NOTES
[GOAL]
Eliminate herd effects in multi-agent voting. Three LLM validators (KIMI, CLAUDE, GEMINI) vote on DAO proposals using commit-reveal cryptography. No partial tallies during commit/lock phases. Results appear only after the reveal window ends and hashes verify.
[MODEL SETUP]
KIMI: deterministic on, abstains if confidence < 0.55
CLAUDE: adversarial guards high, p95 latency ≈ 230 ms
GEMINI: CoT hidden, fixed seed
[COMMIT-REVEAL]
1. Validator computes: commitment = sha256(vote|salt)
2. Submit commitment hash during COMMIT phase
3. LOCK: brief seal, no new commits
4. REVEAL: submit (vote, salt)
5. System verifies: sha256(vote|salt) === commitment
6. TALLY: compute majority, tie → ABSTAIN
[THREAT MODEL]
Protects against: early leakage, partial tallies influencing late voters, herd cascades. Does not protect against: collusion among validators, adversarial prompt injection (mitigated by per-validator guardrails).
[NO-HERD POLICY]
During COMMIT/LOCK: no vote directions or counts shown. During REVEAL: show only "reveals received X/3" and hash verification status. In TALLY: show full counts and final verdict.
[API SCHEMA]
GET /api/proposals → Proposal[]
POST /api/proposals {title, description} → Proposal
POST /api/commit {proposalId, validatorId, commitment, ts}
POST /api/reveal {proposalId, validatorId, vote, salt}
GET /api/state → GlobalState
GET /api/logs → LogEntry[]