# Who checks the machine: and what they found

URL: https://www.orla.finance/en/product/autopilot/supervision
Markdown twin of that page. Append `.md` to any Orla page URL to get one.

The machine's record, kept by the people who correct it. A rule here can only narrow what the machine may do, never widen it.

## How it works

- **Every correction counted.** A changed amount, date, account or rail, a decline, or work done by hand past a proposal. Each counts against that kind of work.
- **The share as it came.** Per kind of work, how much was pressed without a change. A pilot ends at thirty days, twenty decisions and 95%, and then its owner decides.
- **Two mistakes, a new pilot.** Two in thirty days send the kind back to your button and start a new pilot.
- **The supervisor narrows only.** Its rules can pause the autopilot, freeze an agent or raise an exception. Code decides, the model explains, and a monthly report lists every exception.

## Questions

**Does an AI decide what gets flagged?** No. Code checks every act against the mandate and its rules, and the model only explains in words. The supervisor can stop things, never allow more.

**Can a kind of work stay at my button for good?** Yes. Set it to by hand only and it stops proposing at once. Turning it back on loosens control, so it waits like any other loosening.