Reference
Security
Security is structural: separation of powers across evidence, authority, and capital.
Threat model (Agency)
| Threat | Control |
|---|---|
| Unbounded admin key | Allowlist of targets + selectors |
| Model proposes malicious calldata | On-chain filtering + simulation gate |
| Manual response too slow | Bounded autonomous execution |
| High-impact autonomous error | Value-threshold multi-sig escalation |
| Credential leakage | HSM/MPC; keys never in model context |
Oracle guarantees
- Authenticity — recoverable authorized signer
- Integrity — score bound in typed data
- Freshness — timestamp window enforced by consumers
- Availability semantics — missing ≠ zero
Key separation
- Model context: observations in, proposals out — no secrets.
- Signing service: accepts only policy-approved calldata.
- Policy parameters: on-chain, timelocked, L1-anchored where required.
Operational checklist
- Monitor attestation freshness and oracle state.
- Monitor Arbitrum sequencer liveness (stall > 30s → degrade).
- Alert on policy rejection spikes and escalation frequency.
- Run fork tests against production addresses before upgrades.
Disclosure
Report vulnerabilities to security@ukiyo.network. Do not open public issues with exploit details.