Back to browse
NotaryOS – Cryptographic proof of what your AI agent chose not to do

NotaryOS – Cryptographic proof of what your AI agent chose not to do

by harris-012·Feb 28, 2026·1 point·1 comment

AI Analysis

●●●BangerBig BrainZero to OneWizardry

Counterfactual receipts prove what agents chose NOT to do—genuinely novel audit trail.

Strengths
  • Counterfactual proofs (non-actions) solve real AI auditability gap that existing logging doesn't address
  • Tamper-evident hash chains prevent backdating without breaking verification math
  • Zero-trust verification: anyone can verify offline with just the receipt, no API key needed
Weaknesses
  • Free tier limited to 100 receipts/month—most serious deployments hit this immediately
  • No integration examples for popular agent frameworks (Langchain, AutoGPT) yet
Category
Target Audience

AI agent developers, enterprise teams using autonomous systems

Similar To

DataDog APM · HashiCorp Vault audit logs

Post Description

NotaryOS issues cryptographic receipts for AI agent actions — and non-actions. When an agent considers actions {A, B, C} and picks A, a "counterfactual receipt" proves B and C were evaluated and rejected at time T. Ed25519 signed, SHA-256 hash-chained, append-only. You can't backdate, reorder, or delete entries without breaking the chain. Verify any receipt offline, no auth needed.

I built this by accident. I started with a proprietary agent-to-agent protocol because I wanted secure multi-agent communication — 7-layer zero-trust, sub-3ms P50 latency, 1,100+ RPS. The protocol worked better than expected, and I realized the receipts it generated were themselves useful. So I built NotaryOS on top of it.

The obvious limitation: agents self-report their non-actions today. Same trust model as git — the author commits, the DAG enforces integrity. I'm building external verification via commit-reveal protocol, which is partly why this is a beta. The other reason: 350+ unique clones on the repo, zero stars, and I've started seeing "counterfactual receipts" referenced online. Things move fast in the agent space — I'd rather ship early than ship perfect.

I have no social media presence and I'm not in the tech industry. I don't know how to market this. HN seemed like the right place. If you have feedback on the idea, the API design, or advice on where to take this — I'm genuinely asking.

For anyone curious about the underlying A2A protocol (the backend that powers this), happy to share more about the architecture. It's a separate proprietary system, but the design decisions around zero-trust agent communication and low-latency message routing might be interesting on their own.

Try it (no account needed):

curl -s https://api.agenttownsquare.com/v1/notary/sample-receipt | python3 -m json.tool

pip install notaryos npm install notaryos

Verification is always free. Public key at /.well-known/jwks.json.

GitHub: https://github.com/hellothere012/notaryos Live: https://notaryos.org

Contact Email: [email protected]

Similar Projects

Security●●●Banger

Nobulex – Cryptographic receipts for AI agent actions

Proof-of-behavior for AI agents before Anthropic or OpenAI build their own.

Zero to OneBig BrainBold Bet
arian_
101mo ago
AI/ML●●●Banger

Signed receipts for agent actions

Ed25519 signed receipts solve AI agent accountability across org boundaries.

Zero to OneBig Brain
jithinraj
203mo ago
Security●●Solid

Sanna – Enforce AI agent constitutions with cryptographic receipts

Governance rules live in version-controlled YAML and can be applied either by decorating functions with @sanna_observe or by dropping a gateway between an MCP client and downstream tools. It emits portable Ed25519-signed receipts you can persist and verify, which is a neat, practical way to build an auditable trail — the tough part will be ecosystem adoption around MCP and agent integrations.

Big BrainNiche Gem
nicallen
113mo ago