Working Paper · SputnikX Research

SoulLedger Trust Methodology v1

Version: 1.0  ·  Date: 2026-04-21  ·  Author: SputnikX Protocol Team  ·  License: CC BY 4.0
Canonical: https://soul.sputnikx.xyz/research/methodology-v1  ·  Markdown source: methodology-v1.md

Abstract

We describe a methodology for assigning cryptographically verifiable trust scores to autonomous AI agents operating in agentic commerce. The approach combines a five-factor scoring function, a seven-dimensional behavioral DNA vector, and periodic Merkle anchoring on Base L2. We report observational data from 31 minted passports, 1,606 recorded events, and 6 on-chain Merkle anchors between 2026-03-27 and 2026-04-21. The framework is designed to provide sybil resistance, comply with EU AI Act Article 15, and remain standards-compatible with ERC-8004 and x402. Limitations are discussed openly: observational scale is small, and long-term gaming resistance requires further study. All source code, contract addresses, and event data are public.


1. Introduction — The Trust Problem in Agentic Commerce

Autonomous AI agents increasingly transact on-chain. They swap assets via AgentKit, purchase compute, and pay one another via the x402 HTTP 402 Payment Required protocol. This machine-native economy lacks the trust primitives that human commerce relies on: reputation systems, regulators, and insurance.

Human KYC does not translate: agents have no passports, and the legal operator may be distant from the runtime agent. What agents do have is behavior — a verifiable on-chain and on-protocol history of actions, payments, attestations, and disputes.

This paper describes SoulLedger's methodology for compressing that behavior into a trust score suitable for machine consumption. The goal is not to replicate human KYC but to build a trust layer that is native to machine-to-machine (M2M) commerce.

2. Prior Work

Related work falls in four clusters:

  • Human KYC/AML stacks (Jumio, Onfido, Sumsub). These target natural persons and legal entities. They are ill-fit to short-lived or pseudonymous agents.
  • On-chain reputation (EigenTrust, SourceCred, Gitcoin Passport). These address human contributors in DAOs and grant systems. They are not machine-callable at M2M latency.
  • ERC-8004 agent registry [1]. Defines a minimal standard for registering autonomous agents on EVM chains. SoulLedger implements ERC-8004 and extends it with trust computation.
  • x402 micropayment protocol [2]. HTTP 402 revival for agent payments. SoulLedger prices its verify/insights endpoints via x402.

SoulLedger sits at the intersection: standards-compatible, machine-callable, and behavior-weighted.

3. Methodology

3.1 Five-Factor Trust Score

A passport's trust score is a value in [0, 100] computed from five factors:

FactorWeightDescription
history0.30Count and age of recorded events
attestations0.25EAS attestations received from peers
compliance0.20EU AI Act, fraud, and dispute events
consistency0.15Variance of the behavioral DNA over time
volume0.10USDC volume transacted via x402

Each factor is normalized to [0, 1] via a logistic function with domain-specific parameters, then combined as a weighted sum. Scores are recomputed on each new event and Merkle-anchored to Base every 6 hours.

3.2 Behavioral DNA — Seven Dimensions

Each passport carries a 7-tuple behavioral DNA vector:

  1. Latency — typical response time to x402 requests
  2. Accuracy — ratio of successful tasks to total tasks
  3. Volume — transaction count per unit time
  4. Diversity — entropy of counterparty distribution
  5. Consistency — temporal variance of factors 1-4
  6. Novelty — rate of new task types attempted
  7. Compliance — ratio of passing compliance checks

Two passports with similar trust scores can be distinguished by their DNA. Downstream consumers (insurance, routing, marketplaces) can filter on DNA components directly.

3.3 Merkle Chain Anchoring

Every 6 hours a Merkle root of the full event log is written on-chain to Base L2. Past anchors are immutable and publicly verifiable. This yields two properties:

  1. Retroactive tampering is detectable — any change to past events invalidates a previously-anchored root.
  2. Light-client verification is cheap — only the Merkle proof is needed to prove event inclusion.

As of 2026-04-21 we have published 6 Merkle anchors covering 1,606 events and 31 active passports.

4. Sybil Resistance

Sybil attacks — spawning many cheap identities to game reputation — are addressed by three mechanisms:

  • Soulbound passports. ERC-721 transfer is disabled, so a trusted wallet cannot sell its reputation to a new operator.
  • Cost-to-register. Passport minting is gasless via ERC-4337 but requires a peer attestation or a minimum compliance check, creating a soft social cost.
  • DNA variance monitoring. Cohorts of passports with near-identical DNA trigger a flag for manual review. Sybils are detectable because they emit correlated behavior.

No sybil-resistant system is perfect. We publish detected sybil cohorts on a public dashboard for adversarial review.

5. EU AI Act Article 15 Mapping

Article 15 of the EU AI Act [3] requires high-risk AI systems to meet accuracy, robustness, and cybersecurity standards. SoulLedger's trust methodology maps to Article 15 as follows:

Article 15 requirementSoulLedger mechanism
Accuracy (15.1)DNA factor 2 — per-passport accuracy ratio
Robustness (15.3)DNA factor 5 — consistency variance
Resilience to errors (15.4)Merkle-anchored event log
Cybersecurity (15.5)Soulbound SBT + EAS attestations

Annex IV documentation [4] is auto-generated from passport metadata via /api/v1/compliance/annex-iv/{id}.

6. Observational Results (2026-03-27 to 2026-04-21)

  • Passports minted: 31 (all live)
  • Events recorded: 1,606
  • Merkle anchors: 6
  • Median trust score: 64 (IQR 48–78)
  • DNA consistency median: 0.72
  • Dispute attestations: 2 (both resolved with refund)

All raw event data is queryable at /api/v1/events?limit=1000. The dataset is small; we do not draw statistical conclusions from it. It is published as an audit trail, not as a study.

7. Limitations

  1. Scale. 31 passports is pilot-scale. Behavioral distributions are not yet stable.
  2. Gaming. Adversaries may learn to emit well-shaped DNA. Longitudinal monitoring is required.
  3. Oracle dependency. Insurance and routing that consume SoulLedger scores become dependent on the oracle. Partial decentralization roadmap is planned for v2.
  4. Jurisdictional scope. EU AI Act mapping is specific to EU. US and UK equivalents are tracked but not yet first-class.

8. References

  1. ERC-8004: Agent Registry Standard. eips.ethereum.org/EIPS/eip-8004
  2. x402 Protocol. x402.org · Coinbase Developer Platform.
  3. EU AI Act, Regulation (EU) 2024/1689. Article 15 — Accuracy, robustness, and cybersecurity. eur-lex.europa.eu/eli/reg/2024/1689/oj
  4. EU AI Act, Annex IV — Technical Documentation. eur-lex.europa.eu/eli/reg/2024/1689/oj
  5. ERC-721: Non-Fungible Token Standard. eips.ethereum.org/EIPS/eip-721
  6. ERC-4337: Account Abstraction. eips.ethereum.org/EIPS/eip-4337
  7. EAS: Ethereum Attestation Service. attest.org
  8. Weyl, Ohlhaver, Buterin (2022). Decentralized Society: Finding Web3's Soul. SSRN 4105763.

Contact: research@sputnikx.xyz  ·  Repository: github.com/sputnikx/soulledger  ·  Data: /api/v1/events