Epistemic Accountability for Agentic Financial AI: The Transformer Mandate and Evidence Lifecycle Management

Agentic AI systems in finance -- algorithmic traders, autonomous risk assessors, multi-agent portfolio optimizers -- make consequential decisions faster than human operators can audit them. A trading agent that generates a risk assessment, cites its own prior output as supporting evidence, and escalates its confidence creates a closed epistemic loop with no external check. We present an epistemic accountability framework addressing three failure modes specific to agentic financial AI: (1) self-promotion, where agents bootstrap authority by validating their own outputs; (2) trust inflation, where multiple weak signals aggregate into spurious confidence; and (3) evidence staleness, where decisions persist after their supporting data expires. Our central contribution is the Transformer Mandate: a structural constraint requiring that no agent may promote its own epistemic status, enforced at the data model level rather than through prompt instructions. We formalize conservative aggregation via the Godel t-norm with five mathematical invariants guaranteeing that no conclusion exceeds its weakest supporting evidence, and specify formality-dependent evidence validity periods calibrated to financial market dynamics. A retrospective audit found 20-25% of architectural decisions had stale evidence within two months. We discuss alignment with financial regulation including the EU AI Act's auditability requirements and MiFID II algorithmic trading obligations.