Bank vendor onboarding · EvidenceBound

Onboard a vendor without giving the AI the final say.

The AI handles routine vendor work. A person approves decisions that carry real compliance or financial risk. If the evidence changes later, affected work is undone and payment stays blocked.

Real Strands SDK Agent/tool loop. Human authority stays outside the model toolset.
1

AI does the routine work

Vendor packet, tax and bank checks, draft setup and follow-up happen before the human is interrupted.

2

A human decides what carries risk

When vendor activation and payment setup become consequential, the bank employee provides the required authority.

3

Evidence changes

The regulatory evidence gate moves from PASS to HOLD. Stale authority stops working: affected setup is undone, unrelated safe work remains, and first payment cannot proceed.

Technical evidence

Simple banking behavior, backed by explicit control boundaries.

Strands is load-bearingThe public judge path executes a real Strands Agent/tool loop with three model-callable operational tools.
The model cannot approve itselfApprove/revoke authority mutation tools are absent from the Strands tool schema. Human authority is external.
BANK-ECP changes authorityA sanitized DORA vendor-responsibility bridge moves PASS → HOLD and invalidates the stale vendor-risk approval.
Truth boundary. This public page uses a deterministic custom Strands Model provider so anyone can reproduce the judge path without credentials. It proves live Strands orchestration and authority separation. This request does not claim a foundation-model or AgentCore invocation. The BANK-ECP bridge shown here is DEVELOPMENT_ONLY and is not formal benchmark-performance evidence.
Show execution diagnostics
Run the banking scenario to populate the execution ledger.