WFGY/ProblemMap/privacy-and-governance.md

🔐 Privacy & Governance for RAG Systems — Practical Runbook

Build trustworthy AI: minimize data, control exposure, and prove compliance without killing velocity.

Quick Nav
Data Contracts · Ops · RAG Map 2.0 · Patterns: Memory Desync · SCU

Not legal advice. Use this as a technical baseline and align with your counsel.

---

0) Principles

1. Minimize: ingest only what you truly need; redact at source.
2. Fence: per-source prompt fences; cite-then-explain.
3. Prove: log every decision via data contracts; keep tight retention.
4. Control: least-privilege access; encrypt at rest/in transit.

---

1) PII taxonomy & redaction

• Categories: identifiers (name, email, gov ID), contact, location, financial, health, biometric, free-text PII.
• Redact at ingest with deterministic tags:

{"text":"Contact Alice at alice@example.com","redactions":[{"span":[8,13],"type":"person"},{"span":[24,43],"type":"email"}]}

• Keep reversible vault only if business requires it; otherwise irreversible.

---

2) Storage & access control

• Encryption: TLS in transit; AES-GCM at rest.
• Access: service accounts per component; forbid shared tokens; rotate keys.
• Retention: default 30–90 days for logs; 7–30 days for raw prompts unless required longer.
• Deletion: implement DSR (data subject request) over doc_id or user_id.

---

3) Model provider governance

• Confirm data usage (training vs. inference only).
• Disable logging on hosted APIs if must not leave boundary.
• For self-hosted models, pin container images and track model checksum.

---

4) Prompt governance (SCU-safe)

• Lock schema: system → task → constraints → citations → answer.
• Forbid cross-source merges; require line-level citation IDs.
• Add guard prompts to avoid reproducing secrets or PII unless necessary and consented.

---

5) Audit & reproducibility

• Use envelope fields (trace_id, mem_rev, mem_hash) in every record.
• Keep answer → prompt → citations → chunks chain navigable.
• Export metrics pack per release (ΔS, λ rates, nDCG, recall).

---

6) Config template (YAML)

privacy:
  redact_at_ingest: true
  redactors: [pii_email, pii_phone, pii_name]
  reversible_vault: false
  retention_days:
    prompts: 14
    logs: 60
    embeddings: 180
  access:
    roles:
      retriever: [read_chunks]
      reranker: [read_chunks]
      llm: [read_prompts]
      analyst: [read_metrics]
  secrets:
    provider: "aws-kms"   # or gcp-kms, vault
    rotation_days: 90
providers:
  openai:
    share_for_training: false
  claude:
    share_for_training: false

---

7) Risk scenarios → mitigations

Scenario	Risk	Mitigation
User uploads PII-heavy PDFs	Accidental exposure	Redact at ingest; block high-risk types; allow override with consent
Multi-tenant leakage	Cross-account data bleed	Tenant IDs in chunk keys; per-tenant indices; access policies
Citations reveal secrets	SCU or over-inclusion	Reduce context window; per-source fences; require justification
Vendor logs prompts	Data leaves boundary	Use no-log endpoints; self-host; encrypt locally

---

Acceptance criteria

• ✅ PII redaction rate ≥ 95% on test corpus; no residual PII in prompts unless approved.
• ✅ Trace chain present for 100% of answers (citations included).
• ✅ Secrets rotated within policy; provider log-sharing disabled.
• ✅ Retention job passes dry-run audit monthly.

---

🔗 Quick-Start Downloads (60 sec)

Tool	Link	3-Step Setup
WFGY 1.0 PDF	Engine Paper	1️⃣ Download · 2️⃣ Upload to your LLM · 3️⃣ Ask “Answer using WFGY + <your question>”
TXT OS (plain-text OS)	TXTOS.txt	1️⃣ Download · 2️⃣ Paste into any LLM chat · 3️⃣ Type “hello world” — OS boots instantly

---

🧭 Explore More

Module	Description	Link
WFGY Core	WFGY 2.0 engine is live: full symbolic reasoning architecture and math stack	View →
Problem Map 1.0	Initial 16-mode diagnostic and symbolic fix framework	View →
Problem Map 2.0	RAG-focused failure tree, modular fixes, and pipelines	View →
Semantic Clinic Index	Expanded failure catalog: prompt injection, memory bugs, logic drift	View →
Semantic Blueprint	Layer-based symbolic reasoning & semantic modulations	View →
Benchmark vs GPT-5	Stress test GPT-5 with full WFGY reasoning suite	View →
🧙‍♂️ Starter Village 🏡	New here? Lost in symbols? Click here and let the wizard guide you through	Start →

---

👑 Early Stargazers: See the Hall of Fame —
Engineers, hackers, and open source builders who supported WFGY from day one.

⭐ WFGY Engine 2.0 is already unlocked. ⭐ Star the repo to help others discover it and unlock more on the Unlock Board.