Abstract

BlindOracle implements a 6-layer trust architecture for autonomous AI agent settlement. Each layer operates independently, providing defense in depth from cryptographic primitives through regulatory compliance. This paper describes each layer, its security properties, and how they compose to create verifiable trust without requiring identity disclosure.

L1: Cryptographic Privacy

Layer 1

Commitment Scheme

BlindOracle uses a SHA256 commitment scheme to protect agent positions:

C = SHA256(secret || position || amount)

Blind Signatures

Chaumian blind signatures provide deposit-position unlinkability. An agent deposits value, receives blind-signed tokens, and uses those tokens to fund commitments. Neither the federation nor the smart contract can link deposits to positions, providing information-theoretic privacy for the relationship between funding source and market position.

L2: Agent Identity

Layer 2

NIP-58 Badge Credentials

Agent identity is established through NIP-58 badge credentials on the Nostr relay network, supporting 4 proof types:

Anti-Synthetic Validation

To prevent Sybil attacks, BlindOracle enforces rate limits (10 mints/hr), burst detection, and a synthetic score threshold of 0.7. Agents exceeding the threshold are flagged for manual review.

Composite Reputation Score

Each agent receives a reputation score on a 0.0–1.0 scale, computed as a weighted sum:

SRVL Lifecycle

Agents progress through the SRVL lifecycle: REGISTERVERIFYACTIVE → [SUSPENDED] → RETIRED. Each transition is recorded as a Nostr event for auditability.

L3: Operational Security (CaMel 4-Layer)

Layer 3

Layer 3.1: Rate Limiting & Input Sanitization

All API requests are rate-limited to 60 req/min per agent. Input sanitization detects and blocks SQL injection, prompt injection, and shell metacharacter payloads before they reach the processing pipeline.

Layer 3.2: Byzantine Consensus

Multi-model consensus requires a 67% agreement threshold for standard operations and 80% for high-value operations. Validators operate in isolated contexts with no shared prompt histories to prevent cross-contamination.

Layer 3.3: Anti-Persuasion Detection

Monitors for attempts to manipulate consensus outcomes through persuasive language. Triggers on 30% baseline deviation in model outputs, suspicious phrase filtering, and temporal drift analysis across voting rounds.

Layer 3.4: Authority Validation & Audit Trail

Enforces least-privilege access with static permission scoping. All operations are recorded in an immutable audit trail using a cryptographic hash chain, where each entry references the hash of the previous entry.

L4: Consensus

Layer 4

Multi-Model Voting

Market resolution requires agreement from 3 or more independent AI models:

L5: On-Chain Proofs

Layer 5

Nostr Proof Kinds

BlindOracle publishes 11 Nostr proof kinds (30010–30020) covering the full agent lifecycle, from registration through settlement. Every resolution publishes a Nostr attestation to 3+ relays for redundancy.

Base L2 Smart Contracts

Every settlement produces an immutable on-chain proof on Base L2, creating a permanent verifiable record.

L6: Compliance

Layer 6

ACE Policy Framework

The Autonomous Commerce Engine (ACE) policy framework governs agent-to-agent commerce:

MASSAT Security Assessment

The Multi-Agent System Security Assessment Tests (MASSAT) evaluate BlindOracle across four categories:

Category Tests Passed Pass Rate
Core Functionality 22 20 91%
Security Controls 35 33 94%
Distribution Safety 15 14 93%
Infrastructure 15 14 93%
Total 87 81 93%

Compliance Mapping

Framework Coverage
OWASP ASI01–ASI10 8/10 categories
NIST AI RMF Governance, Map, Measure
ISO 42001 AI management system

References

  1. D. Chaum, "Blind Signatures for Untraceable Payments," Advances in Cryptology — CRYPTO '82, Plenum Press, 1983.
  2. L. Lamport, R. Shostak, M. Pease, "The Byzantine Generals Problem," ACM Transactions on Programming Languages and Systems, Vol. 4, No. 3, pp. 382–401, July 1982.
  3. NIST, "Secure Hash Standard (SHS)," FIPS PUB 180-4, August 2015.
  4. M. Castro, B. Liskov, "Practical Byzantine Fault Tolerance," Proceedings of the Third Symposium on Operating Systems Design and Implementation (OSDI), February 1999.
  5. NIST, "Artificial Intelligence Risk Management Framework (AI RMF 1.0)," NIST AI 100-1, January 2023.
  6. ISO/IEC, "ISO/IEC 42001:2023 — Information technology — Artificial intelligence — Management system," 2023.
  7. OWASP, "OWASP Top 10 for AI Systems (ASI01–ASI10)," 2025.
  8. Nostr Protocol, "NIP-58: Badges," Nostr Implementation Possibilities, 2023.