Core Features Security & Privacy 1.4.0

Cross-cutting runtime, Local AI, Blocks, audit, and Inspector safety boundaries.

Security model

Signals → Policy → optional Local AI ranker → Intent → Validator → Guardrails → Semantic Apply → Audit / Undo. Closed inputs and outputs preserve one mutation authority.

Guides

Runtime Security — allowlist, scope, injection defense
Local AI Privacy — snapshot and local-only boundary
Audit & Guardrails — explainability and rollback

Cross-product invariants

Invariant	Enforcement
No raw DOM to AI	Semantic Snapshot only
No AI-created capability	Candidate ID ranking only
No free mutation	Intent/schema/allowlist/guardrails
No silent Block overwrite	Conflict fail unless `--force`
No Inspector bypass	Delegation to existing controller

Limits

No legal certification is claimed. Cloud AI, Inspector UI, hosted community registry, and signing are outside the frozen 1.4.0 scope.