ChangeSafe: An AI Infrastructure Change Airlock

A v0.1 safety boundary that converts AI-generated infrastructure recommendations into typed, schema-validated, transactionally applied patches with deterministic policy checks and hash-bound decision receipts.

Independent project, 2026.

ChangeSafe asks a narrow question: before an AI-generated infrastructure change is applied, how do you prove it is safe to apply?

The v0.1 answer is an airlock. AI-generated recommendations are converted into typed declarative patches, validated against strict schemas and permitted state paths, and applied transactionally against a cloned infrastructure model — never directly against production.

On top of that boundary sit deterministic policy checks for management reachability, protected resources, blast radius, rollback completeness, and required verification. Fail-closed domain states are enforced, inverse rollback is tested, and every decision emits a hash-bound receipt so the outcome is auditable after the fact.

Production execution remains intentionally disabled in v0.1 — the point of this version is the safety boundary itself, not the automation it would eventually gate.