The theory needed a hostile test. So I built a full Trello-style app (auth, teams, boards, lists, cards) with an AI agent playing every role under Lattice governance. And I planted a trap for my own tool. I deliberately recorded a contestable decision, hard delete, built it, and verified it green. Then mid-project I reversed it. Archive only. No physical deletion, ever.
What held: the graph named, in advance, every artifact the reversal had to touch. One requirement, three invariants, three functions, three tests. The matrix read 100% before and 100% after. The whole reversal cost an afternoon, not an archaeology dig. Honestly, better than I expected.
What broke is the note that matters. Nothing enforced the discipline. The agent versioned the BRD, superseded the old decision, retired each test paired with its replacement, all by convention. A lazier agent could’ve deleted the failing delete-tests and walked away green. Nothing would have fired.
So 0.8.1 makes intent changes governed transactions. Priced by blast radius. Classed as wording, widening, narrowing, or contradiction. Decided by a human. A narrowing spawns explicit test-retirement items, the only thing that makes deleting a test legal.
April me was asking exactly this. It just took ten weeks to hear it.