Reversible / Irreversible
One question, applied relentlessly: can this be undone? It is the cleanest line I know between what to automate and what to gate.
State precisely what the system is about to do — not the category, the specific act.
If it can be cleanly reversed, automate it. Speed has almost no downside here.
Anything irreversible or public routes to a human. A pause always costs less than a retraction.
Every gate decision is recorded, so the boundary itself can be reviewed and refined.
Worked example — the framework applied to a real decision, not a hypothetical.
These companion resources are still being written — nothing here is downloadable yet. Everything the framework needs to be run by hand is already on the Steps and Prompt tabs. Want to be told when they land? Say so here.
The test earns its keep the first time it stops something. Applied consistently, the debate about whether a given action "needs approval" disappears — the question is only ever whether it can be undone, and that has one answer.