DoctrineBase is open — a structured reference library for AI doctrine — Browse the base
DoctrineBase

Keep a human decision in the loop for irreversible actions

Automation may prepare an irreversible action, but a human crosses the line. The cost asymmetry of mistakes, not model accuracy, dictates the control.

First recorded July 17, 2026 · Last revised July 17, 2026

Rationale

Reversible actions can be delegated to a model and corrected cheaply when wrong. Irreversible actions — deleting data, sending money, publishing to the world, contacting people — cannot. The cost asymmetry, not the model’s accuracy, is what dictates the control.

The doctrine is a gate, not a veto: automation prepares the action fully (the draft, the diff, the transaction) and a human approves the crossing. This keeps nearly all of the speed while capping the worst case at the cost of one review.

The common failure is gate erosion: approvals become rubber stamps as confidence grows, until the one bad action sails through. Counter it by keeping the approval surface small and legible — a human can meaningfully review one diff, not forty at once.