Candidate Is Not Canon

AI makes it very easy to produce something that looks finished.

A slide deck renders. A plan has headings. A patch passes tests. A research brief contains sources. The artifact is polished enough that the mind quietly moves it from proposal to truth.

That is where a lot of bad work enters a system.

I find it useful to separate four states:

1. Generated: An output exists.
2. Validated: It passes the checks appropriate to its medium.
3. Accepted: A responsible human has reviewed it for the real purpose.
4. Canonical: It has replaced the previous source of truth and downstream work may rely on it.

Each transition requires different evidence.

A generated presentation can be complete and still tell the wrong story. A technically valid plan can still optimize for the wrong goal. A human may accept an artifact for one meeting without deciding that it should become the reusable company template.

So I try to label AI output as a candidate until the next gate is explicit.

This is more than careful wording. State labels shape behavior. If a file is called final, people stop asking questions. If a preview is placed in the canonical directory, later agents treat it as approved context. If a recommendation is written directly into the project plan, the act of proposing it becomes indistinguishable from the act of deciding it.

The safer pattern is visible promotion:

Candidate → technical validation → human acceptance → canonical state

Sometimes one person performs every gate in a few minutes. Sometimes the transitions happen across days and teams. The process can be lightweight. The separation should remain.

This matters most when AI increases the volume and polish of candidate material. The scarce resource is no longer producing a plausible version. It is deciding what earns the right to shape future work.

Candidates are cheap. Canon carries consequences.