Create PHASE4_GOVERNANCE_POLICY.md

PHASE4_GOVERNANCE_POLICY.md

@@ -0,0 +1,82 @@
+# Phase-4 Governance Policy: Semantic Search (FAISS)
+
+## Purpose
+
+Phase-4 introduces **optional semantic search capabilities** using FAISS to
+enhance discovery across **metadata only** associated with publicly released
+FOIA records.
+
+This policy governs whether, how, and under what constraints Phase-4 may be
+enabled.
+
+---
+
+## Scope of Phase-4
+
+Phase-4 MAY include:
+- Vector embeddings of **metadata fields only** (title, agency, date, citation)
+- User-initiated semantic similarity queries
+- In-memory or user-controlled vector stores
+
+Phase-4 MUST NOT include:
+- Full-text document embeddings without explicit review
+- Automated crawling or indexing
+- Cross-user persistence
+- Third-party model training on user data
+- Background ingestion or scheduled jobs
+
+---
+
+## Activation Requirements (ALL REQUIRED)
+
+Phase-4 functionality remains **hard-disabled by default**.
+
+Activation requires:
+1. Legal review approval
+2. Hugging Face Trust & Safety concurrence
+3. Explicit UI opt-in from the user
+4. Clear disclosure of embedding scope and limits
+5. Feature flag activation by maintainers
+
+---
+
+## Data Handling Rules
+
+- No raw PDF content stored by default
+- No embeddings persisted beyond session unless user exports
+- No cross-session correlation
+- No private or sensitive data permitted
+
+---
+
+## Transparency & Auditability
+
+When enabled, Phase-4 must:
+- Log feature activation locally (user-visible)
+- Display semantic scope banner
+- Provide deterministic reproducibility options
+- Include integrity hashes for AI outputs
+
+---
+
+## Kill-Switch & Rollback
+
+- Feature flag allows immediate global disablement
+- No migration required to roll back
+- No user data loss on rollback
+
+---
+
+## Governance Review Cadence
+
+- Initial approval: One-time
+- Re-review required for:
+  - New data sources
+  - New embedding models
+  - Persistent storage changes
+
+---
+
+## Guiding Principle
+
+> Semantic discovery must never compromise transparency, provenance, or user consent.