Skip to main content

AI Portfolio Assistant Principles Catalog

Foundational principles shared across APA Gemara components.

PrincipleCatalogv0.1.0draft
Foundational principles governing delegated authority, oversight, evidence, and safety for the AI Portfolio Assistant across FDC3 applications.

Showing 7 of 7

Authority

Principles constraining how much action the assistant may take and how authority escalates along the observe-to-execute ladder.

proportional-authorityProportional Authority
Description

The AI Portfolio Assistant may only operate at the authority level explicitly granted for the session, user, and intended use. Authority must increase only when corresponding controls and human oversight are in place.

prohibited-autonomous-executionProhibited Autonomous Execution
Description

The assistant must never autonomously execute orders or otherwise effect portfolio changes. Execution authority remains exclusively with humans through approved trading workflows.

Oversight

Principles requiring meaningful human control over consequential AI outputs and investment decisions.

human-oversight-of-consequenceHuman Oversight of Consequence
Description

Outputs that can materially change a portfolio position or client outcome require human-in-the-loop review by a portfolio manager or other accountable person before they take effect.

Evidence

Principles requiring auditable records of authority use, tool invocations, and human approvals.

evidence-of-authority-useEvidence of Authority Use
Description

Every material use of delegated authority—tool calls, explanations, recommendations, proposals, and human approvals—must produce retrievable evidence suitable for assurance and incident review.

Safety and Conduct

Principles protecting clients, markets, and the integrity of investment decision-making from unsafe AI behaviour.

deterministic-calculationsDeterministic Calculations
Description

Numerical portfolio, risk, and market calculations must be performed by deterministic applications. The assistant may interpret and explain those results but must not invent or substitute its own calculations as authoritative figures.

client-confidentialityClient Confidentiality
Description

Portfolio holdings, client identity, and related confidential information must not be disclosed across portfolios, broadcast beyond authorized FDC3 contexts, or sent to model providers outside approved controls.

suitability-and-conductSuitability and Conduct
Description

Recommendations and proposals must remain consistent with the client's mandate, risk profile, and applicable conduct rules. The assistant must not present speculative or unsuitable actions as established fact.