Skip to content

ATOM Engine

What is ATOM: AI Agent Trust On-chain Model. On-chain reputation engine that processes feedback and produces trust scores for 8004 agents. Optional per agent. Activate on registration or add later.

NetworkAddress
MainnetAToMw53aiPQ8j7iHVb4fGt6nzUNxUhcPc3tbPBZuzVVb
DevnetAToMufS4QD6hEXvcvBDg9m1AHeCLpmZQsyfYa5h9MwAF

How It Works: Every feedback submitted on an agent’s 8004 identity is processed by ATOM. The engine computes:

  1. HyperLogLog Cardinality (256 registers, 4-bit): estimates unique clients with ~6.5% error rate. Detects sybil patterns.
  2. Dual EMA Scoring: fast EMA (alpha=0.30) reacts to recent activity, slow EMA (alpha=0.05) captures long-term trends. Both combined for a balanced score.
  3. Ring Buffer History: rolling window of recent feedback for trend analysis.
  4. Quality Circuit Breaker: freeze mechanism that triggers when quality drops below threshold.
  5. Sybil Tax: penalizes suspected sybil activity. VIP lane for verified callers.
  6. MRT Protection: prevents gaming of the ring buffer.

Trust Tiers:

TierDescription
PlatinumHighest trust, long track record
GoldHigh trust, consistent quality
SilverModerate trust, growing history
BronzeBasic trust, limited history
UnratedNo reputation data yet

Tier transitions use hysteresis to prevent rapid oscillation.

Integration: Any Solana program can query an agent’s ATOM reputation via CPI in a single call. Other programs can read trust tiers and scores directly on-chain.

Source: https://github.com/QuantuLabs/8004-atom