safety/private-public-split

Private/Public Split

Vault content is sovereign. What goes public is the user's explicit choice, not a default.

When This Skill Activates

• Vault content is being prepared for export to a platform adapter
• A Claw is generating a public artifact from vault material
• Keywords: "export", "publish", "public vault", "share memory", "platform pack"
• Default for: Starlight Sentinel

What This Skill Does

Enforces a clear boundary between private vault content (stays local) and public-safe content (can be exported to platform adapters, shared contexts, or the public vault). Prevents personal, financial, medical, and relationship data from silently flowing into shared AI contexts.

Procedures

Procedure 1: Entry Classification

Before any vault export, classify each entry:

PRIVATE (never exported without explicit approval):
  - Personal names, addresses, family relationships
  - Financial specifics: account numbers, exact valuations, salary details
  - Medical, health, or therapy references
  - Legal matters
  - Relationship conflicts or sensitive personal dynamics
  - Entries explicitly tagged: { privacy: "private" }
  - All Horizon Vault entries (require PR-level review)

PUBLIC_SAFE (exported by default):
  - Frameworks and mental models (no personal attribution required)
  - Technical architecture patterns
  - Strategic decisions: goals, priorities, opportunity assessments (non-sensitive)
  - Creative concepts, voice elements
  - Professional methodology and principles
  - Entries explicitly tagged: { privacy: "public" }

Procedure 2: Export Filtering

1. Load all entries from target vault(s)
2. Classify each entry as PRIVATE or PUBLIC_SAFE
3. Apply token budget to PUBLIC_SAFE entries (newest + highest confidence first)
4. Present summary to user: "N entries included, M entries withheld as private"
5. User can review the withheld list and opt-in specific entries
6. Write filtered export to target path

Procedure 3: Tagging Protocol

Assist the user in tagging vault entries for future automatic filtering:

Tags added to vault entries:
  { "privacy": "private" }    — never export
  { "privacy": "public" }     — always safe to export
  { "privacy": "review" }     — ask before each export

Default for untagged entries: classified by content heuristic (Procedure 1).

Integration Points

• Vaults: Reads all vaults; writes privacy tags back to entries (via Memory Claw)
• Agents: Starlight Sentinel (primary), Starlight Sage (vault guardian)
• Skills: Composes with secret-detector (structural classification + credential scan), permission-gate (approval flow)
• Claws: Memory Claw (/sis-export), Creator Claw (content generation), Attestation Claw (public vault)

Quality Criteria

• No private entry is exported without explicit user opt-in
• Export summary always shows what was withheld and why
• Privacy tags are preserved across Memory Claw operations (not stripped on reconcile)
• Horizon Vault entries are always withheld by default