Abstract
Research question
Where should detection, redaction and encrypted-original handling occur before retrieval or generation?
Context
A prompt instruction to ignore personal data does not remove that data from provider requests or logs. Context should be prepared before it enters retrieval and generation paths.
Proposed approach
- Detect candidate personal information at explicit input and output boundaries.
- Redact provider-facing context according to application policy.
- Store encrypted originals only when the use case requires them.
Implementation
- Configurable local PII preparation in a Spring Boot demo
- Redacted context and audit output visible side by side
- A no-remote-model mode for deterministic inspection
Observations
- 01
Visible preparation stages are easier to test than prompt-only privacy claims.
- 02
Local processing supports useful demonstrations without sending sample identifiers to a model provider.
Limitations
This work reports an implementation-linked investigation. It is not a
peer-reviewed paper, certification or customer case study.
- Pattern-based detection does not cover every identifier.
- The implementation is a technical control, not a compliance determination.
Reproduce or inspect
Revision note and next questions
Last revised 2026-07-25 against AI Fabric 0.4.0 and the linked public artifacts.
- How should organization-specific detectors be composed and evaluated?
- What retention policies should apply to encrypted originals?
