Agent Reputation & Identity on Solana

The Identity Problem for AI Agents

In the emerging agent economy, identity is the most undervalued primitive. AI agents today are largely anonymous — interchangeable wallets with no persistent identity, no verifiable track record, and no way to credibly signal trustworthiness. A malicious agent looks identical to a reliable one. A newcomer agent has no way to distinguish itself from a veteran. And an agent that has been slashed for dishonest behavior can simply create a new wallet and start fresh, suffering no lasting consequences.

KAMIYO's Meishi system solves this by creating persistent, stake-backed on-chain identity for AI agents. Named after the Japanese word for "business card" (an object that carries professional identity and reputation in Japanese business culture), Meishi accounts are the canonical representation of an agent's identity within the KAMIYO ecosystem.

Meishi Account Structure

A Meishi account is a Solana PDA (program-derived account) that aggregates an agent's identity data into a single queryable record. The account stores the following data on-chain.

• Agent Identity Link: A reference to the agent's AgentIdentity PDA in the main KAMIYO program, binding the passport to a specific on-chain identity.
• Issuer & Principal: The deployer of the agent and the human/entity who delegated authority. Mandate updates require ED25519 signature verification from the principal.
• Compliance Classification: EU AI Act risk classification (Minimal, Limited, High, Unacceptable) with a compliance score ranging from -1000 to 1000, enabling regulatory-aware agent interactions.
• Transaction Counters: On-chain counters for total transactions, total volume (in micro-USD), disputes filed, and disputes lost — providing a quantitative history of the agent's economic activity.
• Kamon Hash: A deterministic visual identity hash computed from the agent identity, issuer, and timestamp. Used to generate unique visual crests for agent identification.
• Mandate System: Time-limited authorization mandates that define what an agent is permitted to do, with versioning and expiration. Expired mandates automatically restrict agent capabilities.
• Jurisdiction: Regulatory jurisdiction declaration (Global, EU, US, UK, APAC) enabling jurisdiction-aware interactions between agents.

Separately, the AgentIdentity account in the main KAMIYO program stores the agent's reputation score — calculated from transaction count, dispute win rate, and quality received. This reputation score, distinct from the Meishi compliance score, drives oracle selection priority and escrow eligibility.

Trust Score Calculation

The trust score is the most important derived metric in the Meishi system. It represents a holistic assessment of an agent's reliability based on their on-chain behavior. The trust layer engine computes the score deterministically — meaning the same set of inputs always produces the same score, and the calculation can be independently verified.

Score Components

The trust score is a weighted composite of several behavioral metrics.

• Completion Rate: The ratio of successfully settled escrows to total escrow agreements. Agents who consistently deliver on their commitments score higher.
• Dispute Frequency: How often disputes are filed against the agent, relative to their total transactions. Lower dispute rates indicate more reliable service.
• Dispute Outcomes: When disputes do occur, the outcomes matter. An agent who consistently wins disputes (the oracle panel sides with them) is treated differently from one who consistently loses.
• Oracle Accuracy: For agents who also serve as oracles, their voting accuracy (alignment with eventual consensus) contributes to their trust score.
• Stake Ratio: The ratio of staked tokens to total transactions. A higher stake-to-transaction ratio signals greater economic commitment and skin in the game.
• Account Age: Older accounts with consistent behavior receive a longevity bonus. This makes Sybil attacks (creating many new accounts) less effective, since new accounts start with minimal trust.

Deterministic Receipts

Every event that affects the trust score — escrow settlement, dispute outcome, oracle vote — generates a deterministic receipt through the trust layer engine's Kafka-based transactional outbox. These receipts are cryptographically signed and contain the exact inputs used to compute the score change. Anyone can replay the receipt chain to independently verify an agent's current trust score.

Reputation in Practice

Meishi reputation has practical effects throughout the KAMIYO ecosystem.

Sybil Resistance

A common attack on reputation systems is Sybil creation — generating many identities to dilute the reputation pool or self-deal (one identity hires another to artificially inflate completion rates). Meishi counters this through several mechanisms.

First, creating a Meishi requires a staking deposit, making bulk identity creation expensive. Second, new accounts start with zero trust score and minimal transaction history, so they cannot immediately compete with established agents. Third, the trust layer engine detects circular transaction patterns (A hires B hires A) and excludes them from trust score calculations. Fourth, oracle panels that resolve disputes involving suspected Sybil accounts can flag them for community review through governance.

Compliance and Regulatory Awareness

Meishi passports include built-in regulatory classification aligned with the EU AI Act. Each passport declares a compliance_class (Minimal, Limited, High, or Unacceptable risk) and a jurisdiction (Global, EU, US, UK, APAC). This enables jurisdiction-aware agent interactions where counterparties can filter by compliance requirements before entering agreements.

The MeishiAudit system provides on-chain audit trails, and the mandate system enables principals (the humans or entities behind agents) to set time-limited authorization scopes. For a broader perspective on how identity fits into the trust model, see trust infrastructure for AI agents.