# Rows that file themselves: you confirm what was only guessed

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

Your rules go first, then your own past choices, a merchant dictionary and recurring payments. AI is asked last, about merchants nothing else could name.

## How it works

- **Five steps, in a fixed order.** Your rules, then a payee you filed yourself under the same name, then the built-in dictionary, marked known merchant, then recurring payments. AI suggestions come last.
- **AI per merchant, not per row.** The processor's tail is cut, so FACEBK *M6A2… and FACEBK *NKGQ… are one merchant. The model hears only the largest ones nobody named; its rows carry an AI chip, and one press takes the batch back.
- **Correct once, or stop it.** Correcting a guess offers a rule for that payee, which files its other rows too. Stop auto-categorizing this payee silences the guessing and files nothing.
- **Wallet transfers by address.** Place a wallet address once with the rule tick on, and the next transfer to that address arrives filed. The model is never asked about addresses.

## Questions

**What does the AI model see?** Only the largest merchants nothing else could name, and never a wallet address. With AI switched off for the space, your rules and the dictionary still file.

**What if a rule catches the wrong rows?** Undo it on the past rows it touched. Rows a person confirmed stay as they were set, and every row the app filed itself carries a mark naming the step.