> ## Documentation index
> Fetch the complete documentation index at: [llms.txt](../llms.txt)
> Use this file to discover all available pages before exploring further.

# Confidence scores and labels

How Sleuth scores every value, and what sourced, inferred, and unknown mean.

[Open a dataset](https://app.sleuth.studio)

Every value in a Sleuth dataset links to its source and has a confidence score from 0 to 100. The score tells you how much to trust a value before you use it.

## How the score is worked out

A fixed set of rules sets the score, so the same evidence always gets the same score and you can see why. Sleuth's model reads the sources and labels what each one says; the rules then score it on three things:

- **Source quality.** A label, a registry, or a filing counts for more than a news story.
- **Agreement.** Independent sources that agree raise the score.
- **Recency,** where it matters, like a trial's status or a drug's phase.

## The labels

| Score | Label |
| --- | --- |
| 85 to 100 | Confirmed |
| 70 to 84 | High |
| 50 to 69 | Moderate |
| 25 to 49 | Low |
| 0 to 24 | Unverified |

## Sourced, inferred, or unknown

- **Sourced:** a document Sleuth links to states the value.
- **Inferred:** Sleuth worked the value out from its sources and shows how.
- **Unknown:** Sleuth couldn't find a public source, so it leaves the cell open instead of filling it in.

Each dataset shows how many of its values fall in each group, and each column has its own score.

## When sources disagree

Sleuth keeps both sources, goes with the one it scores higher, caps the cell's score, and flags it so you see the conflict. In Review mode you see both side by side and can switch to the other one. Your own corrections get a score of 100.

## How Sleuth avoids made-up values

Sleuth checks every value before you see it. Its engineering matches and normalizes every record, AI checks run at each step, disagreements between sources are flagged instead of smoothed over, and Sleuth's data team tests the results. Because every value is labeled and linked, you can check any number in one click, and Review mode shows the few cells Sleuth couldn't settle, and why. [Review mode](review-mode.md)

## Related

- [Check a dataset with Review mode](review-mode.md): See only the cells worth a second look, check their sources, and mark them reviewed.
- [Datasets: rows, columns, and cells](datasets.md): What a Sleuth dataset is, which cells use credits, and how to change one.
- [How fresh the data is](data-freshness.md): How quickly new trials, filings, and deals reach Sleuth, and how to keep your own work current.

*Updated November 9, 2026.*
