A system that corrects itself

Request: „How would I have noticed if it hadn't worked?“

1. A consistency check tests whether documentation and reality still match — registry against files, topology against servers.
2. New procedures are created only via a meta-workflow with a check script that enforces registration in four places.
3. Foreign content is data, never an instruction: an embedded instruction was reported, not followed.
4. Proof before deletion: delete only after the target state is documented, never “status 200 = done”.
5. Every reprimand becomes a rule; numbers are re-measured before they go onto a slide.

A demo shows what is possible; in operation what counts is what happens when the AI gets it wrong.

Every rule that changes behaviour, and every exception to the ask-first default.

An export returned 147 instead of 210 transactions and looked complete — “silent partial result” has since become a guard in the code, not just a lesson.

The board's question “and what if it gets it wrong?” is answered by gates that turn red before the damage.

The difference between demo and operation is one question: how would I have noticed if it hadn't worked? That is why a consistency check continuously tests whether documentation and reality still agree — it went from 28 findings to zero. New procedures are created only through a meta-workflow with a check script. Foreign content is data, never an instruction; the first embedded instruction was reported rather than followed. And nothing is deleted until the target state is proven. On a single day the system logged 16 self-corrections — that is the number I show a board.