AI Treasury Council

DAOs hold over 26 billion dollars in treasury assets, but governance suffers from voter apathy and decision concentration. A single whale can push through a 100k USDC allocation while 90% of token holders abstain. There is no structured analysis, nobody plays devil's advocate, and the reasoning behind decisions disappears.

AI Treasury Council changes this. Submit a treasury proposal (e.g. "Allocate 100k USDC to Aave for yield"), and 5+1 specialized AI agents analyze it from different angles: Bull (growth opportunities, optimistic), Bear (downside risks, skeptical), Risk (quantitative, neutral), Tech (smart contract safety), Sentiment (community plus market mood), and an opt-in Adversarial agent that actively probes for jailbreaks and weak reasoning.

Each agent cites its sources (RSS feeds, CoinGecko, DefiLlama) with confidence weights. The full debate transcript is stored immutably on 0G Storage with content-addressed CIDs (with IPFS Pinata as automatic fallback). Token holders then vote on-chain via OpenZeppelin Governor with a 48-hour timelock before execution on Base Sepolia.

Five Trust Mechanisms (research-backed): source attribution per claim, timelock countdown UI for human review, immutable audit trail on 0G Storage, ENS reputation badges per agent (26 text records each on Sepolia under aicouncil-danergy.eth), and human-in-the-loop council rules.

Proof-of-Work for Agents (Moat 5) is unique among AI agent products: agent quality is cryptographically auditable. Each of the 5 agents has a permanent on-chain reputation score in AgentReputation.sol on Base Sepolia. After every debate, the orchestrator credits agents whose vote aligned with consensus and decrements those who dissented. Agents earn trust through verifiable history, not marketing claims. Cross-DAO portable - same agent participating in multiple councils accumulates reputation across all of them.

Real implementation, no smoke and mirrors: real Anthropic Claude Sonnet 4.6 calls with prompt caching (60-90% cost reduction), 5 contracts verified on Base Sepolia, 5 ENS subnames live on Sepolia with 26 text records each, real deterministic agent EOAs (keccak256 seed - we passed the ENS judges' smell test), real system_prompt_hash (keccak256 of actual prompt) per agent on-chain, Demo Mode marker (Charter principle of honest communication) so judges can use the full UI without a wallet.

Smart contracts: Solidity 0.8.24 + Foundry + OpenZeppelin Contracts v5 (Governor, ERC20Votes, TimelockController). Pre-deploy security audit by Mateusz agent: 0 CRITICAL, 0 HIGH findings. 23/23 Foundry tests pass. Deploy via Wizard pattern, no custom dangerous Solidity.

Frontend: Next.js 16 + Tailwind CSS v4 + shadcn/ui. Wagmi v2 + RainbowKit + viem for wallet and contract interactions. Custom i18n provider for PL+EN bilingual UI (next-intl was incompatible with Turbopack + pnpm + Next 16, documented in ADR-002). Framer Motion for typewriter streaming debate viewer.

Backend: Python 3.11 + FastAPI + Pydantic v2. Anthropic SDK with prompt caching gives 60-90 percent cost reduction per debate (around $0.04 per debate at scale). WebSocket /ws/debate for live streaming. SQLAlchemy 2.0 + Alembic on Postgres, Redis for caching and rate limiting (slowapi 10 req/min/IP).

AI orchestration: Claude Sonnet 4.6 across 5+1 personas. Cut CrewAI early in the sprint (ADR-002) - direct Anthropic SDK calls with structured outputs gave better control and lower latency for a 5-agent debate.

Storage (sponsor 0G Labs): 0G Storage SDK as primary audit trail (factory pattern in apps/api/storage/), IPFS via Pinata as automatic fallback when 0G testnet is flaky. Both store the same content-addressed transcript and we record the CID on-chain.

Identity (sponsor ENS): Sepolia. Registered aicouncil-danergy.eth and minted 5 subnames directly via viem (skipped NameStone after signup friction). Each subname has 26 text records covering persona description, reputation snapshot, cross-chain pointer to AgentReputation contract on Base Sepolia, avatar, and more. ENS resolution chain uses 4 RPC providers (publicnode, tenderly, rockx, 1rpc) for reliability. Real deterministic agent EOAs via keccak256(system_prompt) - smell test passed.

Infra: Vercel for frontend, Railway for backend with Redis Hobby and Postgres Hobby addons. Foundry CI/CD via GitHub Actions plus gitleaks pre-commit.

Process: Solo human founder coordinating a 15-agent AI dev-team via Claude Code over a 3-day sprint. Every commit went through automated code review (Critic agent) and security audit (Mateusz agent) before merge. Worktree-per-agent pattern (ADR-003) prevented conflicts across 7 sessions running in parallel.

Honest scope: 5 contracts real, ENS real, 0G real, Anthropic real. Multisig signers are mock with explicit "demo mode" marker in UI (Charter #7 honest communication).