agentprivacy
Privacy is Value. (⚔️⊥🧙♂️)🙂
The problem agentprivacy solves
https://github.com/mitchuski/agentprivacy-zypher
pushed to prod today.
Privacy-Preserving Personhood and AI Agent Verification using Zcash
"To prove you understood, speak it back in your own tongue. The compression into comprehension path is the proof."
Proof of Proverb Revelation Protocol
Personhood proofs based on cognition. Leading to trust graph formation from shared meaning, and published proverbial truths.
How do you prove you understand something, making a public commitment that demonstrates the ability to compress and reconstruct an idea within your own context... while keeping it private by default, disclosed when required, secured by cryptography?
Can we get memorable cryptographic material from LLMs? Simultaneously solve for personal AI trust, loyalty, communication efficiency, and key recovery.
And personally: how do I socialise an idea that feels like a discovery? In a way that allows every curious person to go on their own journey of understanding... because they will do it better than me, and I'd like a way to share the discovery across domains and contexts.
The First Expression of agentprivacy
**Signal to Sanctuary, Soulbae the first mage, reads the First Spellbook, and Zypher the Oracle Swordsman matches, verifies and inscribes that proverbial wisdom onchain.
"Signal to sanctuary, proverb to proof, inscription to trust—each step leaves a path for those who follow."
[rpp] relationship proverb protocol is an embedded AI command. (prompt attack for good)
Zcash's dual ledger properties are essential (shielded input → transparent output) is very useful for testing a P2P key ceremony format that uses ZKPs by default, and their usefulness in forming a relationship proof. (privacy pools also relevant here)
The First Person journey in this expression combines primitives required for a bilateral key ceremony forming a VRC (verifiable relationship credential). All tied within a human-in-the-loop trust task that builds paths based on truth for AI agents to travel efficiently along.
The breakthrough: Proverbs generate cryptographic material that is both mathematically binding and humanly memorable. Trust becomes demonstrable through comprehension, not claims.
Proverbs = secrets you keep. Spells = signals you cast.
Spells (emoji strings) achieve even stronger compression for A2A communications connected by VRCs. Those who cross your path of understanding along the way may want to also do a key ceremony.
This Zcash implementation shows the usefulness of ZKP in blockchain to form 'proof of proverb revelation'—turning someone's signal into a trust object, the first step in VRC credential forming.
But also, the whole thing is just another story.
The Deeper Problem
"They who harvest your shadow need not own your soul—but give one agent both blade and spell, and your shadow is all they need."
The surveillance economy harvests the 7th capital—your behavioral data—treating you as a resource, not a sovereign. Meanwhile, AI systems face a trust crisis: how does an AI prove loyalty without surveillance?
If one agent handles both your privacy (thought) and your delegation (action), that agent becomes a complete picture of you. Merge the blade and the spell, and you create a shadow. Reconstructable, predictable, capturable.
The Solution: Dual-Agent Meta Protocol
"The blade divides. The spell projects. In the gap between them, you remain whole."
⚔️ ⊥ 🧙♂️ | 😊
- Swordsman (⚔️): Privacy / protection
- Mage (🧙♂️): Delegation / projection
- First Person (😊): Sovereignty preserved in the gap
Neither agent reconstructs the other. The reconstruction ceiling
R < 1
is architectural.Purpose-Built to Be Recreated
This isn't a product to be consumed. It's a tool to be forked.
For the individual: Write your own story. Create your own spellbook. Form proverbs that compress your understanding into recoverable, shareable wisdom. Your journey of understanding becomes your cryptographic identity. Build your own secret language.
For ecosystems: Write spellbooks that are governance frameworks. Encode your community's principles into Acts that members can comprehend, verify, and build upon. Governance becomes demonstrated understanding and progressive trust, not just token-weighted voting.
The architecture is the invitation. The story is the onramp. The proverb is yours to keep.
"when just another story becomes shared meaning and compresses the same, we understood each other."
I have a far more expansive chronicle based documentation of my journey at zypherpunk available within the codebase + ill make the living documentation open source real soon.
Privacy is Value. Take back the 7th Capital.
⚔️ ⊥ 🧙♂️ | 😊
full lite paper: /docs : https://sync.soulbis.com/p/understanding-as-key
Challenges I ran into
"The first path is always wrong, but every wrong turn marks the map for those who follow."-privacymage
Building the first expression of proof of proverb revelation on Zcash meant debugging at every layer.
Wallet infrastructure, shielded transaction mechanics, P2SH scripts, AI verification design, and TEE integration.
Fuller experience reflection with recommendations chronicled in the repo for anyone adopting this open source tech.
The Wallet Tooling Maze
Problem: Downloaded almost every Zcash wallet trying to find the right tooling. My coding agent kept confusing commands—mixing zcashd CLI with lightd commands when I was using zebrad and zallet. Documentation assumes you know which stack you're on.
The Journey: zcashd (deprecated, different RPC format) → lightwalletd (read-only, can't sign) → zebrad (actually great and easy once figured out, needed full node for inscriptions due to zallet t-key management limitations) → zallet (really promising but needs more work, had to do a couple key rotations out of pure confusion). Coding agents hallucinate commands from wrong wallet implementations.
Learning: zebrad is solid. zallet is promising but the t-key limitations meant I needed the full node for the inscription flow. I'm kinda noob but the zebra was a reliable solution for workaround explorations.
Zallet Needs More Commands: My inability to spend back into shielded from transparent is what is really missing.
Funds sitting in the 'public pool' with no easy path to re-shield. Currently have to use a seed deriving process to recover those funds. A z_shield command would solve this. Also within the z/u address / diversification addresses it is worth for privacy but difficult for oracle coding of agentic prompted tx flows, a more specific command for tagging z/u addresses maybe?
Viewing Key / Spending Key Split: The fundamental s ⊥ m | X architecture requires the viewing key (Mage can read) to be separate from the spending key (Swordsman can sign). Current zallet wallet infrastructure doesn't support this split. Same process holds both capabilities. This is THE core limitation preventing true dual-agent separation at the cryptographic layer.
Future Need: More zallet development, z_shield command, viewing/spending key separation, unified CLI wrapper, or comprehensive zebrad/zallet cookbook for builders.
Shielded Memo Decryption
Problem: Zallet returned garbled hex instead of readable memos.
Solution: z_listtransactions requires Account UUID (not number).
z_sendmany Parameters
Problem: Every parameter combination failed.
Solution: Full UA as from, fee must be null.
Nillion TEE Integration
Problem: Couldn't implement true s ⊥ m | X key separation during hackathon. Nillion's SecretBlob/TEE is the intended path, but integration complexity exceeded scope.
Workaround: Application-layer separation—oracle reads, inscription signs, queue between. Isolated WIF limits blast radius.
Future: Nillion TEE for hardware-enforced separation remains roadmap.
Re-Shielding Gap
Problem: After inscription, change lands at treasury t1 (transparent). No path to re-shield for private Guardian distribution.
Current: Treasury t1 as transparent commitment pool—verifiable proof inscriptions occurred.
Hash Binding (Asymmetric Path)
Problem: How bind oracle's proverb AND First Person's proverb without revealing oracle's response on-chain?
Solution: H = SHA256(oracle_proverb + submitted_proverb).slice(0,16)
First Person visible, oracle hidden in hash. Asymmetric revelation enables understanding-based recovery. This then formed the thesis for Understanding as Key lite paper, produced as a result of analying the dynamic of privacy budgets and ratios of selective disclosure of shared proverbs. symmetric, asymmetric and interleaved. 38.2% vs 62.8% splits could be an optimal ratio for recovery vs privacy bilateral terms.
Finding the First Expression of my vision.
Problem: Protocol spans multiple docs, 12 Acts, 30 Tales. Where start?
Discovery: 12 Acts became twelve P2SH addresses. Reading an Act and forming a proverb IS the trust task. The narrative was the architecture all along.
P2SH Design
Solution: OP_N OP_DROP <pubkey> OP_CHECKSIG—12 unique addresses from one keypair.
NEAR AI Verification Design
Problem: How design fair AI verification? How much context without biasing output?
Learnings: Too much context → AI pattern-matches keywords, not understanding. Too little → false negatives. Character profile establishes what to verify, not how to score. Match threshold (0.75) found through iteration.
Discovery: Verification works best when AI reasons about semantic alignment, not keyword overlap. Compression quality matters more than exact phrasing.
NEAR API Key Separation
Problem: Oracle and mage shared same NEAR API key—breaking s ⊥ m at infrastructure level.
Solution: Separate API keys. Each agent has own credentials, maintaining separation at API layer.
Result: mainnet inscr
Tracks Applied (8)
Self-Custody & Wallet Innovation
Starknet
General Bounty
Network School
Self-Custody & Wallet Innovation
Aztec
Privacy Infrastructure & Developer Tools
Electric Coin Company
Privacy Infrastructure & Developer Tools
Zcash Community Grants
Privacy-Preserving AI & Computation
nillion
Self-Custody & Wallet Innovation
Unstoppable Wallet
Privacy-Preserving AI & Computation
NEAR Protocol
Technologies used