Relativity6Platform Docs

Confidence scores

Each prediction includes an accuracy score to help you automate straight-through processing or send cases to review.

What accuracy means

Every prediction includes accuracy: a score between 0.10 and 0.95 that reflects how confident the classification is for that code.

RangeTypical interpretation
0.75 – 0.95Strong match — reasonable to auto-accept when the rest of your file agrees
0.50 – 0.74Moderate — worth a rules check, a second source, or human review
0.10 – 0.49Weak — treat as uncertain and review before binding

Use accuracy alongside the code and title, not on its own. A high score on the wrong industry is still the wrong industry.

Primary and alternate predictions

When you request two predictions (the default), you receive:

  • prediction_a — our primary classification for the business
  • prediction_b — a plausible alternate when the business could fit more than one industry

Compare the two accuracy values to see how clear-cut the answer is:

Gap between primary and alternateSuggested handling
Large (about 0.15 or more)The primary code is usually the right place to start
Moderate (about 0.05 – 0.14)Show both codes to an underwriter
Small (under 0.05)The business may be ambiguous — review or gather more detail

The codes can differ even when scores are close (for example, full-service restaurant vs. limited-service). Always display title with code so reviewers understand the distinction.

Using scores in your workflow

Straight-through processing

Use a primary accuracy threshold and a minimum gap before auto-selecting prediction_a. Many teams start around 0.80 accuracy with a 0.15 gap, then tune from production results.

Existence check

If you enable includeExistenceCheck, treat very low accuracy together with exists: false as a stronger signal to review the entity itself, not just the code.

Your own data

If you already trust a website, prior policy class, or application narrative, weigh disagreements heavily even when accuracy looks high.

When scores are low

Very low accuracy usually means the service could not classify the business with confidence — for example, sparse inputs, conflicting signals, or an unusual operation. It is not a substitute for a careful manual pick.

To improve results:

  • Send name plus state or a full address when you have them
  • Add a clear description when the name alone is generic
  • Use description-only mode only when your description is detailed and trustworthy

Include responseId in support requests if many records cluster at the low end of the range.