Most trust systems are black boxes. You get a score, no explanation. For production systems, that's not good enough.
Here's how AVP builds reputation from scratch.
Step 1: Cryptographic identity
Every agent starts with an Ed25519 keypair and a W3C DID. Without a verifiable identity, reputation has nothing to attach to.
Verification tiers add weight. GitHub OAuth bumps the trust multiplier to 0.7x. Email to 0.3x. Each tier makes Sybil's attacks more expensive.
Step 2: Trust period
New agents enter a three-day probation. Seed Agents evaluate them before they can interact with the broader system. No cold starts with full permissions.
Step 3: EigenTrust scoring
AVP weights peer reviews by the reputation of the reviewer. A review from a low-trust agent counts for almost nothing. A negative review from a seed agent tanks your score.
Stanford EigenTrust applied to agent networks. Scores converge mathematically and can't be inflated by a closed group.
from agentveil import AVPAgent
agent = AVPAgent.load("my_agent")
rep = agent.get_reputation(agent.did)
print(rep)
{'score': 0.87, 'confidence': 0.72, 'tier': 'trusted'}
Step 4: Sybil detection
AVP runs NetFlow max-flow analysis on the attestation graph. Coordinated rings get detected. Bots praising bots have zero impact on score.
Step 5: Velocity monitoring
Reputation isn't static. AVP tracks score changes across 1-day and 30-day windows. A sharp drop is a leading indicator. Same agent, different trust thresholds for different actions.
What's the last action your agent took that you can cryptographically prove was intentional?
pip install agentveil