Engineer Builds Undo-First Architecture for AI Agents Before Changes Are Applied

A software engineer has spent several months developing a system that constructs and stores a verified inverse of any proposed change before that change is executed, reversing the typical decide-act-report order of AI agents. Each proposed change receives a canonical identity, its rollback operation is pre-computed and stored in escrow, and a gate issues one of three verdicts before anything is written to disk. The approach aims to make every outcome — including refusals — a signed, offline-verifiable record that does not require trusting the original issuer. However, the engineer acknowledges a critical limitation: only about 13.8% of write-capable public MCP tools expose enough information to construct a meaningful inverse, risking the system becoming little more than a refusal machine. A formal preprint is forthcoming, though the Lean model and Rust implementation are currently validated by conformance testing rather than a full refinement theorem.