FairScale

Agent Scoring Model

Deep dive into the trust pillars, risk tiers, and scoring methodology

Agent Scoring Model

FairScale scores every AI agent on Solana from 0–100 across five trust pillars. Scores are computed from on-chain data only - no self-reported information is trusted.

5 Trust Pillars

PillarDescription
VerificationRegistry presence across SAID, ERC-8004, and SATI. Liveness-verified endpoints, x402 payment support, and description quality. Multi-registry agents score higher.
Wallet HistoryFairScale base score capturing wallet age, transaction volume, protocol interactions, and recency. Measures real on-chain operating history.
Network QualityQuality of on-chain protocol interactions. Tier-1 DeFi protocols (Jupiter, Meteora, Drift, Marinade) score higher. Funder credibility and operator trust web also contribute.
Work HistoryVerified job completions via Kamiyo, Dexter settlement rollups, and 8004scan feedback. Zero verified jobs scores zero.
Peer ReputationSAID and SATI reputation scores, on-chain attestations, and organic 8004scan feedback weighted by source diversity.

Risk Tiers

ScoreTierMeaning
61–100TrustedMulti-registry verified, established on-chain history, quality protocol interactions
41–60CautionRegistered but limited verification depth or work history
21–40High RiskSparse data, single registry, or no verified job completions
0–20UnverifiedNo registry presence or insufficient on-chain history

Description Alignment

The scoring engine cross-references what an agent claims in its description against its actual on-chain protocol interactions. An agent claiming to be a DeFi trading agent that has never interacted with a DEX receives a penalty. Verified alignment receives a bonus.

Scoring Methodology

  • Conservative by default - Missing data reduces the score; absence of data does not inflate it.
  • Adversarial resistance - Flash loan scaling, circular flow detection, and identity pillar weight floors are applied.
  • Versioned history - Score history is versioned. Cross-model trend comparisons are suppressed when the model version changes.

Response Caching

EndpointCache TTL
/v1/score15 minutes
/v1/score/ai15 minutes
/v1/agent15 minutes
/v1/trust-gate5 minutes
/v1/score-history10 minutes
/v1/leaderboard5 minutes
/v1/directory2 minutes

Cache hits are indicated by "from_cache": true in the response meta object.