Confidence Score: How to Read It and Set Thresholds

Every mostQI reading comes with a score. For each document and each field, it helps you decide whether the data can flow automatically into your system or needs a new capture or human review.

What the score is

mostQI's neural networks produce probabilistic outputs. The score turns those outputs into a number from 0 to 1:

For the document, it reflects the quality of the image and of the reading as a whole.

For a field, it reflects the reliability of that specific reading.

The higher the score, the more reliable the reading.

What the score is not:

It is not an accuracy percentage. A field with a score of 0.9 does not have a "90% chance" of being correct.

It is not linear. The difference between 0.4 and 0.5 is not equivalent to the difference between 0.8 and 0.9.

It is not comparable across endpoints. The extraction score, the Facematch similarity and the Liveness score measure different things.

Where the score appears

Endpoint	Field	Scale	Reference threshold
Content Extraction, MultiPage, Generative, Vio	document score and fields[].score	0 to 1	We recommend using 0.5 as the initial reference value in the POC. The ideal threshold should be defined after testing the service with a sample that represents your company's real documents, taking into account the requirements and particularities of your operation.
Content Classification	classifications[].score	0 to 1	We recommend using 0.5 as the initial reference value in the POC. The ideal threshold should be defined after testing the service with a sample that represents your company's real documents, taking into account the requirements and particularities of your operation.
Content Fullreader	score per page	0 to 1	The image quality standard is based on a reference score of 0.5 or higher. Low-quality (pixelated) images can lower the score.
Content Compare	fields[].score (similarity to the value sent)	0 to 1	We recommend using 0.5 as the initial reference value in the POC. The ideal threshold should be defined after testing the service with a sample that represents your company's real documents, taking into account the requirements and particularities of your operation.
Facematch 1:1 and SearchFaceImage	similarity	0 to 1	0.5 for enrollment or access
AddFaces	face quality score	0 to 1	0.75 to store the face
Liveness Streaming	livenessScore	0 to 1	We recommend considering scores above 0.8 to pass the liveness check.

Threshold, errors and review

The threshold is the minimum score you accept without review. Choosing a threshold means choosing a balance:

Higher threshold: fewer errors get through, but more items go to review.

Lower threshold: more items flow automatically, but more errors get through.

In technical terms, a stricter threshold increases precision (fewer false positives) and lowers recall (more false negatives).

Illustrative example, with 1,000 fields read in a POC:

Threshold	Flow automatically	Go to review	Errors that pass without review
0.3	940	60	12
0.5	880	120	5
0.7	750	250	2

The numbers above are hypothetical and only show the direction of the effect. Actual behavior depends on your documents.

How to choose the threshold in the POC

Start with 0.5. Treat fields scoring below it as candidates for review.

Use real documents. Send a representative sample of your operation, with the quality variations you receive day to day.

Check the results. Compare the extracted values with the correct values and note, for each score range, how many were right.

Adjust per field. Critical fields, such as CPF or amount, may require a higher threshold. Supporting fields may accept a lower one.

Measure the review volume. If it is significant, plan to use mostValid (see below).

Low score: new capture or review?

When the document score is low, the problem is usually the image. mostQI cannot recover a poor image: ask for a new capture and give the user these guidelines:

remove plastic sleeves or protective covers;

avoid folds and creases;

use good lighting, without shadows or glare;

use a neutral background and a steady, focused camera;

leave a margin around the document, with no fingers or objects covering information.

When the document is fine but a field has a low score, or when a new capture is not possible, the way forward is human review.

Human review with mostValid

With mostValid, you send the document through mostValid's own endpoints. It performs the extraction and opens review tasks according to the rules defined for your operation. What the operator corrects is analyzed by MOST's continuous improvement team, which can create new SmartDocs and reduce review in future operations. See all the details about mostValid here.

Frequently asked questions

Can I use 0.5 in production? It can be a good start, but your production threshold should come from your POC results.

Is a field with a score of 1 always correct? No. A high score indicates high reliability, not a guarantee.

Can I compare scores from different endpoints? No. Each endpoint has its own model and its own confidence scale.

Does the order of the fields matter? No. Always locate fields by name.

Confidence score

What the score is

Where the score appears

Threshold, errors and review

How to choose the threshold in the POC

Low score: new capture or review?

Human review with mostValid

Frequently asked questions

See also

Confidence score

What the score is#

Where the score appears#

Threshold, errors and review#

How to choose the threshold in the POC#

Low score: new capture or review?#

Human review with mostValid#

Frequently asked questions#

See also#

What the score is

Where the score appears

Threshold, errors and review

How to choose the threshold in the POC

Low score: new capture or review?

Human review with mostValid

Frequently asked questions

See also