A field-by-field score card: what the automatic reading answered, on how many entries, and how often a person kept that answer or changed it.
When to use it
When you want to know how far you can trust the reading before you rely on it, or which fields are still being corrected by hand.
Before you start
The AI document add-on has to be switched on for your workspace, and you need viewing rights on this screen. Fills must already have been run from the Document Pipeline. Until then the screen says the ledger is empty, so you never see a misleading score.
Where to find it
AI > Accuracy Ledger.
How it works
- There is one line per field, most used first. Each line shows how many answers the field has given, across how many entries, the average of how sure the reading was, and then Accepted, Corrected and Pending.
- A field starts as Pending the moment a fill writes it. It has been answered, but nobody has judged it yet.
- You make the judgment without doing anything extra. Open one of those entries, check it and save it. Every value you left as it was counts as Accepted. Every value you changed counts as Corrected, with your figure recorded beside the original.
- A lot of Pending only means those entries have not been opened and saved yet. It is not a sign that anything is wrong.
- This screen changes nothing. The Autonomy screen reads it as evidence when deciding whether a field has earned more independence.
Tips and common mistakes
- Correcting an entry is not a complaint; it is how the system learns what it gets wrong. The real damage comes from saving a wrong figure unchanged, because that is recorded as agreement.
- The average of how sure the reading was is not accuracy. Accuracy is the Accepted share, and that only exists once people have reviewed and saved.
Related guides
See also Autonomy, Document Pipeline and Backfill Runs.