Banking Shield: Machine-Checked Absence Claims for Privacy-Sensitive AI Explanations

Banking Shield is an author-prepared research preprint presenting a fictional, non-bank, research-only banking-adjacent prototype for machine-checked absence claims in privacy-sensitive AI-style explanations. The paper investigates whether a sensitive-domain prototype can produce useful explanation artifacts while generating machine-checkable evidence about what the explanation layer did not receive, did not transmit, and did not claim. The prototype combines metadata-only integrity evidence, fail-closed forbidden-field filtering, tamper-evident HMAC audit chains, withdrawal semantics, and a deterministic offline AI-style privacy firewall with input allowlisting, output claim control, no-egress checks, and per-response privacy receipts. At the evidence freeze, the implementation passed 417/417 unit tests, 43/43 full banking end-to-end checks, 27/27 security checks, three privacy audits, an AI firewall smoke test, and a no-egress static gate. A formative internal dry run with five trusted internal testers produced 30 sessions and aggregate-only evidence. This manuscript does not claim fraud detection, scam prevention, payment safety, payee verification, financial advice, regulatory compliance, production readiness, bank-grade security, live LLM validation, or real-world banking effectiveness. It is not peer reviewed and has not undergone external banking review or independent security validation. Supplementary preprint v1.2; main Project Simurgh baseline: v0.4.18;