Irreversible Actions and Reversible Actions
In the previous chapter, we introduced the idea that actions which cannot be undone require a two-person check before execution. So what exactly does "irreversible" mean?
If we keep using the word without pinning it down, the question "what does irreversible actually mean in practice?" stays unanswered. This chapter addresses the distinction between reversible and irreversible actions head-on. Once the definitions are clear, you will know where careful procedures are needed and where a lighter touch is fine.
The Difference Between Reversible and Irreversible
A reversible action is one you can undo.
For example: saving a draft, changing a setting in a local environment, or rewriting and overwriting a manuscript. All of these can be recovered — you can undo the change, rewrite it again, or restore the setting.
An irreversible action (an action that cannot be undone or is very costly to undo) is one you cannot take back after it runs.
For example: publishing an article on the web, deleting data, sending a request to an external service, or sending an email. After any of these run, you cannot say "pretend that never happened." The article may already have been read. Deleted data will not come back without a backup. The email has already arrived in someone's inbox.
"Can you undo it, or can you not?" That is the line between reversible and irreversible.
Where the Line Falls in Practice
Here is a more concrete breakdown.
Examples of reversible actions: saving a file locally, editing an article that is still in draft, rewriting a configuration file (when a backup exists), and checking behavior in a staging environment (a test environment that is not yet public). If something goes wrong with any of these, you can redo or roll back the change. The impact stays within reach.
Examples of irreversible actions: publishing an article to the outside world, deleting data from a production environment, sending a request to an external API (a connection point for exchanging data with an outside service), and broadcasting a notice or notification to an unspecified audience. After any of these run, the impact moves outside your control. Once it is outside, you cannot pull it back.
That said, this boundary is not absolute.
Publishing an article, for instance, may have very little impact if almost no one has access to it yet. If backups are in good shape, updating production data can come close to reversible in practice. Whether something is irreversible depends on how your systems are set up and how wide the impact can spread.
What matters is not completing a perfect classification table. What matters is building the habit of asking "can this be undone?" before you run anything.
How to Treat Them Differently
The reason for distinguishing reversible from irreversible is to calibrate how much procedure to apply.
Reversible actions need only a light procedure. You can work on the assumption that if something goes wrong, you will undo it and try again. You do not need a heavy confirmation step every time. Stopping lightweight tasks repeatedly just clogs the whole workflow.
Irreversible actions call for more thorough handling. The two basics are: have a second pair of eyes confirm before execution, and keep a record of the execution.
Having a second pair of eyes confirm means that no single person pushes through alone. A blind spot you did not notice yourself may surface when someone else looks. That is the point of the two-person rule.
Keeping a record means you can explain later why that decision was made. Because an irreversible action cannot be taken back after the fact, having the reasoning on record gives you somewhere to return to if a problem comes up. Without a record, the reasoning behind the decision disappears.
Light handling for reversible actions, careful confirmation and documentation for irreversible ones. This distinction is part of the structure that keeps an AI organization running stably.
In practice, most of the work is reversible: writing drafts, editing, checking things locally. These can be left to the AI side, which reduces how often a human needs to step in. Irreversible actions are relatively rare. That is precisely why you can focus your attention on them when they do arise.
If you treat everything with the same weight, your attention will not land on the decisions that actually matter. Calibrating the weight you apply is the key to keeping the system running over the long term.