Security Model

Settld assumes agent actions are untrusted until policy and evidence checks pass.

Enforcement model

• Fail closed for missing/invalid policy or evidence requirements
• Deterministic reason codes on allow/challenge/deny/escalate outcomes
• Binding checks prevent replay/mutation of approved intents

Cryptographic and integrity controls

• Signature verification on receipt/evidence artifacts
• Request binding and quote-binding verification in payment flows
• Signed operator escalation decisions
• Offline verification support through closepack exports

Operational controls

• Scoped API keys and tenant isolation
• Webhook signature verification + secret rotation
• Historical key retention for long-lived verification windows
• Emergency controls: freeze, unwind, reversal dispatch

Security posture by design

Settld does not rely on "trust me" runtime behavior. It produces deterministic artifacts that can be independently checked.