The agent's runtime contract: what it can and cannot do, even when the user asks. Risk-tiered (low/medium/high/critical) with per-tier confirmation requirements. The constitution is enforced by code, not policy — an admin can read the source and verify behavior.
Documented adversaries: cloud-side eavesdropper, supply-chain compromise, lost device, voice spoofing, prompt injection from shared content. Mitigations are listed inline in the source so reviewers can audit each defense without a separate document. Voiceprint enrollment + Constitution §12 critical-tier confirmation cover the high-risk class.
iPhone (trigger) → Center (cloud router, no models) → Provide (GPU compute) → Mac (executor). All transit is TLS. Models do not run on iPhone or Mac. Memory and identity stay on your devices; only audio streams reach Provide for ASR/LLM. Center never logs voice content — it routes packets and that's it.
Found something? Email security@voxsign.ai with details. We acknowledge within 48 hours, fix critical issues within 7 days, and post a public CVE-style advisory after the fix ships. No bug-bounty cash yet — we'd rather pay engineers fast than gamify disclosures.