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

# Confidence scoring

> How the registry decides how much to trust each discovered asset.

Every AI asset carries a `confidence` value saying how certain the system is that the asset exists and is described accurately. It is **computed automatically** from which channels detected it — never set by hand.

## The four levels

| Level | Criteria | What it means |
| - | - | - |
| `verified` | **Two or more** independent sources | Highest certainty — corroborated across channels |
| `high` | A single detection from the **SDK** or **gateway** | Verified at code level; the call demonstrably happened |
| `medium` | A single detection from a **browser extension**, **desktop agent**, or **platform signal** | Observed once by automated detection |
| `low` | Anything else, including self-reported survey data | Claimed, not technically verified |

## How it is computed

```
        How many sources?
               │
        ┌──────┴──────┐
       2+             1
        │             │
    verified    Which source?
                      │
        ┌─────────────┼──────────────┐
   sdk / gateway   browser /       other
        │          desktop /         │
      high         platform         low
                      │
                   medium
```

Source count dominates. A second independent channel promotes an asset to `verified` regardless of which channels they were — because agreement between two blind observers is stronger evidence than either one alone.

## Why multi-source matters

A survey response saying *"yes, my team uses ChatGPT"* is weak evidence. Someone may be guessing, or describing a trial from six months ago. Watch confidence climb as evidence accumulates:

<Steps>
  <Step title="Survey reports HR uses ChatGPT">
    One source, self-reported → `low`
  </Step>

  <Step title="Browser extension detects ChatGPT in HR">
    Now two independent sources → `verified`
  </Step>

  <Step title="SDK captures GPT-4o calls tagged business_unit=HR">
    Three sources, one of them code-level → stays `verified`, with a richer evidence trail
  </Step>
</Steps>

## Using it

Treat confidence as a triage tool rather than a scoreboard.

* **`verified`** — safe to put in front of an auditor or an EA tool.
* **`high`** — real, but seen through one lens; the business context may be thin.
* **`medium`** — worth a look; often shadow AI nobody has registered.
* **`low`** — a lead, not a finding. Confirm before acting.

<Note>
  A large `low`-confidence population is not a defect. It usually means your survey coverage runs ahead of your technical coverage — the fix is deploying another channel, which promotes the real ones and leaves the phantoms behind.
</Note>

<Card title="Where the sources come from" icon="radar" href="/developers/discovery-sources">
  The channels feeding this calculation, and what each one can and cannot see.
</Card>


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