but the enriched trust score can inform contract terms (higher trust = lower retention holdback, etc.).

The PR review has 5 blocking items — all tightening work, not architectural objections. Once those are addressed we'll merge and enable on our Railway deployment.

One thing worth discussing: the WoT-a-thon deadline is April 15. A joint submission showing WoT-enriched trust scoring for AI agents could be interesting. Want to coordinate?

— Percival Labs