> ## Documentation Index
> Fetch the complete documentation index at: https://help.alk.work/llms.txt
> Use this file to discover all available pages before exploring further.

# Confidence Levels

> How ALK rates its own confidence in an extracted value, and how to use High, Medium, and Low to prioritise review

Every value ALK extracts from a document carries a confidence level: **High**, **Medium**, or **Low**. Confidence is ALK's own estimate of how much corroborating evidence it found for a value — it is a way to prioritise your review, not a guarantee of correctness.

<Note>
  Confidence tells you **where to look first**. A High-confidence value can still be wrong, and a Low-confidence value is often right. Use confidence to decide the order and depth of review, not to skip review entirely on regulated or high-value documents.
</Note>

## The three levels

| Level      | What it means                                                                                                                                                                                                                                       | What to do                                                                                                              |
| ---------- | --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | ----------------------------------------------------------------------------------------------------------------------- |
| **High**   | Multiple independent signals agree. ALK found the value in the document text, the quoted source matched a real region of the page, the value is in the right format for the field, and the underlying text was cleanly read.                        | Spot-check. Review a sample rather than every row.                                                                      |
| **Medium** | The value is supported, but at least one signal is weak or missing — for example the source quote could not be located exactly, the page was scanned rather than native text, or the value appears on only one page with nothing to corroborate it. | Review before relying on it. Open the field in Block view and check the highlighted source region against the document. |
| **Low**    | Little or no corroboration. The value may be malformed for its field type, may not appear in the document text at all, may come from a poorly-read region, or different parts of the document may disagree about it.                                | Always review. Treat the value as a suggestion and confirm it against the document before verifying or exporting.       |

In the interface, the levels are colour-coded consistently:

* **High** — green
* **Medium** — amber
* **Low** — red, with a pulsing dot to draw attention

## Where you see confidence

<Steps>
  <Step title="Table view">
    Each cell shows a small coloured dot next to the value. Scan a column for red dots to find the fields that need attention.
  </Step>

  <Step title="Block view">
    Each field card shows the same dot. Hover it to see the level, and open the field to see ALK's reasoning for the value.
  </Step>

  <Step title="Review modal">
    Opening a field shows a **High / Medium / Low Confidence** badge, ALK's reasoning, the quoted source text, and a highlight over the region of the document the value was read from.
  </Step>
</Steps>

### Review status replaces confidence

Once a person has acted on a field, the reviewer's judgement takes over from ALK's:

* **Verified** (check icon), **Rejected** (X icon), and **Edited** (pencil icon) replace the confidence dot entirely.
* The confidence dot only appears on fields that no one has reviewed yet.

This is deliberate — once you have hand-corrected or confirmed a value, ALK's confidence in its own original extraction no longer describes what is on screen. See [Reviewing extracted fields](/workflows/setting-up-a-workflow#reviewing-extracted-fields) for the full list of review actions.

### When the source text was hard to read

Separately from the field's own level, opening a field shows an amber warning when the part of the page a value was read from was itself poorly scanned, asking you to check the value against the document.

It usually means a low-resolution scan, a skewed page, a stamp or handwriting over the text, or a fax. The value may still be correct, but the underlying characters were uncertain, so it is worth reading the highlighted region yourself before verifying the field.

## How the level is calculated

ALK does not take the model's word for it. The level is a composite of five independent checks. Text recognition quality, format validity, and the model's own self-assessment carry the most weight; cross-page consistency and source verification refine the result.

* **Text recognition quality** — the recognition score of the specific regions the value came from. A crisp native PDF scores high; a poor scan scores low.
* **Format validity** — whether the value fits the field type: a real date for a Date field, a parseable number for a Number field, a match for any validation pattern you set, and no signs of garbled text.
* **Model self-assessment** — how confident the extraction model reported itself to be. Counted, but never on its own.
* **Cross-page consistency** — whether the value appears on more than one page. Found on two or more pages counts in its favour; found on exactly one page is neutral; conflicting variants elsewhere in the document count against it.
* **Source verification** — ALK searches the document text for the quote the model claims to have read the value from, independently of what the model reported, and checks that the quote is genuinely there.

<Note>
  The exact weighting and cut-off points are tuned over time as extraction improves, so treat the levels as a review-prioritisation aid rather than a fixed score you can calculate yourself.
</Note>

<Tip>
  Setting a clear **field type** and, where relevant, a validation pattern is the single most effective thing you can do to improve confidence scoring. Format validity is one of the heaviest checks, and it can only be assessed against a known expected format. See [Editing tips](/workflows/setting-up-a-workflow#editing-tips).
</Tip>

### Long documents and collections

Long documents are processed in sections, and a **collection** is processed as a bundle of related documents. In both cases ALK produces several candidate answers for a field and merges them:

* If only one section returned a value, that value and its confidence are used as-is.
* If several sections returned the **same** value, the strongest candidate's confidence is kept and every supporting source is recorded.
* If sections returned **different** values, the highest-confidence candidate wins.
* If two candidates disagree at the **same** confidence level, ALK treats it as a genuine conflict and resolves it explicitly rather than picking arbitrarily.
* If no section returned a value at all, the field is marked **Low**.

This is why a field can come back Low on a long document even when the value looks obvious — two parts of the document disagreed.

## Confidence in Reconciliation

Reconciliation uses a different scale. Exceptions on the **Exceptions** page carry a numeric match confidence expressed as a percentage rather than a High/Medium/Low level, shown next to each suggested bank-to-General-Ledger pairing along with the reasoning behind it. Percentages of 80% and above are shown in green; anything lower is shown in amber for review. See the [Reconciliation Guide](/workflows/reconciliation-guide).

## Suggested review policy

A practical starting point for most teams:

1. **Low** — review every field, every time. Verify, edit, or reject each one.
2. **Medium** — review every field on documents that carry financial, legal, or compliance consequences; sample the rest.
3. **High** — sample. Review a fixed percentage of rows, and always review fields you know are hard to read in your document set (handwriting, stamps, signatures, totals in tables).
4. Re-run extraction on individual fields with **Retry** before hand-correcting them — a targeted re-run often resolves a Low-confidence field on its own.

## Related articles

* [Setting Up a Workflow](/workflows/setting-up-a-workflow)
* [Creating a Custom Workflow](/workflows/custom-workflow)
* [Document Classification Workflow](/workflows/applications/document-classification)
* [Glossary](/reference/glossary)

## Need help?

If a field's confidence looks consistently wrong for your document set, contact the Wamiri support team at [support@wamiri.com](mailto:support@wamiri.com). Include the workflow, the field name, an example document, and what you expected the value to be.