Local Sandbox Safety Review Packet Template

Static safety boundary

• No live execution
• No model training
• No runtime crawling
• No private-network probing
• No credential validation
• No safety-certification claims

Reviewer checklist

• Name the local sandbox pattern and components.
• Confirm no arbitrary code execution is introduced.
• Confirm no private-network probing or credential validation.
• Confirm no tunnel or webhook replay.
• Record planner/safety/executor/narrator responsibilities.
• State that local execution reduces some dependencies but does not remove all risk.

Example record

{
  "planner": "produces JSON plan",
  "safety": "validates schema and workspace boundaries",
  "executor": "dispatches only validated simulator actions",
  "narrator": "summarizes status without safety-critical authority"
}

QA gates

• Public Output: safe static template only
• Ontology: agent roles must be scoped
• Evidence: runtime capability claims require implementation proof
• Human Review: required for operational claims
• Resource Closure: local costs must be visible
• Auditability: safety decision path must be inspectable

Claim boundaries

• not proven
• not certified
• not deployment-safety evidence
• not consciousness
• not biological equivalence
• not lossless glyph conversion
• no hidden semantic authority
• requires human review for widened claims
• no live execution
• no model training
• no private-network probing
• no credential validation
• no safety-certification claim