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

# Confidence Levels

> Use overall confidence and per-input evidence for record matching, enrichment, and qualification.

Use confidence to decide whether to accept a match, send it for review, or leave a record unchanged. Check both the overall score and the evidence for the inputs your workflow requires.

## Overall Confidence

Each [Search](https://docs.orbitsearch.com/api/search/search.md) result can include `confidence`, a score from 0 to 1, and a `confidence_level`:

| Level | Score | Meaning |
| - | - | - |
| `high` | 0.75–1 | Stronger overall support for the match. |
| `medium` | 0.35 to less than 0.75 | Partial or less conclusive overall support. |
| `low` | Less than 0.35 | Limited overall support. |

These bands summarize match strength. They are not measured accuracy rates or automatic acceptance rules. You can choose a stricter numeric cutoff than the start of a band.

The snapshot's `confidence_kind` tells you whether Orbit assessed an identity match (`identity`) or requested attributes (`criteria`). Each result is scored independently; scores across results do not need to add up to 1.

## Evidence For Each Input

The API returns assessments for the criteria it evaluated in `match_summary`. This includes names, email and phone signals, employment, education, and semantic criteria.

| Field | Meaning |
| - | - |
| `matches` | Evidence supports the input. |
| `partial_matches` | Evidence partly supports the input. |
| `apparent_conflicts` | Available evidence appears to disagree with the input. |
| `unknown` | Available evidence does not establish a match or a conflict. |

Each entry identifies its `field` and `query`. Check both when you have several criteria of the same type. Semantic clauses can be normalized or expanded into employment, education, or other criteria; they do not always map one-to-one to entries with `field: "semantic"`. Separate criteria that remain semantic have their own entries and scores.

An entry can also include its own `confidence` score for that criterion. Higher values mean stronger support that the profile satisfies that input, whichever summary bucket contains the entry. This is separate from the result's overall `confidence`. Names and exact contact checks may have no criterion-level numeric score; use their assessment and matching value. An absent score does not mean zero confidence.

`evidence_strength` and `evidence_level` describe an input's contribution to the overall score. They do not replace the criterion's own `confidence`.

## Require The Attributes That Matter

A high overall score does not establish every supplied attribute. In this example, one semantic clause is supported while the other remains unknown:

```json theme={"dark"}
{
  "confidence": 0.868,
  "confidence_level": "high",
  "confidence_status": "final",
  "match_summary": {
    "matches": [
      {
        "field": "semantic",
        "query": "speaks at developer conferences",
        "outcome": "equivalent",
        "confidence": 0.96,
        "evidence_strength": 0.864,
        "evidence_level": "strong"
      }
    ],
    "partial_matches": [],
    "apparent_conflicts": [],
    "unknown": [
      {
        "field": "semantic",
        "query": "invests in healthcare startups",
        "confidence": 0.03
      }
    ]
  }
}
```

If your qualification rule requires both clauses, this result needs further review. Before automatic qualification, check that the returned criteria preserve every original requirement, including its meaning and time scope. For example, an employment match for a company does not establish that someone invested in it. A missing or differently interpreted requirement needs review.

Require each mandatory attribute to appear in `matches`. For automatic qualification of natural-language criteria, also require a numeric criterion-level `confidence` that meets your validated threshold. For example, `0.90` can be a conservative starting policy to test. Send a criterion without that assessment to review; names and exact contacts use their matching assessments and values instead.

## Choose A Policy For Your Workflow

**CRM Record Matching:** For automatic linking, start conservatively—for example, with overall confidence of at least `0.90`. Require matching identity inputs and no identity conflicts. If you supplied an email or phone to establish ownership, require the matching contact entry to have `evidence_strength: 1`. Additional profile contacts do not weaken that match.

**Enrichment With Review:** Route candidates that meet your review threshold to a person or another validation step before updating records. A `high` label can help prioritize review without authorizing a write.

**Lead Qualification:** Evaluate the required employer, role, location, and semantic clauses individually. A supported criteria match can have an overall score below `0.75`; an overall identity match can be high while a qualification remains unknown. Use the required attributes' assessments rather than applying the CRM identity cutoff to qualification searches.

For automatic identity selection, wait for a `ready` result with `confidence_status: "final"`. Require a clear lead over competing candidates; `0.15` above both the highest `confidence` among all other results and `none_of_above_confidence` is a useful starting margin. Include competing candidates whose profiles are not yet `ready`; array position does not identify the strongest competitor. `none_of_above_confidence` is `1 -` the highest identity result score. Review close or ambiguous candidates rather than selecting one arbitrarily.

Apply these thresholds in your application after Search returns. The API does not accept a minimum-confidence request parameter. Choose your cutoff according to the cost of an incorrect match and validate it on representative records from your workflow.


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.