Authority tiers
Every action is classified T0–T4. Reading is autonomous; committing and spending require recorded human authorization. The AI cannot spend money or make legal commitments on its own.
Security & compliance
Safe by construction, not by instruction.
Portfolio Systems is engineered so the failure mode is “asks a human,” never “acts wrongly.” The properties below are enforced in code and in the database.
Tenant isolation by the database
Row-level security in Postgres means one portfolio cannot read another’s data — even with a deliberately unfiltered query. Isolation is enforced where a bug cannot reach it.
Append-only memory
Records are immutable and attributed. Corrections are supersessions, not edits, so the audit trail of what was promised and when is continuous.
Data minimization
Memory stores operational facts, not private causes. No payment-card numbers, SSNs, immigration status, or medical diagnoses — by design and by validation.
Jurisdiction-aware compliance
Each portfolio carries its own law and language pack, so the AI applies the right rules — consent, notice, habitability — for the jurisdiction it serves.
Resilient, vendor-agnostic
Automatic failover across models and providers keeps the front desk answering when one vendor degrades — no single point of failure in the AI path.